Diogo Camacho

I learned a new word: petavoxel I just came accross an exciting publication in Science by Alexander Shapson-Coe et al. The teams at Harvard University and Google Research, unlocked a new dimension of neural exploration with their "A petavoxel fragment of human cerebral cortex reconstructed at nanoscale resolution." The dataset is so vast, it offers a detailed roadmap of around 57,000 cells and 150 million synapses in just 1 cubic millimeter of brain tissue. To put this in perspective, the world’s largest metropolis, Tokyo, “only” has a population of approximately 40 million people! This 1.4 petabytes of data captures the brain's microstructure in previously unknown detail. https://lnkd.in/e4d8eNcz The implications are profound and uncovering the nuances of brain architecture could revolutionize treatments for conditions like epilepsy, Alzheimer's, and Parkinson's. Also, this is #OpenScience, setting a new bar for neuroscience collaborations and #innovation. Back to #petavoxel : it’s a unit of measurement of tri-dimensional pixels, a whole quadrillion of them! #TheMoreYouKnow #HappyFriday.

8

data cleaning and wrangling (the mundane work) are what you need to be good at. My odyssey of obtaining scRNAseq metadata https://lnkd.in/eB-kDGeD

20

1 Comment

Good morning everyone! Do you like AI and graph neural networks? (see what i did there? :-) ) Our blog today helps you discover how SpatialGlue, a groundbreaking model published in Nature Methods, leverages graph neural networks and dual-attention mechanisms to integrate spatial and multi-omics data, providing unprecedented insights into tissue architecture and cellular interactions. Read our blog for a summary of how this innovative approach outperforms traditional models, offering precise mapping of complex structures and advancing research in fields like cancer and neurology. #SpatialTranscriptomics #ComputationalBiology #Bioinformatics #AI

12

Thoughts on the AlphaFold 3 Controversy Nature's editorial today (https://lnkd.in/e4MD5FZB) regarding its decision to publish Google DeepMind and Isomorphic Labs' paper on the AlphaFold 3 release earlier this month, raises interesting points. Read the AlphaFold paper here: https://lnkd.in/eP3CSr3P It is clear that we are in the midst of a true revolution. AlphaFold 3's ability to now predict interactions of biological molecules is an incredible breakthrough. Yet, the lack of immediate code availability nonetheless raises questions about transparency and reproducibility in AI research. I think that Nature's decision, which falls within the guidelines and discretion of their editors also takes into account the interests of all parties. The editorial mentions that the publication will be updated when the code is made available in 6 months. We will surely discuss these topics on the "AI in Healthcare" panel in which I will represent GENOSKIN and participate during the Biotechnology Innovation Organization convention in San Diego. ➡️ Where & When: San Diego Convention Center Corporation Booth #3823 | June 3rd, 2.00-2.45 PM What are your thoughts on the balance between encouraging scientific breakthroughs and ensuring transparency in AI research? I look forward to exchanging on the innovative ways AI is transforming Healthcare. If you're attending, let's connect. ➡️ Schedule a meeting: https://lnkd.in/eW4nXpk9 See you in sunny San Diego! ☀️ #BIO2024 #AI #AlphaFold3 #OpenScience #Innovation #Revolution Business France SoCal French-American Chamber of Commerce ▶ Stephane Richard Bettina ExpertonSteve LevineJean-Frederic Petit-Nivard Brian Atwood Sylvie ALMERI Camille Risso Jerome Revole Nam-Hee LEE

18

3 Comments

The CHAIMELEON Project aims to build an EU-wide cloud-based infrastructure for health imaging data to facilitate and contribute to developing and validating AI tools for improved cancer management. Key points: 💡 We have built a pan-cancer repository guaranteeing data standardization and image harmonization of patient data from different European hospitals to facilitate the development of reproducible and generalizable AI models. 🔏 We apply the European data privacy standards to avoid patient re-identification. 📈 We have built an infrastructure that combines storage and processing capabilities to train and validate AI solutions to accelerate lab-to-market transition. 🏆 We have organized an Open Challenge to open the developed infrastructure to the research community. More info: https://lnkd.in/dudFz-Kn 🎥 Do not miss the following video if you want more details of the project 👇 👇 👇

23

There are now many providers of metadata of scientific papers: details about license, bibtex, retraction status, author list, COI statements. We aggregated five of these into one client with fallbacks in PaperQA2. We use it for adding context to LLM calls, but this could be a great tool for exploring science of reserach papers

