PROTOCOL

https://doi.org/10.1038/s41596-022-00728-0

I-TASSER-MTD: a deep-learning-based platform

for multi-domain protein structure and function
prediction
Xiaogen Zhou 1,2, Wei Zheng 1, Yang Li 1
, Robin Pearce1, Chengxin Zhang1, Eric W. Bell 1
,
Guijun Zhang2 and Yang Zhang 1,3 ✉
Most proteins in cells are composed of multiple folding units (or domains) to perform complex functions in a cooperative
manner. Relative to the rapid progress in single-domain structure prediction, there are few effective tools available for
multi-domain protein structure assembly, mainly due to the complexity of modeling multi-domain proteins, which involves
higher degrees of freedom in domain-orientation space and various levels of continuous and discontinuous domain
assembly and linker refinement. To meet the challenge and the high demand of the community, we developed I-TASSER-
MTD to model the structures and functions of multi-domain proteins through a progressive protocol that combines
sequence-based domain parsing, single-domain structure folding, inter-domain structure assembly and structure-based
1234567890():,;
1234567890():,;

function annotation in a fully automated pipeline. Advanced deep-learning models have been incorporated into each of the
steps to enhance both the domain modeling and inter-domain assembly accuracy. The protocol allows for the
incorporation of experimental cross-linking data and cryo-electron microscopy density maps to guide the multi-domain
structure assembly simulations. I-TASSER-MTD is built on I-TASSER but substantially extends its ability and accuracy in
modeling large multi-domain protein structures and provides meaningful functional insights for the targets at both the
domain- and full-chain levels from the amino acid sequence alone.

Introduction
Much progress has been made in protein structure prediction as a result of decades of effort1–4. The
progress has been particularly notable in recent years owing to the introduction of coevolution-based
contact prediction5–7 and deep neural-network learning techniques8–10. In particular, the end-to-end
sequence-to-structure training approaches, such as AlphaFold2 (ref. 11), built on the attention and
equivariant transformer networks, have achieved unprecedented modeling accuracy in the protein
structure prediction as witnessed in the recent CASP14 experiment12. However, most of the advanced
methods have mainly focused on the modeling of individual domain structures, which are the
minimum folding units of proteins that fold and function independently. In fact, more than two-
thirds of prokaryotic proteins and four-fifths of eukaryotic proteins contain two or more domains13,
where many proteins perform higher-level cellular functions through cooperative domain interac-
tions14,15. Therefore, determining the full-length structures of multi-domain proteins is a crucial step
towards elucidating their full functions and designing new drugs to regulate these functions.
A common approach for multi-domain protein structure modeling is to split the query sequence
into domains and generate models for each individual domain separately16,17. The individual domain
models are subsequently assembled into full-length models, usually under the guidance of other
homologous multi-domain proteins from the Protein Data Bank (PDB)18. However, many multi-
domain proteins have been solved only as single-domain proteins, and just 35.3% of proteins in the
PDB contain multi-domain structures. The lack of homologous multi-domain structures makes
the template-based domain-assembly approach infeasible for most multi-domain protein targets. On
the other hand, template-free (or ab initio) domain structure assembly is challenging, owing to the
fact that multi-domain proteins have a high degree of freedom in domain-orientation space and we
do not have reliable force fields to accommodate the domain–domain interactions. In a recent study,

1
Department of Computational Medicine and Bioinformatics, University of Michigan, Ann Arbor, MI, USA. 2College of Information Engineering,
Zhejiang University of Technology, Hangzhou, China. 3Department of Biological Chemistry, University of Michigan, Ann Arbor, MI, USA.
✉e-mail: [email protected]

2326 NATURE PROTOCOLS | VOL 17 | OCTOBER 2022 | 2326–2353 | www.nature.com/nprot

 NATURE PROTOCOLS PROTOCOL
Zhou et al. proposed a hybrid approach, called DEMO16, which first detects analogous structure
templates through domain-complex structure alignment19. Next, a knowledge-based force field is
combined with deep-learning contact/distance restraints to refine the multi-domain analogous
templates through replica-exchange Monte Carlo (REMC) simulations. As structure alignment can
often detect more distant templates beyond sequence-based alignments, DEMO remarkably enhances
the coverage and ability of the homology-based approaches. Large-scale benchmark tests demon-
strated advantages of DEMO over other state-of-the-art approaches (either homology or ab initio) on
proteins containing both continuous and discontinuous domain structures16.
Although several methods have been proposed for domain boundary prediction20–22 and domain
model assembly16,17,23, there have been very few protocols dedicated to automated multi-domain
protein structure and function prediction from sequence alone. One reason is that the complexity of
modeling multi-domain proteins, which involves various levels of continuous and discontinuous
domain assembly and linker refinement, makes the pipeline development difficult to automate. In
addition, although domains are usually basic units with distinct biological functions, the function of
multi-domain proteins is often predicted using the same strategies used for single-domain proteins;
this is especially true for the modeling of protein-level functions such as Gene Ontology (GO) terms,
while the commonly used GO predictors, such as MetaGO24, NetGO25 and INGA26, do not attempt
to split the sequence into domains and annotate functions at individual domain level. Some studies27
have attempted to assign GO terms to specific regions of the protein, but they are either not
specifically optimized for multi-domain structure models28 or do not use protein structure at all27.
These methods are therefore less useful for the function annotation of multi-domain proteins.
Furthermore, several important questions should be considered to improve the modeling accuracy
when developing such unified multi-domain protein structure and function prediction pipelines.
First, coevolutionary analysis and deep learning have been successfully applied to single-domain
protein structure prediction and have notably improved its accuracy29. How can we extend similar
techniques to guide domain parsing and domain model assembly? Second, is there any intrinsic
correlation between the similarity of single-domain proteins and the global similarity of multi-
domain proteins? If yes, how can such correlation be exploited through analogous multi-domain
protein template detection informed by the structural alignment of individual domains? Third, how
can we incorporate the sparse restraints, such as cryo-electron microscopy (EM) density maps and
cross-linking data, into the simulation to guide the full-length structure modeling? Finally, given that
the distinct functions of individual domains usually contribute towards the overall protein function15,
how can we draw a more comprehensive function annotation by combining domain-level and full-
length function prediction?
To address the above questions and meet the high demand of the community, we developed
I-TASSER-MTD (with MTD standing for ‘multi-domain’), a fully automated pipeline for multi-domain
protein structure assembly and structure-based function annotation from sequence alone (the I-TASSER-
MTD platform is freely available at https://zhanggroup.org/I-TASSER-MTD/)16,30,31. The core of
I-TASSER-MTD is built on I-TASSER4 and DEMO16, where the former has consistently been ranked as
the top method for automatic protein structure prediction in the last eight iterations of the community-
wide CASP experiments31–39, while the online server for I-TASSER has completed structure and function
predictions for >680,000 proteins submitted by >160,000 registered users from 159 countries40. With the
integration of cutting-edge inter-domain assembly methods, as well as advanced deep neural-network
techniques41, I-TASSER-MTD dramatically extends the ability and capacity of I-TASSER for modeling
large multi-domain protein structures. Meanwhile, unlike pure deep-learning methods, which largely act
as a block box, the accessibility and interpretability of the domain structure assembly process allows for
the I-TASSER-MTD server to effectively incorporate the restraints from cryo-EM and cross-linking
experiments as input by users. As a validation of the pipeline, I-TASSER-MTD participated in the most
recent CASP14 (under the group name ‘Zhang-Server’) and achieved the best performance on modeling
the multi-domain proteins among all server groups in the experiment30. To the best of our knowledge,
this represents the first protocol devoted specifically to the modeling of multi-domain proteins, which we
believe will benefit the biological and biomedical communities.

The I-TASSER-MTD pipeline

I-TASSER-MTD is built on multiple tools that we recently developed for sequence-based domain
boundary prediction, deep-learning spatial restraint prediction, single-domain structure folding,
multi-domain structure assembly and structure-based function annotation. The pipeline is depicted

NATURE PROTOCOLS | VOL 17 | OCTOBER 2022 | 2326–2353 | www.nature.com/nprot 2327

 PROTOCOL NATURE PROTOCOLS

MNELVDTTEMYLRTIYDLEEEGVTPLRARIAERLDQSGPTVSQTVSRMERDGLLRVAGDRLELTEKGRALAIAVMR
KHRLAERLLVDVIGLPWEEVHAEACRWEHVMSEDVERRLVKVLNNPTTSPFGNPIPGLDELGVGPEPGADDANLV
LOMETS2
LOMETS2
threading BioLiP
structure–function
database
Structure database Threading templates Template fragments

Single-domain or coverage ≥Cov ?

Easy target and coverage >Cov? Final full-length model
UniProt-GOA
Domain N Domain C
MNELVDTTEMYLRTIYDLEEEGVT sequence–function
PLRARIAERLDQSGPTVSQTVSR
MERDGLLRVAGDRLELTEKGRAL database
Template AIAVMRKHRLAERLLVDVIGLPWE
EVHAEACRWEHVMSEDVERRLV
KVLNNPTTSPFGNPIPGLDELGVG
Query PEPGADDANLVRLTELPAGSPVAV
VVRQLTEHVQGDIDLITRLKDAGV
VPNARVTVETTPGGGVTIVIPGHE
Linker NVTLPHEMAHAVKVEKVMNELVD Iterative
fragment assembly STRING
Domain conservation score Domain boundary

DeepPotential
PPI–function
ThreaDom database
Domain assembly
G T N V P L T F
D3
G S N V P L A F
G S D V P M T F
G T N I P L A I
G S N V P L A I D2
G S D V P M A S EC number GO terms
G S V V P M A F
G S V V D M A F D1 CC BP MF
Sequence MSA
database Folding unit detection Clustering
Hard target or Multi-domain and
coverage ≤Cov? coverage <Cov? Analogous templates

Binding site
50 residual
blocks Spatial restraints Individual domain models Multi-domain library Function prediction
Query
sequence DeepPotential FUpred D-I-TASSER DEMO COFACTOR

Fig. 1 | Overview of the I-TASSER-MTD protocol for multi-domain protein structure and function prediction. Cov is the cutoff of the alignment
coverage for assessing if the query needs to be modeled as a single-unit or multi-unit target, where a unit can contain a single domain or multiple
domains if the latter is fully covered by the LOMETS2 threading alignments. CC, BP and MF represent, respectively, the cellular component, biological
process and molecular function in GO. It should be noted that, following their implementation on full-length sequence, LOMETS2 and DeepPotential
will be run again in D-I-TASSER to generate templates/restraints for each individual domain if the query is deemed as a multi-unit target.

in Fig. 1. Starting from the query sequence, multiple threading templates are first collected by
LOMETS2 (ref. 42) a meta-server approach that combines up to 11 sequence/hidden Markov model
profile-based methods and deep-learning threading programs—from the PDB, and domain bound-
aries are predicted by FUpred22 and ThreaDom20,43. Meanwhile, residue–residue spatial restraints are
created by DeepPotential44 through residual convolutional network training. If the query sequence is
deemed a multi-domain protein by FUpred or ThreaDom, and none of the top ten template align-
ments can cover all domains (i.e., one or more domains with an alignment coverage below the cutoff
Cov = 95%), the ‘multi-domain assembly’ mode will be initiated, where models for each domain will
be independently constructed by D-I-TASSER30, a crucially improved version of I-TASSER powered
by the DeepPotential spatial restraints. Subsequently, the domain models will be assembled into the
full-length model by DEMO16 based on the structurally analogous templates. Otherwise, if the query
is deemed a single-domain protein or one or more top template alignments can cover all domains, the
‘multi-domain assembly’ mode will be turned off and the full-length structure will be directly
modeled by D-I-TASSER. Finally, the protein function annotations, including the Enzyme Com-
mission (EC) numbers, GO terms and ligand-binding sites are predicted by COFACTOR45 for all
individual domains and the full-chain protein, based on the modeled structures, sequences and
protein–protein interactions (PPIs). The main individual programs employed in the I-TASSER-MTD
pipeline are listed in Supplementary Table 1. Next, we briefly describe the main components of the
I-TASSER-MTD pipeline.

Sequence-based domain boundary prediction

Accurate domain boundary assignment is crucial for multi-domain protein structure and function
prediction. In the first stage of I-TASSER-MTD, the query sequence is split into domains by com-
bining a threading template-based method, ThreaDom20,43, with a deep-learning contact-based

2328 NATURE PROTOCOLS | VOL 17 | OCTOBER 2022 | 2326–2353 | www.nature.com/nprot

 NATURE PROTOCOLS PROTOCOL
program, FUpred22 (Supplementary Fig. 1). For doing this, the query sequence is first threaded
through the PDB by LOMETS2 (ref. 42), which utilizes and combines 11 state-of-the-art threading
programs, to create multiple template alignments, where the query is assigned as an Easy (or Hard)
target if outstanding templates are (or are not) identified. If the protein is defined as an Easy target by
LOMETS2 and the alignment coverage is >95%, ThreaDom will be applied to predict the domain
boundary. In ThreaDom, the domain conservation score (DCS), which linearly combines the tem-
plate domain linker score and the gap penalty score, is calculated according to the LOMETS2
template alignments. The domain linker score is evaluated on the basis of the domain boundary
definition of the templates in CATH46 or defined by DomainParser47, while the gap penalty score is
measured by the number of gaps in the multiple template alignments. Finally, the domain boundaries
are determined by the DCS using a target-specific scoring cutoff.
If the protein is defined as a Hard target by LOMETS2 or the alignment coverage is ≤95%, the
domain boundary will be predicted by FUpred by maximizing the number of intra-domain contacts
and minimizing the number of inter-domain contacts that are predicted by a deep residual con-
volutional neural network model. To create the deep-learning model, the multiple sequence align-
ment (MSA) is first created by iteratively searching the query against four metagenome sequence
databases (Metaclust48, BFD49, Mgnify50 and IMG/M51) and two whole-genome databases (Uni-
clust3052 and Uniref9053) using DeepMSA54. Then the secondary structure is predicted by PSSpred55,
and the contact map is generated by the deep-learning-based predictor ResPRE10, which has been
included in DeepPotential, using features directly extracted from the MSA. To deduce the domain
boundaries from the contact map, a folding unit score (FU-score)22 is calculated for continuous
domain and discontinuous domain detection (for continuous and discontinuous domain definition,
see Supplementary Fig. 2). For a continuous two-domain protein, the FU-score of each residue is
defined in relation to the number of intra-domain contacts and the number of inter-domain contacts
for the two domains by considering the residue as a domain boundary. To calculate the FU-score of a
residue pair (i, j) in a discontinuous two-domain protein, the C-terminal contact map (j + 1, L) is
shifted to the N-terminal contact map (1, i) to convert the discontinuous domain into a continuous
one to form a new contact map, where i < j, and L is the sequence length of the query protein
(Supplementary Fig. 3). The FU-score for the residue pair (i, j) is calculated according to the number
of intra-domain contacts and inter-domain contacts in the new map. The domain boundary is
determined according to a trained cutoff of the FU-score, and the predicted secondary structure is
also used to avoid the boundary being located within a helix or a strand. As the FU-score is defined
for the two-domain case, the above procedure is recursively performed for the determined domains
until no additional domains can be detected.

Deep neural network-based spatial restraints prediction

DeepPotential44, a newly developed deep residual neural-network-based predictor, is used to create
multiple spatial restraints, including distance maps for both Cα and Cβ atoms, Cα-based hydrogen-
bonding networks4 and Cα–Cβ torsion angles. Different from the contact map prediction, DeepPo-
tential, whose architecture is depicted in Supplementary Fig. 4, predicts the probability that inter-
residue distances fall within 36 equal-width bins from [2, 20] Å, as well as two additional bins with
distances <2 Å and >20 Å. The DeepPotential program was trained on a nonredundant dataset of
26,151 solved protein structures obtained from the PDB. Starting from the query sequence, three 1D
features (i.e., field parameters of the Potts model, a one-hot representation of the sequence, and a
hidden Markov model) and two 2D coevolutionary features (i.e., the pseudolikelihood maximized
Potts model56 and mutual information) are extracted from the MSA created by DeepMSA. The 1D
features are tiled in two dimensions and concatenated with the 2D features before being fed into the
neural network, which includes ten 1D residual blocks41 and ten 2D residual blocks. The neural
network model was trained by the Adam optimization algorithm to marginally minimize cross-
entropy loss. In addition to the 20 Å Cα/Cβ distance cutoff, distance maps with multiple thresholds
(i.e., 10, 13, 16 and 19 Å), inter-residue torsion angles57 and hydrogen-bonding networks are also
considered in a multi-task learning strategy58. The contact maps and domain–domain interface maps
are extracted from the predicted distances by the summation of the cumulative probability of dis-
tances <8 Å and distances <18 Å, respectively.

Individual domain structure modeling with D-I-TASSER

The structural model for each domain is built using D-I-TASSER, which is an extended version of
I-TASSER4,40,59–61 that integrates the DeepPotential spatial restraints with iterative threading

NATURE PROTOCOLS | VOL 17 | OCTOBER 2022 | 2326–2353 | www.nature.com/nprot 2329

 PROTOCOL NATURE PROTOCOLS

assembly simulations (Supplementary Fig. 5). For each domain, LOMETS2 is used again to identify
domain-level structural templates from a nonredundant PDB structural library. Meanwhile, distance
maps, hydrogen-bonding networks4 and inter-residue torsion angles are predicted by DeepPotential.
The contact maps are also predicted by four deep-learning-based methods (ResPre10, DeepPLM31,
ResTriplet62 and TripletRes63) and a naïve Bayes-based contact predictor (NeBcon64) (see the
description and benchmark results reported in Supplementary Note 1). Then, the domain-level
structure models are constructed using REMC simulations under the guidance of the deep-learning-
based spatial restraints and the I-TASSER potential4, the latter of which contains generic statistical
potentials and threading template-based restraints. Five REMC simulations are performed in parallel
for each protein, where the structural decoys from 8 (or 3 for hard targets) low-temperature replicas
are clustered by SPICKER65. Finally, the decoy of the center of the largest cluster is selected as the
final model, and the side-chains of the final model are then repacked by FASPR66, which is further
refined by FG-MD67. It should be noted that some programs (e.g., LOMETS2 and DeepPotential) that
have been implemented at full-chain level are performed again in this step to create D-I-TASSER
models for each domain. As the programs and the D-I-TASSER force field have been trained mainly
on the domain level and many template structures solved in the PDB contain only single domains,
the domain-level modeling often generates more reliable results than that starting from the full-chain
sequences; this is also one of the major motivations for the development of the I-TASSER-
MTD pipeline.

Domain assembly through analogous global structural alignments

After all individual domain models are created, they are assembled into full-length models
using DEMO16 under the guidance of the analogous templates detected by domain-complex
structural alignments19 (Supplementary Fig. 7). For this, a nonredundant multi-domain protein
library, which is updated weekly, has been constructed by collecting all multi-domain protein
structures from the PDB. The individual domain models are first aligned to each template of the
library by TM-align19, and the harmonic mean of the TM-score68 of all domains is defined as the
score of the template (T-score). Then, five initial full-length models are respectively constructed on
the basis of the top five templates with the highest score by searching the best position of each domain
on the template through a sliding-window-based procedure. For each initial model, REMC simula-
tions with all domains randomly rotated and translated as rigid bodies are performed to optimize the
orientation of each domain under the guidance of an energy function (Supplementary Note 2) that
includes Cα clashes between domains, domain boundary connectivity, inter-domain distances and
domain–domain interfaces predicted by DeepPotential, a generic inter-domain contact potential,
local domain distance restraints from the initial domain-template superpositions and inter-domain
distance restraints deduced from the top templates. Owing to errors in the tail of domain models,
linkers between domains may be disconnected after the assembly. To reconnect the domain struc-
tures, the linker residues of the full-length model with the lowest energy are relaxed, where the Cα
atoms are regenerated by self-avoiding random walks, followed by adding the N, C, O atoms, and
side-chain centers using FASPR66. Next, Metropolis Monte Carlo simulations are performed to refine
the linker model guided by a potential containing a Cα clash term between the linker and domain
structures, a statistical torsion-angle potential from Ramachandran plots3,69, an orientation-
dependent side-chain contact potential, and a statistical N–Cα–C bond angle potential. Finally, the
model with the lowest linker energy is selected for side-chain reconstruction and refinement by
FASPR66 and FG-MD67.

Structure-, sequence- and PPI-based function annotation

The protein function, including GO terms, EC numbers and ligand-binding sites of the full-length
protein structure and the individual domains are annotated by our latest version of COFACTOR45.
This new version of COFACTOR includes not only the structure-based function prediction pipeline
inherited from the old COFACTOR program70, but also new methods for sequence- and PPI-based
function prediction (Supplementary Fig. 8). In the structure-based pipeline, the predicted model is
searched by TM-align through the BioLiP library71 for structure templates with known GO terms, EC
numbers and/or ligand-binding residues. Template residues in known active sites and/or ligand-
binding sites are realigned by local structure similarity. The functions are transferred from structure
templates to target structure by both global and local structural similarity. In the sequence-based
pipeline, the target sequence is used to search through the UniProt-GOA database for BLAST and
PSI-BLAST hits with GO annotations. GO terms are transferred to the target through a weighted

2330 NATURE PROTOCOLS | VOL 17 | OCTOBER 2022 | 2326–2353 | www.nature.com/nprot

 NATURE PROTOCOLS PROTOCOL
K-nearest-neighbor approach, where K is the number of (PSI-)BLAST hits with E-values ≤0.01 and
the weight for each template is equal to the global sequence identity of the target-template alignment.
The PPI-based pipeline of the latest version of COFACTOR is ported from MetaGO24. In this
pipeline, the target sequence is mapped by BLAST to the STRING database to identify PPI partners.
The sequences of PPI partners are then mapped to UniProt-GOA by BLAST. GO terms are also
predicted by a weighted K-nearest-neighbor method, where the weight of each UniProt-GOA
sequence is calculated as the product of two values: the sequence identity between the UniProt-GOA
sequence and the STRING PPI partner, and the STRING score between the target sequence and the
PPI partner. Finally, the structure-, sequence- and PPI-based predictions are combined using
weighted averaging to derive the final consensus function prediction.

Estimation of model quality

Estimating the accuracy of a predicted model is essential to decide how users will use the model in
their research. In I-TASSER-MTD, the accuracy of the kth predicted model is estimated by the
estimated TM-score, eTM-score(k), which is calculated on the basis of the convergence of the domain
assembly simulations, the confidence of the full-length templates identified for domain assembly, the
satisfaction rate of the predicted inter-domain distances and the estimated accuracy of the individual
domain model by D-I-TASSER, using the following equation:

  10   T 
P T -scoreðiÞ P  pre 

eTM-scoreðkÞ ¼ w1 ln MMðtotkÞ ´ 1
hRMSDik þ w2 ln 10
1
T -score0 þ w3 wneff ln T
1
dt  dt
model
ðkÞ
i¼1 t¼1
 pre model  NP
OðI ;I Þk dom
þw4 wneff ln N ðI Þ
pre þ w5
1
Ndom eTM-scoredom ðDÞ þ w6
D¼1
ð1Þ

The first term in Equation (1) evaluates the degree of convergence of the domain assembly
simulations, where Mtot is the total number of full-length decoys generated in the domain assembly
simulations, M(k) is the number of structure decoys with root-mean-square deviation (RMSD) <1.5 Å
to the kth full-length model and hRMSDik denotes the average RMSD between these decoys and the
kth reported model. The second term assesses the quality of the full-length template, where T-score(i)
is the template score of the ith full-length template, which is calculated as the harmonic mean of the
TM-scores between the domain models and the full-length template that is used for DEMO-based
domain assembly, and T-score0 = 0.85 is the cutoff used to distinguish good from bad templates. The
third term assesses how closely the distances in the reported model match the predicted distances by
DeepPotential, where T is the number of predicted inter-domain distances used to guide the domain
pre
assembly, and dt and dtmodel ðkÞ are the distances of the tth residue pair in the predicted distance map
and the kth reported model, respectively. The fourth term accounts for the domain–domain interface
satisfaction rate of the predicted interface map in the reported model, where N(Ipre) is the number of
predicted domain–domain interfaces and OðI pre ; I model Þk is the number of overlapped interfaces
between the predicted interface map and the kth reported model. As restraints in the third and fourth
terms are predicted using MSAs, wneff is a weight associated with the quality of the MSA and
calculated on the basis of the number of effective sequences (neff; Supplementary Eq. (S12)). Finally,
the fifth term accounts for the quality of individual domain models from D-I-TASSER, where Ndom is
the total number of domains and eTM-scoredom ðDÞ is the estimated TM-score of the Dth domain
model from D-I-TASSER (Supplementary Note 3). w1 = 0.065, w2 = 0.063, w3 = −0.08, w4 = 0.01,
w5 = 0.96 and w6 = 0.1 are the weighting factors, which are optimized using an improved differential
evolution algorithm72–74 to minimize the error between the eTM-score and the actual TM-score of
the decoys to the native structure on the DEMO training set of 425 nonredundant multi-domain
proteins. In addition, the estimated RMSD (eRMSD) of I-TASSER-MTD models is also calculated
on the basis of the same terms in Equation 1 but with an additional term to count for the protein
length (L) (Eq. (S14) in Supplementary Note 4), where the weighting factors for eRMSD are w1 = −1.40,
w2 = −2.74, w3 = 4.78, w4 = −1.19, w5 = −16.43, w6 = 0.0 and w7 = 2.66. Meanwhile, we define a new
score to quantitatively assess the relative populations of the assembled conformations for the kth
reported model by

pðkÞ
P-scoreðkÞ ¼ P ð2Þ
k pðkÞ

NATURE PROTOCOLS | VOL 17 | OCTOBER 2022 | 2326–2353 | www.nature.com/nprot 2331

 PROTOCOL NATURE PROTOCOLS

where pðkÞ ¼ MMðtotkÞ is the normalized number of structural decoys of kth model in the I-TASSER-MTD
assembly simulations.
The accuracy of eTM-score and eRMSD were examined on the DEMO benchmark set, which
includes 356 multi-domain proteins with different domain types that are nonhomologous to the
DEMO training dataset. As shown in Supplementary Fig. 9, the eTM-score has a high Pearson
correlation coefficient (PCC = 0.85) with the actual TM-score, where the average error of the
eTM-score is 0.07. Compared with the eTM-score, the eRMSD and RMSD have a slightly lower
correlation (PCC = 0.82), where the average error between eRMSD and RMSD (2.2 Å) is relatively
high (see the distribution in Supplementary Fig. 9d). It should be noted that RMSD is not the best
measurement for the accuracy of predicted models when the modeling accuracy is low; as all residue
pairs have the same weight in the RMSD calculation, this renders the RMSD value sensitive to local
variations, such as tails or loops, rather than the global fold. In this regard, it is recommended to use
TM-score as a more reliable measurement for model accuracy assessment; because the residue pairs
with smaller distance errors are weighted more heavily than the residues with larger errors in the TM-
score calculation, the TM-score value is generally more sensitive to the accuracy of the global fold of
the predicted models68.
As shown in Supplementary Fig. 9a, if we use an eTM-score cutoff of 0.5 to select models with
correct global topologies, both the false-negative and false-positive rate are <0.15, indicating that the
fold-level prediction by the eTM-score is correct in >85% of the cases. As illustrative examples, six
models from two targets are included in Supplementary Fig. 10, showing that the eTM-score is highly
correlated with the actual TM-score of the models in different ranges of model quality. In Supple-
mentary Fig. 11, we also show the eTM-score and thus the model quality will be reduced when more
and more random mutations are introduced to the target sequence of PDBID 1we3F.
In addition to the global quality assessment, to show the local accuracy of each individual domain
model, I-TASSER-MTD estimates the residue-level distance error of the predicted model relative to
the native structure using ResQ75, a method for estimating residue-level quality in protein structure
prediction on the basis of local variations of modeling simulations and the uncertainty of homologous
template alignments.