137

3 Comments

Can Polaris become the guiding star in this drug development #rAIvolution? A recent correspondence in Nature Portfolio's Nature Machine Intelligence, led by Cas Wognum of Valence Labs and Recursion, calls for "an industry-led initiative to critically assess machine learning for real-world drug discovery." This collaborative effort is both timely and crucial for our industry's future. Here's a free copy of the paper: https://lnkd.in/eNkQGqUR The authors, representing a consortium of ten leading biotech and pharmaceutical companies, argue that despite exciting innovations, we lack standardized benchmarks and evaluation methods tailored to drug discovery's unique challenges. This gap is delaying ML adoption and potentially misleading us about real-world impact. The fact that competitors are joining forces underscores the importance of this initiative. These include: Johnson & Johnson, Bayer, Novartis, Blueprint Medicines, Merck, Nimbus Therapeutics, Pfizer, AstraZeneca, and Relay Therapeutics. Why does this matter? The challenges we face in implementing ML for drug discovery are multifaceted: • Inconsistent datasets hinder progress • High-stakes decisions require reliable methods • Current benchmarks may not reflect real-world conditions • Interdisciplinary expertise is crucial but often siloed The proposed solution is an open-science, cross-industry initiative focusing on curating representative benchmark datasets, developing rigorous evaluation guidelines, and creating open-source tools to implement best practices. We have a responsibility to bridge the gap between perceived progress and actual impact. This Polaris - Benchmarks for methods that matter effort could become our guiding star in the vast sea of data and methodologies. By pooling our expertise and resources, we can establish a fixed point of reference to guide ML innovation in our field. The question now is: how can we all contribute to this initiative? What unique insights could each of us bring to the table? Our collective expertise can chart a more accurate course for innovation, potentially revolutionizing how we approach drug development and ultimately benefiting patients worldwide. Let's start a conversation about shaping the future of ML in drug discovery. #MachineLearning #DrugDevelopment #Innovation #Biotech #Collaboration Illustration EMxMJ

20

2 Comments

It's all LLM training all the time now, nothing else matters 😳 "The Tahoe-100M data set is the world's largest atlas of single-cell transcriptomic data, mapping how drugs impact patient cells at scale. It comprises 100 million cells and 60,000 experiments, mapping 1,200 drug treatments across 50 different tumor models. It is larger than all public single-cell datasets combined, and 50x larger than all publicly available, drug-perturbed single-cell data put together." "The NVIDIA team will contribute machine learning and data engineering expertise to TRAIN MODELS on this data and package for use."

118

4 Comments

What's all the Fas about? As a consultant I'm often asked to narrow down a long list of potential molecules to a more manageable number that can be made and tested in the lab. In an ideal world you would produce and test them all. Biology is complex and we seem to be pretty rubbish at predicting it. However at some point reality and resource limitations kick in and you have to make some reasoned choices. So it's nice to see a paper where researchers did produce an extensive list of proteins and find the unexpected! Fas ligand is a trimer that binds to Fas (CD95) and activates cell death by trimerizing the receptor. In some cancers tumor associated macrophages and other cells can activate the Fas pathway and lead to the death of activated T-cells. Blockade of FasL is currently being explored in a ph2 trial of asunercept, a soluble Fas-Fc fusion protein. However, this is dimeric Fas trying to block trimeric FasL. Would a trimeric molecule not be better and maybe a hexamer would be even better? Researchers at University Hospital Wurzburg set out to test this by making monomeric, dimeric, trimeric, tetrameric and hexameric forms of Fas. If I was trying to predict a winner I would have maybe gone with the hexamer or said that everything from trimer upwards would be equal. Not for the first time I'm wrong. Trimer and hexamer are about equal but the tetramer is far superior. For a proposed model as to why take a look at figure 6 and associated text. I'm a big fan of the advances in AI but I hope that in this new era of computer aided biology and drug development we don't forget that sometimes we have to acknowledge our lack of understanding and just make and test lots of things. DOI: https://lnkd.in/gR3eibGJ ----- I'm Ian, I post about antibody engineering, recombinant proteins and my journey to bootstrap Gamma Proteins into a leading supplier of Fc receptors. If you like my content please reshare with your network and follow me to see more.