Experimental design incorporating user-specific data

User-specified domain boundaries
While the I-TASSER-MTD server is capable of fully automated domain partition, it optionally allows
users to provide their own domain definition. If a complete and effective domain definition is
provided, the server will use the provided domain information to split the query sequence into
multiple parts for independent structure modeling, rather than running FUpred or ThreaDom for the
domain boundary prediction. If only a partial domain definition is provided, for example, with a
mixed input sequence of defined and unknown domains, the server will keep the user-defined
domains and predict the domains of unknown regions by FUpred or ThreaDom. After inputting the
domain definition, the server will automatically check the effectiveness of the given domain defini-
tion. Specifically, a domain or a segment should be represented by a starting residue index and ending
residue index with a hyphen between them. Domains should be separated by semicolons, while
segments of a discontinuous domain should be separated by commas. The length of a domain should
be between 30 and 1,000 amino acids, which is the range of domain lengths for the majority of CATH
domains46. Missing or overlapping residue indices are not allowed in the domain definition if users
want to completely use their own domain definition rather than one determined by the server. All
domains should be written on one line and end with semicolons (for details, see Box 1).

User-specified full-length structure templates

The I-TASSER-MTD server assembles multiple domain structures using analogous full-length tem-
plates followed by knowledge-based domain structure refinement simulations. By default, the server
detects the analogous full-length templates from the multi-domain protein library through structural
alignments by TM-align, where the library includes multi-domain proteins at a pair-wise sequence
identity <70% or sequence identity ≥70% but with TM-scores <0.5. Alternatively, the server allows
users to specify solved protein structures or separately predicted models as templates. The user-
specified templates must be in PDB format. Users can provide up to 20 templates and compress them
into one file to upload to the server. If more than five templates are provided, each template will be
evaluated by the harmonic mean of the TM-score between all domain models and the template, and

2332 NATURE PROTOCOLS | VOL 17 | OCTOBER 2022 | 2326–2353 | www.nature.com/nprot

 NATURE PROTOCOLS PROTOCOL
Box 1 | Providing the domain definition
The three cases listed below should be considered in the provided domain definition.
1 Proteins with a complete continuous domain definition can be pasted in the text box like this example (PDBID:
2qbuA, Supplementary Fig. 2a):

1  131; 132  228;

where 1–131 indicates that residue 1 to residue 131 belong to the first domain and residues 132–228 to the
range for the second domain. Note that domains must be separated by semicolons, and a semicolon should be
included at the end of the definition.
2 Proteins containing discontinuous domains in a complete domain definition should be prepared similarly to
the following example (PDBID: 1atgA, Supplementary Fig. 2b):

1  81; 191  232; 82  190;

where ‘1–81,191–232’ represents the definition of the first domain, which is a discontinuous domain; 1–81
indicates that residues from positions 1 through 81 make up the first domain segment; and 191–232 indicates
that residues from 191 to 232 belong to its second segment. 82–190 is the range of the second continuous
domain, which is inserted in the first discontinuous domain. Note that different parts of a discontinuous
domain must be separated by commas.
3 The partial domain definition should be prepared like the following example (PDBID: 1h88C, Supplementary
Fig. 2c):

1  51;
where ‘1–51’ indicates that one of the domains should include residues from 1 to 51. The total sequence length
of the protein is 152, and the complete domain definition should be ‘1–51;52–105;106–152;’ as defined in the
CATH database. The server will keep the domain provided by users when predicting the domain boundary.
When providing the defined domain, please ensure that the length of every undefined domain region is be
larger than the minimum length (30) of a domain.

the top five templates with the highest scores will be selected for the initial full-length model
generation. If only one template is specified, the top four analogous templates detected from the
library and the uploaded template will be used to create the initial full-length model. For both cases,
the domain-template alignments are determined by a sliding-window procedure based on TM-score
as used by DEMO16. Furthermore, all templates provided by the user will be utilized to extract the
inter-domain distance profiles, which will be used to mainly guide the domain assembly simulations
along with restraints deduced from the templates identified from the library. If the provided template
comes from computationally modeled structures (such as those from AlphaFold2 predictions) and
includes the confidence score (pLDDT)11 in the ‘Temperature factor’ column, only the distances of
the residue pairs with pLDDT >70 for both residues are extracted from the template to guide the
assembly simulations.

Incorporating experimental cross-linking data or contact/distance restraints

Distance restraints and spatial contact information for a protein subunit or domain derived from
chemical cross-linking by mass spectrometry can be very useful for protein structure modeling. Given
this, the I-TASSER-MTD server provides an option to incorporate user-specified cross-linking data as
an additional restraint to guide the domain model assembly. Users can upload the cross-linking data
to the server, and the server will automatically check and confirm that the data are in the correct
format: the first column and the second column in the provided file should be the residue indices with
Cα distances less than a cutoff, which should be specified in the third column; headers or footers are
not allowed in the data file. Once effective cross-linking data are given, they are converted to a three-
gradient contact potential16 and added to the DEMO assembly energy function, where the weight of
the cross-linking potential was optimized over a training set. The other parts of the DEMO assembly
process will remain unchanged. It should be noted that, since the cross-linking data are implemented
as residue–residue contacts, the provided data are not limited to cross-linking. Instead, contact or
distance restraints obtained from any methods can be written in the specified format and input to the
server to guide the domain model assembly (for details, see Box 2). To allow I-TASSER-MTD to
better implement the restraints, users are recommended to provide a confidence score of each residue
pair in the fourth column when providing the predicted contact or distance restraints. All confidence
scores should be in the range of [0, 1], and they will be considered as 1 if not provided.
The confidence score of each residue pair will be used to determine the weight of the residue pair in
the contact potential (Supplementary Note 4). As residue pairs with low confidence scores usually

NATURE PROTOCOLS | VOL 17 | OCTOBER 2022 | 2326–2353 | www.nature.com/nprot 2333

 PROTOCOL NATURE PROTOCOLS

Box 2 | Incorporating residue–residue contact restraints

Residue–residue contacts from experimental cross-linking data or predicted by any other programs can be
prepared and pasted in the text box on the main page of the server similar to the following example (PDBID:
1fx7A):

15 165 28 0:85
63 220 21 0:96
94 168 17 0:79
where the first and the second columns are the residue indices, and the third column is the maximum Cα distance
for the residue pair. The fourth column is the confidence score that the distance of the residue pair is less than
the given distance listed in the third column. The confidence score is optional, and it will be set to 1 if not
provided. Values in the same row should be separated by tabs or spaces. For example, the first row ‘15 165 28
0.85’ indicates that the Cα distance between residues 15 and 165 has a confidence of 0.85 being less than 28 Å.
Note that the residue index is for the full-chain sequence rather than each individual domain, and the residue
index starts from 1 rather than 0.

have relatively low accuracy in the contact/distance prediction, users are recommended to provide
restraints with confidence score >0.5 to reduce noise.

Integrating experimental cryo-EM data

Cryo-EM has become established as powerful method for macromolecular structure determination in
recent years76. The I-TASSER-MTD server is equipped to integrate cryo-EM density maps to assist
domain assembly and refinement (Supplementary Fig. 12). Specifically, when a cryo-EM density map
is available, each domain model is first matched into the density map by performing limited-memory
Broyden–Fletcher–Goldfarb–Shanno (L-BFGS) optimization with different starting positions to
identify the best location and orientation with the highest density correlation to the density map. The
top poses of each domain are combined to build multiple initial full-length models. The full-length
models generated on the basis of the templates identified by TM-align are also fit into the density map
using the same L-BFGS algorithm. According to the density correlation score77 between the density
probed from the full-length model and the experimental density, the top full-length models are
selected to optimize the orientations of all domains by DEMO rigid-body assembly under the gui-
dance of the density correlation score and the inherent DEMO assembly force field. To further
improve the individual domain models and the full-length model, the top models are selected from
the rigid-body assembly results according to the density correlation score, and atom-, segment- and
domain-level refinement are performed using REMC simulations guided by a knowledge-based
force field, inter-domain distances predicted by DeepPotential, and the density correlation score.
Finally, the full-length model with the lowest energy is selected for side-chain repacking by FASPR
and FG-MD.

Comparison with other methods

Although a number of webservers have been developed to predict the structures and functions of
proteins, very few are devoted to fully automated multi-domain protein structure and function
prediction from sequence alone. The I-TASSER-MTD server distinguishes itself from other protein
structure prediction servers in several major aspects pertaining to multi-domain proteins (Table 1).
First, unlike other protein structure prediction servers where all query proteins are considered as
single-domain proteins or the domain boundaries are roughly determined using external online
programs such as PFAM78 or the NCBI Conserved Domain Database search service79, I-TASSER-
MTD automatically detects the domains according to a combination of threading- and deep-learning
based domain boundary prediction algorithms depending on the target type (homologous or non-
homologous). Of particular significance, I-TASSER-MTD can effectively predict complex dis-
continuous domains that consist of multiple separate segments at the sequence level. Second, for
multi-domain proteins, almost all servers can generate individual domain models by submitting the
individual domain sequences independently, but few can automatically assemble the full-length
model. The I-TASSER-MTD server not only can construct the models of individual domains, but can
also assemble all domain models into the full-length model. Third, an important unique feature of
I-TASSER-MTD is that it provides options for several types of experimental restraints to guide the
domain model assembly, including homologous templates, cross-linking data, cryo-EM density maps
and other sources of residue–residue contact or distance information. Fourth, unlike other servers

2334 NATURE PROTOCOLS | VOL 17 | OCTOBER 2022 | 2326–2353 | www.nature.com/nprot

 NATURE PROTOCOLS PROTOCOL
Table 1 | Comparison of I-TASSER-MTD with other protein structure prediction servers

Name URL DBP IDM MDA DF, FF EDM

I-TASSER-MTD https://zhanggroup.org/I-TASSER-MTD/ Yes Yes Yes DF + FF Yes

RoseTTAFold82 http://robetta.bakerlab.org/ No Yes No No No
RaptorX103,104 http://raptorx.uchicago.edu/ Yes Yes No DF/FF No
trRosetta57,105 https://yanglab.nankai.edu.cn/trRosetta/ No Yes No No No
PSIPRED106 http://bioinf.cs.ucl.ac.uk/psipred/ Yes Yes No DF/FF No
HHpred107 https://toolkit.tuebingen.mpg.de/tools/hhpred No Yes No No No
Phyre108 http://www.sbg.bio.ic.ac.uk/phyre2/ Yes Yes No DF/FF No
DBP, domain boundary prediction; DF, domain level function prediction; EDM, experimental data-assisted modeling; FF, full-length level function
prediction; IDM, individual domain modeling; MDA, multi-domain protein structure assembly.

that mostly predict the function of the individual domains, I-TASSER-MTD performs the function
annotation of the query protein at both the domain level and full-length level.
As a blind test, I-TASSER-MTD (as ‘Zhang-Server’) participated in the most recent community-
wide CASP experiment for fully automated protein structure prediction. In Fig. 2a, we present a
summary of the five best performing servers in CASP14, in which we sorted the servers according to
the average global distance test (GDT) score of the full-length models for all multi-domain proteins
with one or more template-free modeling (FM) or template-free modeling/template-based modeling
(FM/TBM) domain. The average GDT score of I-TASSER-MTD for the multi-domain proteins was
the highest among all participating servers. The accuracy of the individual domain models for multi-
domain proteins was also higher than that of other servers. For example, I-TASSER-MTD achieved
an average GDT score of 61.4 for all individual domain models of the multi-domain proteins, which
was 19.4% higher than that of the second-best server, ROSETTA (51.4). This is mainly due to the
incorporation of the highly accurate deep-learning-based restraints from DeepPotential in the
I-TASSER-MTD simulations. The average GDT score of I-TASSER-MTD was also ranked the highest
for the structure modeling of single-domain proteins in CASP14.
Owing to the employment of FUpred and ThreaDom, I-TASSER-MTD can accurately distinguish
multi-domain proteins from single-domain proteins and predict the domain boundaries with rea-
sonable accuracy. Here, since we cannot obtain the domain definitions used by other servers in CASP,
I-TASSER-MTD is compared with two state-of-the-art methods ConDO21 and DoBo80 on all
CASP14 targets. As shown in Fig. 2b, the accuracy of I-TASSER-MTD domain boundary prediction
is significantly higher than the that of the two control methods in terms of normalized domain
overlap (NDO) score81 for the protein domain boundary prediction, as well as accuracy (ACC) and
Matthew’s correlation coefficient (MCC) for protein classification. For example, the NDO score of
I-TASSER-MTD for multi-domain protein was 0.86, which was 65.4% and 79.2% higher than that of
ConDO (0.52) and DoBO (0.48), respectively.
We also compared I-TASSER-MTD with AlphaFold2 (ref. 11) and RoseTTAFold82 on all CASP14
targets. As the AlphaFold2 results reported in CASP14 are based on the human-expert group, while
the results of I-TASSER-MTD (as ‘Zhang-Server’) are based on the automated server group, we
regenerated all models by running the standalone AlphaFold2 package for a fair comparison. The
average TM-score of AlphaFold2 is 0.84, which is considerably higher than that of I-TASSER-MTD
(0.65). For RoseTTAFold, which has two options, I-TASSER-MTD’s TM-score is slightly higher than
the RoseTTAFold end-to-end version (0.63) but slightly lower than the RoseTTAFold pyRosetta
version (0.69).
To highlight the effectiveness of I-TASSER-MTD on some protein targets, we list in Fig. 2c–e three
examples of multi-domain models built by I-TASSER-MTD that had a significantly higher TM-score
than models built with the state-of-the-art programs. First, Fig. 2c shows the comparison between the
native and predicted structures of human complement component C6 (Uniprot ID: P13671). Although
AlphaFold2 almost correctly predicted all domain models (TM-score 0.78, 0.93, 0.93, 0.88 and 0.87),
the domain orientations were not correctly generated, resulting in a full-length model with a poorer
TM-score/RMSD of 0.63/31.1 Å. The I-TASSER-MTD model obtained a TM-score/RMSD of 0.95/
3.2 Å since it correctly generated both domain models and inter-domain orientations after the
assembly. Figure 2d shows the second example from Sarcoplasmic/endoplasmic reticulum calcium
ATPase 2 (Uniprot ID: P16615), where I-TASSER-MTD generated a better-quality model (TM-score/

NATURE PROTOCOLS | VOL 17 | OCTOBER 2022 | 2326–2353 | www.nature.com/nprot 2335

 PROTOCOL NATURE PROTOCOLS

a Full-length model of Domain model of Single-domain b

multi-domain protein multi-domain protein protein Zhang-Server ConDo DoBo
70
GDT score (average)

1
60

Score (average)
0.8
50
0.6
40
30 0.4

20 0.2
Multi NDO Single NDO Overall ACC Overall MCC
Ya SE er
-S TA
ap r
rX
d

Ya SE er
-S TA
ap r
rX
d

Ya SE er
-S TA
ap r
rX
d
R rve

R rve

R rve
ol

ol

ol
v

v
to

to

to
ng T

ng T

ng T
tF

tF

tF
RO er

RO er

RO er
e

e
S

S
g-

g-

g-
an

an

an
Zh

Zh

Zh
c d e
P13671 P16615 T1092

I-TASSER-MTD model
TM score = 0.95
RMSD = 3.2 Å
I-TASSER-MTD model I-TASSER-MTD model
TM score = 0.97 TM score = 0.82
AlphaFold2 model RMSD = 1.7 Å RMSD = 5.3 Å
TM score = 0.62
RMSD = 31.1 Å

AlphaFold2 model RoseTTAFold model (pyRosetta)

TM score = 0.65 TM score = 0.53
RMSD = 11.5 Å RMSD = 14.2 Å

Gray: native structure Color: predicted model

Fig. 2 | Comparison between I-TASSER-MTD and other methods. a, Comparison between I-TASSER-MTD (Zhang-Server) with the other top four
servers of CASP14 on modeling the full-length multi-domain targets assessed by the GDT score. b, Comparison of I-TASSER-MTD with ConDo and
DoBo for the protein domain boundary prediction on the CASP14 targets, where the y-axis is the NDO score of the multi-domain protein, the NDO
score of the single-domain protein, and the ACC and MCC for the four subparts from left to right. c–e, Representative examples in which I-TASSER-
MTD generated better quality full-length models than AlphaFold2 and RoseTTAFold, where the gray and color cartoon represent native structure and
predicted model, respectively, and different colors indicate different domains: human complement component C6 (P13671) (c); human sarcoplasmic/
endoplasmic reticulum calcium ATPase 2 (P16615) (d); DNA-directed RNA polymerase beta′ subunit (T1092 of CASP14) (e).

RMSD 0.97/1.7 Å) than AlphaFold2 (0.65/11.5 Å), while the latter misfolded the larger-size domain
resulting in an incorrect overall domain orientation. Finally, Fig. 2e presents an example of
the CASP14 target (T1092) from DNA-directed RNA polymerase beta′ subunit. Although the
RoseTTAFold pyRosetta version generated a correct fold for the domain models (TM-score 0.77 and
0.87), the domain orientations were not correctly modeled, resulting a full-length TM-score/RMSD
of 0.53/14.2 Å. Again, I-TASSER-MTD correctly constructed both the domain models and domain
orientations, thus obtaining a full-length model with TM-score/RMSD of 0.82/5.3 Å. These data
demonstrate that I-TASSER-MTD is complementary to these state-of-the-art programs, especially for
long protein sequences with multiple domains, although the overall performance on many other
targets of I-TASSER-MTD, the average TM-score of which is still lower than AlphaFold2, needs
further improvement.
In addition, compared with the deep-learning-based end-to-end models (i.e., AlphaFold2 (ref. 11)
and RoseTTAFold82), which are largely a block box to both developer and users83, I-TASSER-MTD
has several advantages due to the fact that its simulation process is accessible and interpretable. First,
I-TASSER-MTD reports the templates used to model each of the regions (domains) and full-length
protein, which can help users better understand where the predictions come from and therefore
provide functional insights for further studies on the protein. In fact, I-TASSER-MTD offers a
separate section for protein function annotation built on the structural modeling results.

2336 NATURE PROTOCOLS | VOL 17 | OCTOBER 2022 | 2326–2353 | www.nature.com/nprot

 NATURE PROTOCOLS PROTOCOL
Second, I-TASSER-MTD simulations often generate different models for a query protein, which can
be crucial for the protein folding and function study since there are some proteins that adopt
alternative conformations for different states. For example, the human protein Pin1 contains separate
regulatory and catalytic domains which sample ‘extended’ and ‘compact’ states with different
structures84. Experimental studies have shown that there is an equilibrium between ‘compact’ and
‘extended’ states in a rough approximation with populations of 50:50 (ref. 85). Supplementary Fig. 13
shows that the top five models generated by I-TASSER-MTD are highly diverse and include both
‘extended’ and ‘compact’ states, while models constructed by AlphaFold2 converge to a single
‘compact’ state. Accordingly, I-TASSER-MTD provided a P-score to quantitatively assess the relative
populations of assembled conformations (Equation 2)). Finally, owing to the accessibility of the
modeling process, different sources of user-provided information, including cross-linking and cryo-
EM data, could be readily incorporated into the I-TASSER-MTD pipeline to improve the multi-
domain structure assembly.

Further applications of the I-TASSER-MTD protocol

In addition to its capacity as a powerful protein structure prediction server, the advancements
for multi-domain structure prediction by I-TASSER-MTD provide several additional useful
applications.
First, the domain definition and the corresponding domain models predicted by I-TASSER-MTD
can be used for protein family detection. The classification of evolutionary relationships of proteins is
crucial for protein structure and function studies. Several databases, such as the widely used SCOPe86
and CATH87 databases, have been developed to classify protein structures on the basis of individual
domains, which are considered to be the structural, functional and evolutionary units of proteins88.
As structure is more conserved than sequence89, almost all databases group protein domains in the
light of both structural similarity and sequential information. However, the domain definition and
classification for them are only performed on the basis of the solved proteins in the PDB. Therefore,
for proteins without solved structures, the domain boundaries and models predicted by I-TASSER-
MTD can be used to assist protein classification.
Second, I-TASSER-MTD supports cryo-EM data-assisted modeling, which can be used to help
experimental biologists construct protein structures from cryo-EM density maps. Cryo-EM has
become an indispensable method for determining structures of large proteins. For high-resolution
cryo-EM density maps, atomic structures can be constructed by programs traditionally used for X-ray
crystallography90, but these programs perform relatively poorly for medium-to-low-resolution den-
sity maps91. A common method for the structure modeling of these challenging density maps is to fit
a homologous structure into the density map, followed by atomic-level structural refinement.
However, the success of the approach highly depends on the quality of the initial models, while many
proteins, especially multi-domain proteins, have no homologous proteins with previously solved
structures. For these cases, I-TASSER-MTD can be employed for modeling as it constructs full-length
models by assembling independently predicted domain models, without requiring homologous
full-length structures.
Finally, an important aspect of I-TASSER-MTD is the annotation of protein functions at both the
domain- and full-chain level based on predicted structures using its integrated COFACTOR algo-
rithm. COFACTOR was shown to be accurate and scalable for full proteome-scale structure-based
annotation of microbes92 and higher organisms93 alike. For this purpose, the I-TASSER/COFACTOR
pipeline was incorporated into the neXtProt database for automated human protein function mod-
eling94. It was also applied to the JCVI-syn3.0 minimal genome where it found that a substantial
number of previously unannotated proteins are putative vitamin transporters92. In this regard,
COFACTOR predictions can be useful for function hypothesis generation when planning low-
throughput experiments. For example, the COFACTOR GO prediction was recently used to guide the
characterization of C9orf72, a guanine nucleotide exchange factor, which lead to a better under-
standing of its molecular role in amyotrophic lateral sclerosis95. Similarly, the ligand-binding
prediction of COFACTOR was used for identifying heme-binding sites in HemJ, the poorly studied
Protoporphyrinogen IX oxidase in cyanobacteria96.

Limitations
The domain boundary prediction method employed in the I-TASSER-MTD server for Hard targets
with template alignment coverage ≤95% is based on the deep-learning predicted contact maps that

NATURE PROTOCOLS | VOL 17 | OCTOBER 2022 | 2326–2353 | www.nature.com/nprot 2337

 PROTOCOL NATURE PROTOCOLS

require MSA collection. For extremely large proteins (>2,000 residues), contact map prediction and
MSA collection will require a high amount of random access memory (RAM) that the current
computing server cannot provide for some cases, which will result in failed domain partitioning by
FUpred due to memory limit. However, this limitation can be overcome by using ThreaDom for all
targets with sequence lengths >2,000 residues. Users can download the FUpred standalone package
and run it locally to predict the domain boundaries for a query sequence if they want to use the
FUpred-predicted domain definition for these cases. Furthermore, users also can provide the domain
definition predicted by other external programs such as NCBI Conserved Domain Database or
PFAM. This will also speed up the I-TASSER-MTD structure modeling process as the server will not
take time to predict the domain boundaries. See the ‘Experimental design’ section for detailed
instructions on how to provide the domain definition.
One highlight of the I-TASSER-MTD server is that it independently creates the model of each
individual domain and assembles all domain models into the full-length model. However, the quality
of the final full-length model is dependent on the accuracy of the individual domain models.
Although the domain assembly process and scoring function can accommodate some degree of
structural uncertainty, an incorrect domain model (e.g., TM-score <0.5) may affect the full-length
template identification based on structural alignment and misguide the domain assembly. This will
probably result in a poor final full-length model with a low eTM-score because each model is
considered as a rigid body during the domain assembly. For cases with low eTM-scores (e.g., <0.5),
users are advised to provide other sources of structural information, such as cross-linking experi-
mental data, restraints from mutagenesis, high-confidence full-length templates or inter-domain
contacts/distances determined by alternative programs, to guide the domain assembly.

Materials
Equipment
● A personal computer with Internet connection and a web browser with JavaScript enabled
(the I-TASSER-MTD server is compatible with popular web browsers, including Google Chrome,
Firefox, Microsoft Edge and Safari)
CRITICAL The amino acid sequence should be in
c

● The amino acid sequence of the protein of interest

FASTA format, in which only characters from the single-letter code of the 20 standard amino acids are
allowed. Spaces, line breaks and header lines starting with ‘>’ will be ignored and will not affect the
prediction.

Software
● A web browser such as Google Chrome, Firefox, Microsoft Edge or Safari
● (Optional) A molecular visualizing software, such as Jmol, RasMol or PYMOL, for viewing the 3D
structure of the modeled protein and the predicted functional sites locally

Procedure
Query sequence submission ● Timing 5 min
1 Navigate to the I-TASSER-MTD website at https://zhanggroup.org/I-TASSER-MTD/.
2 Provide the amino acid sequence by copying and pasting the sequence into the provided form or
directly uploading a plain text file containing the sequence.
CRITICAL STEP The sequence should contain only one chain. If the provided sequence includes
c

multiple chains, only the first chain will be used. At present, the I-TASSER-MTD server accepts
protein sequences with a length between 30 and 3,000 amino acids.
3 Input an email address in the text box to receive the results when the job is completed.
CRITICAL STEP It is crucial to provide a correct email address. Otherwise, the user will not be
c

able to receive any notifications or the results.