130

12 Comments

Analysis pipelines for biological data have already turned into a commodity. Especially with great projects like nf-core. For most types of biological analysis there are now a dozen+ oss implementations on github. But an end-to-end solution to easily go from data to results... That is still an open problem. “Scientific questions are often complex, and no single platform can address every issue.” That is, in my perspective, the crux of the issue. Computational biologists juggle between half a dozen point solutions to go from raw data to results. Nextflow is an important, but small piece of that whole equation. It still needs a workflow executor to run the pipelines, then hosted notebooks to do downstream analysis, a database to keep track of metadata, and easy plotting interfaces for scientists to explore the results. Those need to be built AND maintained. Suddenly between time, salaries, and un-optimized cloud compute the costs of the analysis skyrockets. But it does NOT have to be that way. Scientists seek solutions, bioinformaticians seek infrastructure, so creating a great analysis experience for both requires a platform that has all those components built-in, interconnected, and with great GUI, CLI, and API interfaces. The balance of covering all the bases, as to not contribute to the point-solution deluge, while providing a delightful experience to wet-lab and computational users alike, is incredibly complicated to achieve. Far from commoditized. I do not believe ANY of the existing data infrastructure platforms have achieved that well. LatchBio is obviously working incredibly hard to try, but the jury is still out.

40

1 Comment

Integrating multiomics and prior knowledge: a study of the Graphnet penalty impact https://lnkd.in/ewk6UnQz

22

A useful article: Sequence alignment algorithms are essential tools that enable scientists to compare biological sequences such as DNA, RNA, and proteins. These algorithms help identify regions of similarity that may indicate functional, structural, or evolutionary relationships between the sequences. Since one major and important task in bioinformatics is sequence alignment and choosing the right tool can impact the project, I have published an article in my blog related to classifying "sequence alignment algorithms" into different groups. You can read it from here: https://lnkd.in/dSMkAsin #bioinformatics

40

2 Comments

Wow - 2 ex-Meta employees launch EvolutionaryScale landing $142m to advance AI in biology. EvolutionaryScale has introduced ESM3, a cutting-edge language model for biology that predicts protein sequences, structures, and functions. This model can simulate millions of years of evolution, generating new proteins with potential applications in medicine, research, and sustainability. ESM3 represents a significant advancement in AI's ability to design and engineer biological systems. There will be comparisons with AlphaFold3, intrigued where this goes... https://lnkd.in/dhtQDtsA #AIinDrugDiscovery #genomics #proteomics

3

1 Comment

Swarms, gratitude and good karma In March 2000, Guy Theraulaz and I made the cover of Scientific American for "Swarm Smarts", a sort of companion article to our book with Marco Dorigo, "Swarm Intelligence" (https://lnkd.in/gqJwRTvF). We were elated, not just for making the cover of SciAm but because it was a terrific article that greatly benefited from the work of our super-talented editor -the article was so good that is was republished in 2008 (https://lnkd.in/gyyaGi5e). Working with SciAm editor Alden Hayashi through successive iterations of the manuscript, I could see our somewhat tedious exposition become alive and exciting. And I told him so. When the article came out, so much better than how it came in, I was effusive in my gratitude: Alden had really elevated our prose. What he told me was surprising and a bit sad: authors are usually upset about one thing or the other and my recognition of his essential role was a unique event. I did not consider myself a particularly kind or grateful person, or perhaps it was a lack of awareness, but I was touched by the impact gratitude can have. About one year passes and I had moved on from the academic universe to the business world. Then Alden writes: he had also moved, from SciAm to Harvard Business Review, as an editor. He thought that as a scientist I had probably never heard of HBR and asked me whether I knew someone who would write a business-oriented article on Swarm Intelligence. Well, not only had I heard of HBR, but I was his man. Then-editor-in-chief Suzy Welch gave the go-ahead. And I got to work with Alden again, with my dear late friend Chris Meyer as a co-author. Alden supported me as editor on a couple more projects, and with each one he elevated the quality of the writing and the clarity of the concepts. This all lead to an inflection point in my career. Just because I expressed my admiration and gratitude with sincerity, not a common occurrence in my life then. I have been working on it since then. Don't hide your sense of awe and shower people who impress you with compliments and thanks. Even for little things. #gratitude