4 (Optional) Enter a name for the protein. The protein will be named ‘query protein’ if no name is
provided.
5 (Optional) Provide the domain definition. Enter the domain definition in the corresponding form.
The maximum length of a sequence accepted by the server will be increased to 3,000 if the domain
definition is provided. Read about specifying the domain definition in the ‘Experimental design’
section or click the question mark to read the explanation.
? TROUBLESHOOTING

2338 NATURE PROTOCOLS | VOL 17 | OCTOBER 2022 | 2326–2353 | www.nature.com/nprot

 NATURE PROTOCOLS PROTOCOL
6 (Optional) Upload the full-length templates to guide the assembly. Put all templates (each template
in one independent PDB file) in one compressed file (*.tar.gz, *.zip, *.tar or *.tar.bz2), and upload
it by clicking the corresponding button. Read about specifying the full-length templates in the
‘Experimental design’ section of this protocol or click the question mark to read the explanation.
! CAUTION The server uses a maximum of 20 templates. If users provide more than 20 templates,
only the first 20 templates will be utilized during the domain assembly simulations.
? TROUBLESHOOTING
7 (Optional) Exclude some templates from the library for both individual domain modeling and
domain model assembly. Users are advised to keep the default option to use all templates. In rare
scenarios for benchmarking purposes, users can choose ‘remove templates sharing >30% sequence
identity with target’ to remove all the templates that are highly homologous to the target sequence.
! CAUTION Users are advised to keep the default selection of using all templates during the
modeling. In general, excluding homologous templates will make structure prediction harder.
Therefore, this option is only for benchmarking purposes.
8 (Optional) Decide whether or not to predict the protein function at the domain and full-length
level. If users need to analyze the function of the protein, choose ‘YES’ to predict the function.
Otherwise, keep the default option.
! CAUTION Selecting the option to predict the protein function will greatly increase the job runtime
because the protein function prediction cannot be performed until the model is generated.
9 (Optional) Provide experimental data to assist the domain assembly. Prepare the cross-linking data
as described in the ‘Experimental design’ section, and select the corresponding box to paste the data
in the form or directly upload the file containing the data. To provide cryo-EM data, select the
corresponding box to show the ‘upload’ button, and upload the cryo-EM density map (in MRC or
CCP4 format) by clicking the button. Users should also enter the resolution of the density map in
the given text box.
CRITICAL STEP As we showed using the DEMO benchmark set, cross-linking data and cryo-EM
c

density map will significantly improve the quality of the final full-length model16,77. In addition,
cross-linking data can also be replaced by contact or distance restraints determined by any contact
or distance prediction program.
? TROUBLESHOOTING
10 Click the ‘Run I-TASSER-MTD’ button to submit the job.
? TROUBLESHOOTING

Job monitoring ● Timing 6–12 h

11 Monitor the job status. Once the job is submitted, the browser will be directed to a new page
displaying a confirmation of the length of the submitted sequence, a job identification number and
an estimated time to complete the job. The page is automatically refreshed every 10 s, and the
results will be shown in this page when the job is finished. Users may choose to bookmark this link
to retrieve the results. Meanwhile, users will receive an email confirmation containing a link to the
page when the job is successfully submitted.
PAUSE POINT Once the sequence is successfully submitted, it will be put in a queue until all
j

other jobs before it are processed on the computer cluster. Users may choose to close the job status
page. When the prediction is done, an email notification containing the link to the results page will
be sent to the user. The results can be accessed through this link or the bookmarked page.
12 Click the link in the email notification or open the link bookmarked in Step 11 to visit the page
containing the results. The page starts with a title and a link to download the tarball file including
all results listed on the page (Fig. 3a). An example results page is available at https://zhanggroup.
org/I-TASSER-MTD/example/.
! CAUTION The results will be stored on the server for 1 month. Users are recommended to
download the results to their computers.

Query sequence and predicted domain definition ● Timing 3 min

13 View the first section of the results page to check the submitted amino acid sequence and the
predicted domain definitions (Fig. 3b). The sequence of each predicted domain is represented by
different colors in the full-length query sequence. The illustration of the color and the sequence
range for each domain are listed below the query sequence. See Box 1 to follow the representation of
the domain range.

NATURE PROTOCOLS | VOL 17 | OCTOBER 2022 | 2326–2353 | www.nature.com/nprot 2339

 PROTOCOL NATURE PROTOCOLS

a
1 I-TASSER-MTD Results for job ITM118

2 [Click on ITM118_results.tar.bz2 to download the tarball file including all results listed on this page]

3
4

Fig. 3 | Example of the I-TASSER-MTD results page (Sections 2 and 3). a, Title of the results page and the link to download all results shown in the
page. b, Query sequence in FASTA format submitted by the user, where the predicted domains are marked by different colors, and the range of each
domain is listed below the sequence. c, Top final full-length models predicted by the server and their estimated accuracy (right). Model 1 is shown
(left), where different domains are represented by different colors. d, Predicted individual domain models that are used to assemble the full-length
model and the estimated distance error of each domain model.

2340 NATURE PROTOCOLS | VOL 17 | OCTOBER 2022 | 2326–2353 | www.nature.com/nprot

 NATURE PROTOCOLS PROTOCOL
Full-length structure prediction ● Timing 10 min
14 View the second section of the results page to analyze the full-length structure predictions (Fig. 3c).
The default structure shown in the JSmol applet97 is the first model, where different domain models
are marked by different colors. Users can rotate and zoom the model by dragging the mouse on the
image. The table to the right of the image summarizes the top five predicted models. In the first
column, click the circle in front of the model name to display the corresponding full-length model
in the JSmol applet.
CRITICAL STEP If only one full-length model is reported in this section, it indicates that the

c
model was directly generated by D-I-TASSER as the threading templates could cover all domains,
and the top templates identified by LOMETS2 had consistent topologies. In these cases, the final
model usually has a relative high eTM-score, indicating a high-quality final model.
15 View the second column of the table to analyze the eTM-score for each full-length model. As
defined in Equation. 1, the eTM-score is calculated on the basis of the confidence of individual
domain models and the confidence of the inter-domain assembly simulations. The eTM-score
usually ranges from 0 to 1, where a higher score indicates a model of better quality. In general,
models with eTM-score >0.5 have a correct global fold.
? TROUBLESHOOTING
16 View the eRMSD to the native structure shown in the third column of the table. As defined in
Eq. (S14), eRMSD is estimated in a similar way as the eTM-score but with the sequence length
incorporated.
! CAUTION Since the top five models are ranked by energy or by cluster size, in rare cases it is
possible that the lower-rank models have a higher eTM-score or lower eRMSD. Accordingly,
although the first model has a better quality on average, it is also possible that the lower-rank
models have a better quality than the higher-rank models as seen in our benchmark tests. Users can
also estimate the total local distance difference test (lDDT) of the model for reference by using
deep-learning based quality assessment methods, such as DeepAccNet98 and DeepUMQA99.
17 View the P-score reported in the fourth column of the table. Here, P-score is used to assess the
relative populations of complex conformations under the assumption that the relative populations
of the complex conformations are approximately proportional to their entropy variations in the
domain structural assembly simulations. The P-score value ranges between [0, 1], and a higher
value indicates the structure occurs more often in the simulation trajectory.
18 Download the predicted models in PDB format by clicking on the ‘Download model’ link shown in
the corresponding row of the column marked as ‘PDB file’. Users can interactively view the
predicted structure on their computer using the programs mentioned in the ‘Materials’ section.
19 View the predicted probability of the inter-domain interaction for every two domains, which is
listed under the table (see label 1 in Fig. 3c). Here, an inter-domain interaction is defined as one or
more residue pairs with distance <8 Å apart from the linker region. The domain interaction
probability is estimated by the average probability score of the top 1% of the inter-domain residue
pairs predicted by the deep-learning models. It ranges from 0 to 1, with a higher value indicating
that the two domains have a larger probability of interaction.
20 Click the link labeled ‘More about eTM-score’ (see label 2 in Fig. 3c) to open a new page containing
more information about the eTM-score and eRMSD.

Individual domain modeling ● Timing 5 min

21 Scroll down further to see the individual domain models that are independently created by
D-I-TASSER (Fig. 3d) and used to assemble the full-length models shown in Fig. 3c. The 3D model
of each domain is independently displayed using the JSmol applet.
22 Download the PDB file of the predicted individual domain structure by clicking on the ‘download
domX.pdb’ link below the image of each individual domain model.
23 View the eTM-score of the domain modeling shown below the download link for each domain
model (see label 1 in Fig. 3c), and the distance error to native (in angstrom) for each residue
estimated by ResQ75 which is displayed in the chart (see label 2 in Fig. 3c) under the structure.
Users can view the estimated error of each residue by moving the mouse on the chart and download
the file summarizing the predicted distance error of all residues by clicking on the ‘Download
distance error of domain X’ link (see label 3 in Fig. 3c) below the chart.
24 View the predicted function of each individual domain. If the query protein is predicted to be a
multi-domain protein and the user chose the option for function prediction, the predicted function

NATURE PROTOCOLS | VOL 17 | OCTOBER 2022 | 2326–2353 | www.nature.com/nprot 2341

 PROTOCOL NATURE PROTOCOLS

1 2 3

Fig. 4 | Example of the I-TASSER-MTD results page (Sections 4–6). a, Predicted secondary structure of the full-length sequence, where different
types of secondary structure are marked by different colors. b, Predicted solvent accessibility of the full-length sequence, where a larger confidence
score indicates a higher probability of exposure. c, Results of the domain boundary prediction, which includes the contact map used to guide the
domain boundary prediction (1), the curve of the FU-score for continuous domain detection (2), the heatmap of FU-score for discontinuous domain
detection (3) and the predicted domain definitions (4).

of each domain can be viewed by clicking the link labeled ‘Click to view the predicted function’
below the corresponding predicted domain distance error chart (see label 4 in Fig. 3c). The function
results page will be interpreted in Steps 38–52.

Secondary structure prediction ● Timing 2 min

25 View the next section to analyze the predicted secondary structure of the full-length sequence
(Fig. 4a). The results contain three typical states: helix (H), β-strand (S) and loop/coil (C), which are
determined according to the confidence scores of each residue shown in the next row of the
predicted results. The confidence score ranges from 0 to 9, where a higher value represents a state
with higher confidence. The predicted secondary structure results are employed to guide domain
parsing and domain modeling.

Solvent accessibility prediction ● Timing 2 min

26 Scroll down further to see the section for the predicted solvent accessibility of the
full-length sequence (Fig. 4b). For each residue, a confidence score ranging from 0 (buried
residue) to 8 (highly exposed residue) is shown below the sequence. These results are used by the
threading methods during the individual domain modeling in D-I-TASSER to guide the template
identification.

Domain boundary prediction ● Timing 5 min

27 View the next section of the results page to analyze the predicted domain boundaries (Fig. 4c). If the
user does not provide the domain definition, the results of the domain boundary prediction will be

2342 NATURE PROTOCOLS | VOL 17 | OCTOBER 2022 | 2326–2353 | www.nature.com/nprot

 NATURE PROTOCOLS PROTOCOL
reported in four columns. The first column displays the contact map applied to guide the prediction
of domain boundaries (see label 1 in Fig. 4c), according to which different domains (or segments of
the discontinuous domain) are separated by red lines. Each domain or segment is marked by a
name on the contact map. For example, ‘D1’ indicates the first domain (continuous). If a ‘D1-1’ is
marked on the figure (Supplementary Fig. 14), it indicates that the second domain is a
discontinuous domain, and ‘D1-1’ represents its first segment (Box 1). Users can click ‘Download
contact map’ to download the contact data for further study.
! CAUTION If the domain boundary is predicted by ThreaDom, only two columns will be shown in
this section (Supplementary Fig. 15), where the first column is the curve of the DCS, and the second
column is the predicted domain definition.
28 View the second column of the domain boundary prediction results (see label 2 in Fig. 4c). The
figure shows the FU-score curve of all residues for the continuous domain detection, on which
vertical red lines correspond to the indices of the predicted domain boundary residues with
FU-scores below the cutoff (the horizontal dotted line). Users can download the FU-score file by
clicking ‘Download the FU-score (continuous)’.
! CAUTION For some cases, the domain boundary does not correspond to the residue index with
the lowest FU-score. This is because the predicted boundary is located within a strand or a helix,
and so the boundary is shifted to the coil region as domain boundaries occur mainly at loop regions
rather than on regular secondary structures.
29 View the FU-score heatmap for discontinuous domains shown in the third column of the domain
boundary prediction results (see label 3 in Fig. 4c). The figure is generated according to the
FU-score matrix for the discontinuous domain detection, where the colors ranging from blue to red
indicate low to high scores, and the black dotted lines represent the predicted continuous domain
boundaries. Users can download the FU-score matrix by clicking ‘Download the FU-score
(discontinuous)’.
30 View the predicted domains reported in the fourth column of the domain boundary prediction
results (see label 4 in Fig. 4c), where each domain is written in one line. The ‘Modeling’ and
‘Without linker’ parts refer to the domain definitions used for the domain structure modeling and
the domain boundary definition without including the linker, respectively.

Full-length templates for domain assembly ● Timing 5 min

31 View the next section of the results page to analyze the top ten full-length templates identified by
the TM-align-based structural alignments on the basis of the domain models (Fig. 5a). These
templates are employed to construct the initial full-length model, and the structurally aligned
regions of these templates are used to deduce the inter-domain distance profiles to guide the
domain assembly simulations.
32 View the sequence identity between the query and the template proteins (column SeqId), where a
higher value indicates a strong evolutionary relationship between the query and template proteins.
33 View the score of the full-length template (column ‘TplScore’), which is the harmonic mean of the
TM-scores between the domain models and the template. The higher the score, the higher the
structural similarity between the template and the query protein.
! CAUTION As described in the section entitled ‘The I-TASSER-MTD pipeline’, the full-length
model will be directly created by D-I-TASSER if the LOMETS2 threading templates cover all
domains. Therefore, in this case, the templates shown here will be those identified by LOMETS2
threading, and the TplScore will be the harmonic mean of the TM-scores between all domain
models and the LOMETS2 template.
34 View the structural alignments to identify similar regions in the query and the template proteins.
The query sequence is colored by domains, and the aligned residues in the template that are
identical to the corresponding query residues are colored on the basis of their amino acid properties
in the alignment.
35 Click on the PDB code and chain identifier for the templates in the ‘Template’ column. The browser
will be directed to the Research Collaboratory for Structural Bioinformatics website showing
information about the template protein.

Predicted distance/interface map for domain assembly ● Timing 5 min

36 Scroll down to see the predicted distance/interface map section for the domain model assembly
(Fig. 5b). The first column and the second column show the deep-learning predicted distance maps

NATURE PROTOCOLS | VOL 17 | OCTOBER 2022 | 2326–2353 | www.nature.com/nprot 2343

 PROTOCOL NATURE PROTOCOLS

Fig. 5 | Example of the I-TASSER-MTD results page (Sections 7 and 8). a, Top ten full-length templates identified by the global structural alignment,
which are used to guide the domain model assembly. b, Predicted residue–residue distance maps and domain–domain interface maps for domain
model assembly.

with Cα and Cβ distances <20 Å, respectively. The subsequent two columns show the interface maps
with Cα and Cβ distances <18 Å predicted by deep learning, where all residue pairs are included in
the map. In these maps, only the inter-domain distances and inter-domain interface probabilities
are employed to guide the domain model assembly.
? TROUBLESHOOTING
37 Click the link below the figure to download the corresponding predicted distance map for further
analysis. For example, click the link labeled ‘Download CA distance map’ to download the predicted
distance map with Cα distances <20 Å.

Analogous proteins of the predicted full-length model ● Timing 5 min

38 View the structurally analogous proteins of the predicted top full-length model, which is
identified from the PDB by TM-align structural alignment (Fig. 6a). The table to the right of the
model summarizes the analogous proteins, where the proteins are ranked according to the
TM-score (shown in the fourth column) between the predicted full-length model and the TM-align
detected templates. The analogous proteins with a TM-score >0.5 can be used to determine the
structural family of the query protein100.
39 View the ‘RMSD’, ‘IDEN’ and ‘Cov.’ columns in the table to analyze the RMSD, sequence identity
and coverage of the aligned regions determined by TM-align, which indicates the conservation of
spatial motifs in the model and the structurally analogous proteins.
CRITICAL STEP Users are recommended to read the explanation of each criterion, which is
c

provided below the table.

40 Click the link in the last column of ‘Download Alignment’ to download the PDB file with the top
model structurally aligned to the corresponding analogous protein.

2344 NATURE PROTOCOLS | VOL 17 | OCTOBER 2022 | 2326–2353 | www.nature.com/nprot

 NATURE PROTOCOLS PROTOCOL
a b Predicted Gene Ontology (GO) Terms

c
2

d 3

Fig. 6 | Example of the I-TASSER-MTD results page (Sections 9–12). a, Top ten analogous structures that are structurally close to the top model of
the query protein. b, Results of the predicted GO terms including MF (1), BP (2) and CC (3). c, Results of the predicted EC numbers from the top five
homologous enzyme templates. d, Results of the predicted ligand-binding site from the top five homologous templates.

GO term prediction ● Timing 5 min

41 View the GO terms predicted according to the I-TASSER-MTD full-length model, full-chain
sequence and PPI networks (Fig. 6b). The predicted results include three subsections: molecular
function (MF) (Figs. 6b-1), biological process (BP) (Figs. 6b-2) and cellular component (CC)
(Figs. 6b-3).
42 For each GO aspect, view the directed acyclic graph by Graphviz101 shown in the left panel of the
corresponding subsection, which is plotted by combining the predicted GO terms and their parent
terms. On the graph, the predicted GO terms are colored by the confidence score CscoreGO
(Supplementary Note 6), where the color code for the range of CscoreGO is illustrated in the bottom
of the right panel. The CscoreGO values range between 0 and 1, and a higher value usually indicates
a better confidence in predicting the function using the template. The precision of each range of
CscoreGO is shown in Supplementary Fig. 16a. For example, BP has a 71% probability of being
correctly predicted if CscoreGO >0.9.
43 View the table for the summary of predicted results shown on the right of the graph.
The three columns of the table show the predicted GO terms, the confidence score and the
corresponding common name of the GO term. Click on each GO term in the first column to

NATURE PROTOCOLS | VOL 17 | OCTOBER 2022 | 2326–2353 | www.nature.com/nprot 2345

 PROTOCOL NATURE PROTOCOLS

visit the Amigo website (http://amigo.geneontology.org/amigo) to analyze the definition and lineage
of the term.
44 Click on the link ‘full result’ below the table to download the results presented in the table.

EC number prediction ● Timing 5 min

45 Scroll down further to analyze the top five EC number predictions (Fig. 6c). The left panel shows
the first I-TASSER-MTD model, as well as the predicted active sites in case that they exists, in an
accompanying JSmol applet.
46 View the table in the right panel, which summarizes the results of the predictions. In the table, the
EC predictions are ranked on the basis of the CscoreEC (confidence score of EC, Supplementary
Note 6) reported in the third column. Here, CscoreEC values range between 0 and 1, where a higher
score indicates a more reliable EC number prediction The relationship between the precision and
CscoreEC is reported in Supplementary Fig. 16c. For example, the CscoreEC in the range of [0.6, 0.7]
indicates that the first digit, first two digits, first three digits and all four digits of the EC number
have probabilities of 76%, 66%, 63% and 44% of being correctly predicted, respectively. Users can
click the circle in the first column to see the corresponding predicted active sites in the left
JSmol panel.
47 View the TM-score between the predicted model and the enzyme analogs (column ‘PDB Hit’),
RMSD of aligned regions, sequence identity, coverage of aligned regions, predicted EC number and
predicted active sites in the table.
CRITICAL STEP Although the CscoreEC is used to rank the predicted EC number,
c

users are advised to consult both the CscoreEC and the TM-score. For example, if most
of the identified functional analogs with similar folds (i.e., TM-score >0.5) have the same EC
number digits and the CscoreEC is relatively high, the likelihood of the prediction being correct is
very high.
48 Click on the EC numbers to visit the ExPASy enzyme database (https://enzyme.expasy.org/) to
further learn about the enzyme families, such as reactions catalyzed by the enzyme, the cofactors
required and the metabolic pathway in which they function.

Ligand-binding site prediction ● Timing 5 min

49 In case that there exist predicted ligand-binding sites, view the last section of the results page to
analyze the best predicted ligand-binding sites in the predicted full-length model (Fig. 6d). The full-
length model, together with the predicted binding site residues and the corresponding ligand, is
displayed in the left panel using the JSmol applet. The binding site residues of the query protein are
highlighted as ‘ball and stick’ with the corresponding residue numbers labeled in magenta, whereas
ligand atoms are represented by colored spheres.
50 View the table reporting the top identified functional analogs and the derived binding site residues
in the right panel. The predicted binding site residues are ranked according to the confidence score
CscoreLB (confidence score of ligand-binding site, Supplementary Note 6) listed in the third column
of the table. The CscoreLB value ranges from 0 to 1, with a higher value corresponding to a more
reliable ligand-binding site. The precision versus CscoreLB is reported in Supplementary Fig. 16b to
show how reliable the results are. For example, CscoreLB >0.7 indicates the ligand-binding site has
>82% probability of being correctly predicted. Users can click on the PDB links to track the bound
ligands in the functional analogs.
51 View the TM-score, RMSD, sequence identity and coverage of the aligned regions between the
query model and the identified functional analogs, as well as the binding site score and predicted
binding site residues in the table. The binding site score is a measure of local similarity calculated
according to the local structural similarity of the ligand-binding sites and the sequence identity
between the I-TASSER-MTD model and the structural analogs. A binding site score >1 indicates a
outstanding local match between the predicted and template binding sites based on a large-scale
benchmark study45,70.
52 Download the ligand–protein complex. Users can click on the link listed in the column ‘Download
Complex’ to download the PDB file containing the predicted model and the bound ligand. The file
can be viewed interactively using any molecular visualization tool. Rendering the ligand as colored
spheres and depicting predicted binding site residues as ‘ball and stick’ will aid in visualizing the
binding site cleft and help with analyzing the ligand–protein interactions.

2346 NATURE PROTOCOLS | VOL 17 | OCTOBER 2022 | 2326–2353 | www.nature.com/nprot

 NATURE PROTOCOLS PROTOCOL
Troubleshooting
Troubleshooting advice can be found in Table 2.

Table 2 | Troubleshooting table

Step Error message or problem Possible reason Solution

5 Error! Some residues are missed in Some residue indices are not included in the Check the domain definition to make sure
the domain definition domain definition that the domain definition includes all residue
indices
Error! Residue overlaps in the Some residue indices are included in multiple Check the domain definition to make sure
domain definition domain ranges each residue index is included in only
one domain
Error! Wrong format of the The domain definition cannot be recognized Read the instructions on how to correct the
domain definition by the server format of the domain definition or follow the
formatting shown in the example input
Error! Too many domains in the Maximum number of domains accepted by the Merge short domains in the domain
domain definition server is 20 definition or model the first 20 domains and
the rest of the domains independently using
the server, and assemble them using DEMO
Error! Too large domain in the Maximum length of the domain accepted by Further split large domains into multiple parts
domain definition the server is 1,000
Error! Too short domain in the Minimum length of the domain accepted by Merge short domains with other domains
domain definition the server is 30
6 Error! Incorrect template name The template name cannot be recognized by Rename the template according to the
the server instructions
Error! Wrong tarball format (XXX) The tarball format is not supported by Repackage the templates as *.tar.bz2,
of the templates the server *.tar.gz, *.tar or *.zip
9 Error! Wrong cross-linking data The cross-linking data cannot be read by Check the instructions or the example input
the server to ensure the cross-linking data is in the
correct format. Ensure that the residue index
is within the sequence length range
Error! Wrong density map file The cryo-EM density map data cannot be read Make sure that the density map is in the 8-
by the server bit, 16-bit, or 32-bit MRC and CCP4 format,
and fix the error according to the instructions
10 The protein sequence is too short The range of sequence length is not within Check the sequence to make sure that the
or too long [30, 2000] length is within [30, 2000]. Users who want
to model larger proteins with sequence
lengths >2,000 residues should specify their
own domain boundary
15 Low eTM-score full-length model Low-quality individual domain models, As described in the ‘Limitations’ section,
full-length templates or distance map users are advised to seek other sources of
structural information, such as experimental
data, full-length templates and contact/
distance restraints predicted by other tools
35 No predicted distance maps The sequence length is large, which causes a Users can download the standalone package
larger distance map than can be predicted by and predict the distance locally
the server owing to RAM limitations

Timing
Steps 1–10, query sequence submission: 5 min
Steps 11–12, job monitoring: 6–12 h
Step 13, query sequence and predicted domain definition: 3 min
Steps 14–20, predicted full-length structures: 10 min
Steps 21–24, individual domain modeling: 5 min
Step 25, secondary structure prediction: 2 min
Step 26, solvent accessibility prediction: 2 min
Steps 27–30, domain boundary prediction: 5 min

NATURE PROTOCOLS | VOL 17 | OCTOBER 2022 | 2326–2353 | www.nature.com/nprot 2347

 PROTOCOL NATURE PROTOCOLS

Steps 31–35, full-length templates for domain assembly: 5 min

Steps 36–37, predicted distance/interface map for domain assembly: 5 min
Steps 38–40, analogous proteins of the predicted full-length model: 5 min
Steps 41–44, GO term prediction: 5 min
Steps 45–48, EC number prediction: 5 min
Steps 49–52, ligand-binding site prediction: 5 min
Prediction of the 3D structure and function for a medium-size multi-domain protein (~300–500
residues) requires 6–10 h using the I-TASSER-MTD server. Longer proteins with more domains will
take more time as the server independently models the structure of each domain, and the domain
assembly can be performed only when all individual domain models are generated. The actual pro-
cessing time, however, also depends on the number of jobs in the queue. Currently, the I-TASSER-MTD
server is working on a supercomputer cluster that consists of 59 20-core IBM NeXtScale nx360 M4
nodes running at 2.8 GHz, and users can receive the results within 1–3 d in most cases.

Anticipated results
Here we used the iron-dependent regulator from Mycobacterium tuberculosis (PDB ID: 1fx7A) as an
example to predict the structure and function of the protein using the I-TASSER-MTD server. In
Step 8, we chose the ‘Option IV’ as ‘YES’ to predict the function of all individual domains and the
full-length protein, while other options were kept as default. Once the job is finished, the user will
receive an email containing a link to the results page. Clicking on the link will open the results page
containing the following results, which are shown in Figs. 3–6:
● The title of the results page: I-TASSER-MTD results for job jobid (see label 1 in Fig. 3a).