19

2 Comments

Fitbit (now part of Google) dashboards only give you correlation—not causation. Same goes for any LLMs. Here's how to add #causalinference #digitaltwins to such #timeseries analyses: https://lnkd.in/gByQ-EUR All time series models that follow the general form Y~f(X,W) are welcome! (stats, machine learning, deep learning / AI, etc.) (2-min video explainer: https://lnkd.in/gx5Mj-Zj ) #switchback experiment #nof1 #singlecasedesign #digitaltwin #esametry #esametrics *** ⏩ Subscribe to statsof1.org for more n-of-1 posts and podcast episodes (https://lnkd.in/gGuzJenr). ⏪ 📚 N-of-1 / SCD resources (papers, links, etc.)—and how to add your own: statsof1.org/resources 📖 (How to cite our resources page: https://lnkd.in/gfAVwCgZ)

57

9 Comments

Tempus AI stock on a tear ($7B market cap today!) gives me an idea for a fun deal 💡🤪 I've long been intrigued about the potential value of GRAIL's phenotype+genotype+methylome database - no way to know really but could be worth far more than Galleri (which seems to have very little value as GRAL market cap << cash). So, what if Tempus bought Grail? What would it get, and what would it cost? We can assume from this week's big Grail layoff that it now operates free and clear of EU restrictions, so I would think pretty much anyone except Illumina could buy it and do with what they wanted (yay capitalism!) The amazing thing buying Grail is that it has $1B in cash but has a market cap of just $500M. So you could offer a 30% premium and still come out of the deal with $350M profit and a fantastic database that could add a couple of billion to Tempus market cap. Tempus could sell Grail data to its pharma customers, restart development of the MRD and DAC tests just halted (Tempus rents its MRD test from Personalis), and save an massive amount of money by pausing Galleri sales - could be restarted in the (unlikely) event Galleri gets FDA approval and CMS reimbursement. Alas there is a fly in the ointment of this whacky plan! There's a clawback clause that requires that if Grail is bought in the 15 month period after the spinout, a bunch of the cash has to get back to Illumina. I think it is: $clawback = $500M x (15 - months-since-spin)/15. So, doesn't work out to buy Grail right now as Illumina would get $500M back and we think we need $650M to buy. But (yay algebra! can you tell I'm on a long flight today? 🤣) if Tempus waits 5 months then Illumina only gets back $350M leaving $650 in bank 👏 So, my unsolicited advice is for Tempus AI is buy Grail in December 2024 and it won't cost you a dime and you'll have an amazing data asset and your own MRD test 🤣🤣🤣

90

11 Comments

Researching on applying large language models to genomics is booming - so much so that this team from University of California, Berkeley has written a survey paper https://lnkd.in/gW9XgBY7 covering 30 different models. (And this must be big because genomics models have their own acronym now 🤣) "Large language models (LLMs) are having transformative impacts across a wide range of scientific fields, particularly in the biomedical sciences. Just as the goal of Natural Language Processing is to understand sequences of words, a major objective in biology is to understand biological sequences." "Genomic Language Models (gLMs), which are LLMs trained on DNA sequences, have the potential to significantly advance our understanding of genomes and how DNA elements at various scales interact to give rise to complex functions." "In this review, we showcase this potential by highlighting key applications of gLMs, including fitness prediction, sequence design, and transfer learning." "One of the key pillars of this development is self-supervised learning, in which training on massive amounts of unlabeled data enables the learning of complex features and their interactions." There should already be a gLM on a li'l GPU in every Illumina sequencer! (my post from a year ago 😠 https://lnkd.in/gGk4cA6h) Then Illumina could have announced a deal with NVIDIA and be one of the cool kids like Thermo Fisher Scientific this week 🥳 (or...Molly He? Francisco Garcia? ElementAI 🤣)

76

7 Comments

I just want to be a farmer with kittens, butterflies and animals running around! Why? Spectral leakage. Yup, it sounds innocuous, but it's a bug/condition that occurs in Fourier Neural Operators. Have you ever faced a bug that seems to grow more elusive the deeper you dig into it? I'm knee-deep in that battle after spending all of Sunday on it, and I'm still at it now. AI research is exhausting! The crux of the issue is improper windowing in the Fourier transform, which leads to spectral energy leaking into adjacent frequencies. The result is inaccurate predictions, unstable training, and a nagging sense of unease every time you rerun the model. What's gnarly about this bug is that it's invisible until it's not. No errors, no crashes, just degrading accuracy or predictions that don't quite fit. Debugging it requires an intersection of signal processing intuition, neural network know-how, and a healthy dose of patience. I've been untangling it bit by bit—implementing windowing, validating against known solutions, and questioning every decision made in the code. It’s exhausting, but I guess that’s the nature of chasing spectral ghosts in complex systems. #Kittens #Rainbows #Butterfly #AI

1

2 Comments