● The link to download the results files. Click on the link to download a compressed file containing all

results shown on the results page (see label 2 in Fig. 3a).

● The user submitted sequence in FASTA format with predicted domains marked by different colors

(Fig. 3b). Users can confirm the sequence length shown in the parentheses after the query name.
● The top five full-length models predicted by I-TASSER-MTD (Fig. 3c). The left panel shows the

predicted model using the JSmol applet, while the right panel summarizes the information of the
predicted models.
● The predicted individual domain models (Fig. 3d). Each domain model is shown in an independent

JSmol applet, with a link to download the model, and the predicted functions are given below the model.
The predicted functions for each individual domain are shown in Supplementary Figs. 17–19.
● The predicted secondary structure of the full-length sequence (Fig. 4a). The results contain three rows.

The first row shows the query sequence, while the second and the third row report the predicted
secondary structure and the confidence score, respectively.
● The predicted solvent accessibility of the full-length sequence, which contains two rows (Fig. 4b). The first

row and the second row display the query sequence and the confidence score for the predicted solvent
accessibility, respectively.
● The predicted domain boundary (Fig. 4c). This section contains the deep-learning-predicted contact map

for the domain boundary prediction, the FU-score curve for the continuous domain detection, the
FU-score heatmap for the discontinuous domain detection and the predicted domain definition.
● The top ten full-length templates identified by TM-align structural alignment (Fig. 5a). This section

reports the sequence identity, template score and the alignment between the query and the template.
● The distance/interface map predicted by deep learning (Fig. 5b). The C /C distance maps with distances
α β
<20 Å, the Cα/Cβ interface maps with distances <18 Å and the corresponding link to download the
distance/interface maps are shown in this section.
● The top ten proteins structurally close to the query protein (Fig. 6a). The predicted full-length models

with the aligned analogous models are depicted in the left panel using the JSmol applet, while the
information for the structural analogs is shown in the right panel.
● The predicted GO terms (Fig. 6b). This section contains the predicted MF, BP and CC. For each section,

the predicted GO terms are plotted using a directed acyclic graph displayed in the left panel, while a table
with the predicted GO terms is shown on the right panel.
● The predicted EC numbers (Fig. 6c). The JSmol panel on the left shows the full-length model and the

predicted active site residues. The table on the right summarizes the predicted active sites.
● The predicted ligand-binding sites (Fig. 6d). The full-length model, predicted binding sites and

corresponding ligands are displayed on the left panel using the JSmol applet, and a summary of the
predicted ligand-binding sites is reported in a table on the left panel.

2348 NATURE PROTOCOLS | VOL 17 | OCTOBER 2022 | 2326–2353 | www.nature.com/nprot

 NATURE PROTOCOLS PROTOCOL
It should be noted that only the domain definition will be shown in the predicted domain
boundary section if the user provides the domain definition. In addition, the link to the predicted
functions of each individual domain will not be generated if the query protein is predicted as a single-
domain protein or the user did not choose the option for function prediction. Proteins structurally
close to the query protein, predicted GO terms, EC numbers and ligand-binding sites for the full-
length protein will not be reported in the results page if the user kept the default option to not include
protein function prediction.

Data availability
The raw data and example files are available at https://zhanggroup.org/I-TASSER-MTD/ or from the
corresponding author upon reasonable request.

Code availability
The I-TASSER-MTD standalone package is freely available for academic use at https://zhanggroup.
org/I-TASSER-MTD/.

References
1. Sali, A. & Blundell, T. L. Comparative protein modelling by satisfaction of spatial restraints. J. Mol. Biol.
234, 779–815 (1993).
2. Simons, K. T., Kooperberg, C., Huang, E. & Baker, D. Assembly of protein tertiary structures from
fragments with similar local sequences using simulated annealing and Bayesian scoring functions. J. Mol.
Biol. 268, 209–225 (1997).
3. Xu, D. & Zhang, Y. Ab initio protein structure assembly using continuous structure fragments and opti-
mized knowledge-based force field. Proteins 80, 1715–1735 (2012).
4. Yang, J. et al. The I-TASSER Suite: protein structure and function prediction. Nat. Methods 12, 7–8 (2015).
5. Weigt, M., White, R. A., Szurmant, H., Hoch, J. A. & Hwa, T. Identification of direct residue contacts in
protein–protein interaction by message passing. Proc. Natl Acad. Sci. USA 106, 67–72 (2009).
6. Marks, D. S. et al. Protein 3D structure computed from evolutionary sequence variation. PLoS ONE 6,
e28766 (2011).
7. Mortuza, S. et al. Improving fragment-based ab initio protein structure assembly using low-accuracy
contact-map predictions. Nat. Commun. 12, 5011 (2021).
8. Wang, S., Sun, S., Li, Z., Zhang, R. & Xu, J. Accurate de novo prediction of protein contact map by ultra-
deep learning model. PLoS Comput. Biol. 13, e1005324 (2017).
9. Senior, A. W. et al. Improved protein structure prediction using potentials from deep learning. Nature 577,
706–710 (2020).
10. Li, Y., Hu, J., Zhang, C., Yu, D.-J. & Zhang, Y. ResPRE: high-accuracy protein contact prediction by
coupling precision matrix with deep residual neural networks. Bioinformatics 35, 4647–4655 (2019).
11. Jumper, J. et al. Highly accurate protein structure prediction with AlphaFold. Nature 596, 583–589 (2021).
12. Kryshtafovych, A., Schwede, T., Topf, M., Fidelis, K. & Moult, J. Critical assessment of methods of protein
structure prediction (CASP)—Round XIV. Proteins 89, 1607–1617 (2021).
13. Chothia, C., Gough, J., Vogel, C. & Teichmann, S. A. Evolution of the protein repertoire. Science 300,
1701–1703 (2003).
14. Apic, G., Huber, W. & Teichmann, S. A. Multi-domain protein families and domain pairs: comparison with
known structures and a random model of domain recombination. J. Struct. Funct. Genomics 4,
67–78 (2003).
15. Han, J.-H., Batey, S., Nickson, A. A., Teichmann, S. A. & Clarke, J. J. N. R. M. C. B. The folding and
evolution of multidomain proteins. Nat. Rev. Mol. Cell Biol. 8, 319 (2007).
16. Zhou, X. G., Hu, J., Zhang, C. X., Zhang, G. J. & Zhang, Y. Assembling multidomain protein structures
through analogous global structural alignments. Proc. Natl Acad. Sci. USA 116, 15930–15938 (2019).
17. Xu, D., Jaroszewski, L., Li, Z. & Godzik, A. AIDA: ab initio domain assembly for automated
multi-domain protein structure prediction and domain–domain interaction prediction. Bioinformatics 31,
2098–2105 (2015).
18. Berman, H. M. et al. The Protein Data Bank. Nucleic Acids Res. 28, 235–242 (2000).
19. Zhang, Y. & Skolnick, J. TM-align: a protein structure alignment algorithm based on the TM-score. Nucleic
Acids Res. 33, 2302–2309 (2005).
20. Xue, Z., Xu, D., Wang, Y. & Zhang, Y. ThreaDom: extracting protein domain boundary information from
multiple threading alignments. Bioinformatics 29, i247–i256 (2013).
21. Hong, S. H., Joo, K. & Lee, J. ConDo: protein domain boundary prediction using coevolutionary infor-
mation. Bioinformatics 35, 2411–2417 (2019).
22. Zheng, W. et al. FUpred: detecting protein domains through deep-learning based contact map prediction.
Bioinformatics 36, 3749–3757 (2020).
23. Wollacott, A. M., Zanghellini, A., Murphy, P. & Baker, D. Prediction of structures of multidomain proteins
from structures of the individual domains. Protein Sci. 16, 165–175 (2007).

NATURE PROTOCOLS | VOL 17 | OCTOBER 2022 | 2326–2353 | www.nature.com/nprot 2349

 PROTOCOL NATURE PROTOCOLS

24. Zhang, C., Zheng, W., Freddolino, P. L. & Zhang, Y. MetaGO: predicting Gene Ontology of non-
homologous proteins through low-resolution protein structure prediction and protein–protein network
mapping. J. Mol. Biol. 430, 2256–2265 (2018).
25. Yao, S. et al. NetGO 2.0: improving large-scale protein function prediction with massive sequence, text,
domain, family and network information. Nucleic Acids Res. 49, W469–W475 (2021).
26. Piovesan, D. & Tosatto, S. C. INGA 2.0: improving protein function prediction for the dark proteome.
Nucleic Acids Res. 47, W373–W378 (2019).
27. Koo, D. C. E. & Bonneau, R. Towards region-specific propagation of protein functions. Bioinformatics 35,
1737–1744 (2019).
28. Gligorijević, V. et al. Structure-based protein function prediction using graph convolutional networks.
Nat. Commun. 12, 1–14 (2021).
29. Pearce, R. & Zhang, Y. Toward the solution of the protein structure prediction problem. J. Biol. Chem. 297,
100870 (2021).
30. Zheng, W. et al. Protein structure prediction using deep learning distance and hydrogen‐bonding restraints
in CASP14. Proteins 89, 1734–1751 (2021).
31. Zheng, W. et al. Deep-learning contact-map guided protein structure prediction in CASP13. Proteins 87,
1149–1164 (2019).
32. Battey, J. N. et al. Automated server predictions in CASP7. Proteins 69 (Suppl.), 68–82 (2007).
33. Croll, T. I., Sammito, M. D., Kryshtafovych, A. & Read, R. J. Evaluation of template-based modeling in
CASP13. Proteins 87, 1113–1127 (2019).
34. Zhang, Y. Template-based modeling and free modeling by I-TASSER in CASP7. Proteins 69
(Suppl.), 108–117 (2007).
35. Zhang, Y. I-TASSER: fully automated protein structure prediction in CASP8. Proteins 77 (Suppl.), 100–113
(2009).
36. Xu, D., Zhang, J., Roy, A. & Zhang, Y. Automated protein structure modeling in CASP9 by I-TASSER
pipeline combined with QUARK-based ab initio folding and FG-MD-based structure refinement. Proteins
79 (Suppl.), 147–160 (2011).
37. Zhang, Y. Interplay of I-TASSER and QUARK for template-based and ab initio protein structure prediction
in CASP10. Proteins 82 (Suppl.), 175–187 (2014).
38. Zhang, W. et al. Integration of QUARK and I-TASSER for ab initio protein structure prediction in CASP11.
Proteins 84 (Suppl.), 76–86 (2016).
39. Zhang, C., Mortuza, S. M., He, B., Wang, Y. & Zhang, Y. Template-based and free modeling of I-TASSER
and QUARK pipelines using predicted contact maps in CASP12. Proteins 86 (Suppl.), 136–151 (2018).
40. Roy, A., Kucukural, A. & Zhang, Y. I-TASSER: a unified platform for automated protein structure and
function prediction. Nat. Protoc. 5, 725–738 (2010).
41. He, K., Zhang, X., Ren, S. & Sun, J. Deep residual learning for image recognition. in Proceedings of the IEEE
conference on computer vision and pattern recognition 770–778 (2016).
42. Zheng, W. et al. LOMETS2: improved meta-threading server for fold-recognition and structure-based
function annotation for distant-homology proteins. Nucleic Acids Res. 47, W429–W436 (2019).
43. Wang, Y. et al. ThreaDomEx: a unified platform for predicting continuous and discontinuous protein
domains by multiple-threading and segment assembly. Nucleic Acids Res. 45, W400–W407 (2017).
44. Li, Y. et al. Protein inter‐residue contact and distance prediction by coupling complementary coevolution
features with deep residual networks in CASP14. Proteins 89, 1911–1921 (2021).
45. Zhang, C., Freddolino, P. L. & Zhang, Y. COFACTOR: improved protein function prediction by
combining structure, sequence and protein–protein interaction information. Nucleic Acids Res. 45,
W291–W299 (2017).
46. Sillitoe, I. et al. CATH: increased structural coverage of functional space. Nucleic Acids Res. 49, D266–D273
(2021).
47. Xu, Y., Xu, D. & Gabow, H. N. Protein domain decomposition using a graph-theoretic approach.
Bioinformatics 16, 1091–1104 (2000).
48. Steinegger, M. & Söding, J. Clustering huge protein sequence sets in linear time. Nat. Commun. 9, 2542 (2018).
49. Steinegger, M., Mirdita, M. & Söding, J. Protein-level assembly increases protein sequence recovery from
metagenomic samples manyfold. Nat. Methods 16, 603–606 (2019).
50. Mitchell, A. L. et al. MGnify: the microbiome analysis resource in 2020. Nucleic Acids Res. 48, D570–D578
(2020).
51. Chen, I.-M. A. et al. The IMG/M data management and analysis system v. 6.0: new tools and advanced
capabilities. Nucleic Acids Res. 49, D751–D763 (2021).
52. Mirdita, M. et al. Uniclust databases of clustered and deeply annotated protein sequences and alignments.
Nucleic Acids Res. 45, D170–D176 (2017).
53. Suzek, B. E. et al. UniRef clusters: a comprehensive and scalable alternative for improving sequence
similarity searches. Bioinformatics 31, 926–932 (2015).
54. Zhang, C., Zheng, W., Mortuza, S., Li, Y. & Zhang, Y. DeepMSA: constructing deep multiple sequence
alignment to improve contact prediction and fold-recognition for distant-homology proteins. Bioinformatics
36, 2105–2112 (2020).
55. Yan, R., Xu, D., Yang, J., Walker, S. & Zhang, Y. A comparative assessment and analysis of 20 representative
sequence alignment methods for protein structure prediction. Sci. Rep. 3, 1–9 (2013).

2350 NATURE PROTOCOLS | VOL 17 | OCTOBER 2022 | 2326–2353 | www.nature.com/nprot

 NATURE PROTOCOLS PROTOCOL
56. Ekeberg, M., Hartonen, T. & Aurell, E. Fast pseudolikelihood maximization for direct-coupling analysis
of protein structure from many homologous amino-acid sequences. J. Comput. Phys. 276, 341–356
(2014).
57. Yang, J. et al. Improved protein structure prediction using predicted interresidue orientations. Proc. Natl
Acad. Sci. USA 117, 1496–1503 (2020).
58. Thrun, S. in Advances in Neural Information Processing Systems 640–646 (Morgan Kaufmann Publishers, 1996).
59. Zheng, W., Zhang, C., Bell, E. W. & Zhang, Y. I-TASSER gateway: a protein structure and function
prediction server powered by XSEDE. Future Gener. Comput. Syst. 99, 73–85 (2019).
60. Zhang, Y. I-TASSER server for protein 3D structure prediction. BMC Bioinforma. 9, 40 (2008).
61. Zheng, W. et al. Folding non-homologous proteins by coupling deep-learning contact maps with I-TASSER
assembly simulations. Cell Rep. Methods 1, 100014 (2021).
62. Li, Y., Zhang, C., Bell, E. W., Yu, D. J. & Zhang, Y. Ensembling multiple raw coevolutionary features with
deep residual neural networks for contact-map prediction in CASP13. Proteins 87, 1082–1091 (2019).
63. Li, Y. et al. Deducing high-accuracy protein contact-maps from a triplet of coevolutionary matrices through
deep residual convolutional networks. PLOS Comput. Biol. 17, e1008865 (2021).
64. He, B., Mortuza, S., Wang, Y., Shen, H.-B. & Zhang, Y. NeBcon: protein contact map prediction using
neural network training coupled with naïve Bayes classifiers. Bioinformatics 33, 2296–2306 (2017).
65. Zhang, Y. & Skolnick, J. SPICKER: a clustering approach to identify near‐native protein folds. J. Comput.
Chem. 25, 865–871 (2004).
66. Huang, X., Pearce, R. & Zhang, Y. FASPR: an open-source tool for fast and accurate protein side-chain
packing. Bioinformatics 36, 3758–3765 (2020).
67. Zhang, J., Liang, Y. & Zhang, Y. Atomic-level protein structure refinement using fragment-guided molecular
dynamics conformation sampling. Structure 19, 1784–1795 (2011).
68. Zhang, Y. & Skolnick, J. Scoring function for automated assessment of protein structure template quality.
Proteins 57, 702–710 (2004).
69. Ramachandran, G. T. & Sasisekharan, V. in Advances in Protein Chemistry, 23 283–437 (Elsevier, 1968).
70. Roy, A., Yang, J. & Zhang, Y. COFACTOR: an accurate comparative algorithm for structure-based protein
function annotation. Nucleic Acids Res. 40, W471–W477 (2012).
71. Yang, J., Roy, A. & Zhang, Y. BioLiP: a semi-manually curated database for biologically relevant
ligand–protein interactions. Nucleic Acids Res. 41, D1096–D1103 (2012).
72. Zhou, X. G., Peng, C. X., Liu, J., Zhang, Y. & Zhang, G. J. Underestimation-assisted global-local cooperative
differential evolution and the application to protein structure prediction. IEEE Trans. Evol. Comput. 24,
536–550 (2020).
73. Zhou, X. G. & Zhang, G. J. Abstract convex underestimation assisted multistage differential evolution. IEEE
Trans. Cybern. 47, 2730–2741 (2017).
74. Zhou, X. G. & Zhang, G. J. Differential evolution with underestimation-based multimutation strategy. IEEE
Trans. Cybern. 49, 1353–1364 (2018).
75. Yang, J., Wang, Y. & Zhang, Y. ResQ: an approach to unified estimation of B-factor and residue-specific
error in protein structure prediction. J. Mol. Biol. 428, 693–701 (2016).
76. Glaeser, R. M. How good can cryo-EM become? Nat. Methods 13, 28–32 (2016).
77. Zhou, X. G. et al. Progressive assembly of multi-domain protein structures from cryo-EM density maps.
Nat. Comput. Sci. 2, 265–275 (2022).
78. Mistry, J. et al. Pfam: the protein families database in 2021. Nucleic Acids Res. 49, D412–D419 (2021).
79. Lu, S. et al. CDD/SPARCLE: the conserved domain database in 2020. Nucleic Acids Res. 48, D265–D268 (2020).
80. Eickholt, J., Deng, X. & Cheng, J. DoBo: protein domain boundary prediction by integrating evolutionary
signals and machine learning. BMC Bioinforma. 12, 1–8 (2011).
81. Tai, C. H., Lee, W. J., Vincent, J. J. & Lee, B. Evaluation of domain prediction in CASP6. Proteins 61,
183–192 (2005).
82. Baek, M. et al. Accurate prediction of protein structures and interactions using a three-track neural network.
Science 373, 871–876 (2021).
83. Pearce, R. & Zhang, Y. Deep learning techniques have significantly impacted protein structure prediction
and protein design. Curr. Opin. Struct. Biol. 68, 194–207 (2021).
84. Born, A., Henen, M. A. & Vögeli, B. Activity and affinity of Pin1 variants. Molecules 25, 36 (2020).
85. Born, A. et al. Reconstruction of coupled intra-and interdomain protein motion from nuclear and electron
magnetic resonance. J. Am. Chem. Soc. 143, 16055–16067 (2021).
86. Chandonia, J.-M., Fox, N. K. & Brenner, S. E. SCOPe: manual curation and artifact removal in the structural
classification of proteins—extended database. J. Mol. Biol. 429, 348–355 (2017).
87. Lam, S. D. et al. Gene3D: expanding the utility of domain assignments. Nucleic Acids Res. 44, D404–D409
(2015).
88. Yu, L. et al. Grammar of protein domain architectures. Proc. Natl Acad. Sci. USA 116, 3636–3645 (2019).
89. Chothia, C. & Lesk, A. M. The relation between the divergence of sequence and structure in proteins. EMBO
J. 5, 823–826 (1986).
90. Emsley, P., Lohkamp, B., Scott, W. G. & Cowtan, K. Features and development of Coot. Acta Crystallogr.
D. Biol. Crystallogr. 66, 486–501 (2010).
91. DiMaio, F. et al. Atomic-accuracy models from 4.5-Å cryo-electron microscopy data with density-guided
iterative local refinement. Nat. Methods 12, 361–365 (2015).

NATURE PROTOCOLS | VOL 17 | OCTOBER 2022 | 2326–2353 | www.nature.com/nprot 2351

 PROTOCOL NATURE PROTOCOLS

92. Zhang, C. et al. Functions of essential genes and a scale-free protein interaction network revealed by
structure-based function and interaction prediction for a minimal genome. J. Proteome Res. 20, 1178–1189
(2021).
93. Zhang, C., Wei, X., Omenn, G. S. & Zhang, Y. Structure and protein interaction-based gene ontology
annotations reveal likely functions of uncharacterized proteins on human chromosome 17. J. Proteome Res.
17, 4186–4196 (2018).
94. Zhang, C., Lane, L., Omenn, G. S. & Zhang, Y. Blinded testing of function annotation for uPE1 proteins by
I-TASSER/COFACTOR pipeline using the 2018–2019 additions to neXtProt and the CAFA3 challenge.
J. Proteome Res. 18, 4154–4166 (2019).
95. Iyer, S., Subramanian, V. & Acharya, K. R. C9orf72, a protein associated with amyotrophic lateral sclerosis
(ALS) is a guanine nucleotide exchange factor. PeerJ 6, e5815 (2018).
96. Skotnicová, P. et al. The cyanobacterial protoporphyrinogen oxidase HemJ is a new b-type heme protein
functionally coupled with coproporphyrinogen III oxidase. J. Biol. Chem. 293, 12394–12404 (2018).
97. Hanson, R. M., Prilusky, J., Renjian, Z., Nakane, T. & Sussman, J. L. JSmol and the next‐generation web‐
based representation of 3D molecular structure as applied to proteopedia. Isr. J. Chem. 53, 207–216 (2013).
98. Hiranuma, N. et al. Improved protein structure refinement guided by deep learning based accuracy esti-
mation. Nat. Commun. 12, 1–11 (2021).
99. Guo, S.-S., Liu, J., Zhou, X. & Zhang, G. DeepUMQA: ultrafast shape recognition-based protein model
quality assessment using deep learning. Bioinformatics 38, 1895–1903 (2022).
100. Xu, J. & Zhang, Y. How significant is a protein structure similarity with TM-score = 0.5? Bioinformatics 26,
889–895 (2010).
101. Ellson, J., Gansner, E.R., Koutsofios, E., North, S.C. & Woodhull, G. in Graph Drawing Software 127–148
(Springer, 2004).
102. Towns, J. et al. XSEDE: acceleratingscientific discovery. Comput. Sci. Eng. 16, 62–74 (2014).
103. Xu, J. Distance-based protein folding powered by deep learning. Proc. Natl Acad. Sci. USA 116,
16856–16865 (2019).
104. Källberg, M. et al. Template-based protein structure modeling using the RaptorX web server. Nat. Protoc. 7,
1511–1522 (2012).
105. Du, Z. et al. The trRosetta server for fast and accurate protein structure prediction. Nat. Protoc. 16,
5634–5651 (2021).
106. Lobley, A., Sadowski, M. I. & Jones, D. T. pGenTHREADER and pDomTHREADER: new methods for
improved protein fold recognition and superfamily discrimination. Bioinformatics 25, 1761–1767 (2009).
107. Zimmermann, L. et al. A completely reimplemented MPI bioinformatics toolkit with a new HHpred server
at its core. J. Mol. Biol. 430, 2237–2243 (2018).
108. Kelley, L. A., Mezulis, S., Yates, C. M., Wass, M. N. & Sternberg, M. J. The Phyre2 web portal for protein
modeling, prediction and analysis. Nat. Protoc. 10, 845–858 (2015).

Acknowledgements
This work is supported in part by the National Institute of General Medical Sciences (GM136422 and S10OD026825 to Y.Z.), the
National Institute of Allergy and Infectious Diseases (AI134678 to Y.Z.), the National Science Foundation (IIS1901191 and DBI2030790
to Y.Z.), the National Nature Science Foundation of China (62173304 and 61773346 to G.Z.), the ‘New Generation Artificial Intelligence’
major project of Science and Technology Innovation 2030 of the Ministry of Science and Technology of China (2021ZD0150100 to G.Z.)
and the Key Project of Zhejiang Provincial Natural Science Foundation of China (LZ20F030002 to G.Z.). This work used the Extreme
Science and Engineering Discovery Environment (XSEDE)102, which is supported by the National Science Foundation (ACI1548562).

Author contributions
Y.Z. conceived and designed the project. X.Z. developed the pipeline and performed the test. W.Z. developed the method for domain
boundaries prediction. Y.L. developed the method for contacts and distances prediction. C.Z. developed the method for protein function
prediction. Y.Z., W.Z., Y.L., C.Z. and R.P. developed the method for individual domain modeling. X.Z. developed the method for multi-
domain protein structure assembly. X.Z. and E.B. tested the server. G.Z. helped supervise the research. X.Z. and Y.Z. wrote the
manuscript, and all authors read and approved the final manuscript.

Competing interests
The authors declare no competing interests.

Additional information
Supplementary information The online version contains supplementary material available at https://doi.org/10.1038/s41596-022-00728-0.
Correspondence and requests for materials should be addressed to Yang Zhang.
Peer review information Nature Protocols thanks Ruben Sánchez-García, Beat R. Vogeli and the other, anonymous, reviewer(s) for their
contribution to the peer review of this work.
Reprints and permissions information is available at www.nature.com/reprints.
Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Springer Nature or its licensor holds exclusive rights to this article under a publishing agreement with the author(s) or other
rightsholder(s); author self-archiving of the accepted manuscript version of this article is solely governed by the terms of such publishing
agreement and applicable law.

2352 NATURE PROTOCOLS | VOL 17 | OCTOBER 2022 | 2326–2353 | www.nature.com/nprot

 NATURE PROTOCOLS PROTOCOL

Received: 4 February 2022; Accepted: 24 May 2022;

Published online: 5 August 2022

Related links
Key references using this protocol
Zhou, X. et al. Proc. Natl Acad. Sci. USA 116, 15930–15938 (2019): https://doi.org/10.1073/pnas.1905068116
Zhang, C. et al. Nucleic Acids Res. 45, W291–299 (2017): https://doi.org/10.1093/nar/gkx366
Zheng, W. et al. Cell Rep. Methods 1, 100014 (2021): https://doi.org/10.1016/j.crmeth.2021.100014
Hermes, C. et al. Nat. Commun. 12, 144 (2021): https://doi.org/10.1038/s41467-020-20418-3

NATURE PROTOCOLS | VOL 17 | OCTOBER 2022 | 2326–2353 | www.nature.com/nprot 2353