Brandon Rice

I wonder if change in a mature industry is incremental or disruptive? I've been on the road last 3 weeks. One of my realizations is that the pharma/biotech industry is having a really hard time understanding the implications of AI in their field. Our conversations are centered on how AI might augment our old way of doing things. Not so much on will there be newer ways of doing things? Biological research is a complex balancing exercise between mechanistic understanding and pure empirical evidence. We are progressively better at understanding - or as I would say hypothesising - about how a disease works and how a drug might 'cure' it. A submission of mechanism of action is more common in IND applications but as far as I can tell it's not a pre-requisite. The greatest successes of drug research are solid empirical hit and trial. Very successful ones at that. The hay day of drug research in the 60s to 90s was because big pharma figured out how to do compound screening effectively at scale. Will AI give us new ways to do this in more effective ways? I would be extremely surprised if that were not to be the case. a. we have never had access to the kind of data that we do now b. we have never had a tool like modern AI to sift through data Those two combined will give us ways to do empirical research like we haven't done before. That much is clear. Will it change the relative dearth of blockbuster drugs? That is a whole another subject. What is clear to me is that the current focus on mechanistic understanding will have to evolve to incorporate AI. We will have newer ways to do things - not necessarily change or abandon old ways. Link to a blog that Abhishek wrote in 2020 which seems truer than ever before.

36

1 Comment

I'm delighted to share news of this new AI-focused drug discovery company, Xaira, built on foundational technology from David Baker and led by Marc Tessier-Lavigne. The company applies AI tools to unlock opportunities in drug discovery for aspects of biology that we believe are not readily obtainable with conventional methods. I co-led NEA's investment in Xaira, along with my NEA partner and the firm’s co-CEO Mohamad Makhzoumi, and I will be joining Xaira’s board. Our enthusiasm for Xaira based on our belief that applying AI and large language models to biology holds immense promise and is still in its early stages. Currently, the application of AI to biology relies on model architectures borrowed from the tech world. A lack of well-curated clinical training sets hinders building and training these models effectively. This limits the advancement of AI in solving biological problems compared to other fields. Xaira aims to bridge this gap. We now have a clear understanding of the data needed to answer fundamental biological questions. By iterating on leading life science models and using focused perturbation data generated through our large-scale experiments, Xaira will generate data specifically suited to building advanced model architectures that are optimized for biology -- and hopefully, unlock therapeutic opportunities for patients who are not being optimally served by existing therapeutics.

490

13 Comments

The Silent Tutors: How LLMs Are Bridging Knowledge Gaps in Cell Therapy Manufacturing

3

Make data immediately and efficiently available so folks can focus on fighting diseases instead of battling outdated systems. It is not okay to hinder biotechs from obtaining their clinical trial data in cumulative CSV or Excel files.  Imagine this: Hours go by, and you're at your desk painstakingly sifting through individual data points. Why? Because someone decided that sending cumulative data was too much to ask. No one should have to: 🔹 Request data files from a vendor contact who may not reply in a timely manner. One client experienced this when their primary contact was out for a week, and they were unaware. 🔹 Have to wait for a weekly or monthly file because that's what the data transfer agreement dictates. Another client could only access their data via PDFs, leaving no room for immediate data exploration or analysis. 🔹 Pay an extra cost for more real-time data updates/transfers. Access to all data should be immediate. #ClinicalTrials #Medicine #RealTimeData #Innovation

14

#data #biology #biotech #food #ai #ML This is such an important issue! First, if you’re not already subscribed to Jesse's newsletter Jesse Johnson, you should be! It’s packed with extremely interesting and educational content, especially for #data and #R&D managers and users (which means all of us, right?). This specific post is incredibly relevant in our company too, especially with teams that aren't traditionally data-oriented. Our solution in my data team along with Omer Genkin? We co-create internal tools with our users, leveraging ML and other cutting-edge techniques to solve real day-to-day challenges. Our mission is to make these tools as intuitive and accessible as possible – with the simplest UI/UX (if any at all) and ensuring the insights are crystal clear. It's all about #empowering everyone, not just the data scientists, to harness the power of #data and #AI.

11

A new company called Xaira Therapeutics just launched with over $1B+ billion in funding for using AI in drug discovery R&D. To give you some context, the last time we saw this level of fundraising was for Roivant Sciences, which focuses on creating companies around abandoned drugs, and Altos Labs, with $3B for cellular rejuvenation programming. But this is the 1st for an AI drug discovery R&D company. Interestingly, a significant portion of the company's technology originates from the Institute for Protein Design at the University of Washington, which is renowned for its work in protein science. This institute is led by Prof. David Baker, the co-founder of this and also the Charm Therapeutics. Read the full story here: https://lnkd.in/eCZfV4CW #drugdiscovery #artificialintelligence

14

1 Comment

By leveraging AI with vast genetic databases, we are not only quickening the pace of identifying potential drug targets but also enhancing the accuracy of clinical diagnostics and predictive modeling. Recently, I had the privilege of discussing the profound impact AI is having on pharmaceutical R&D with Kim Branson, GSK's SVP Global Head of Artificial Intelligence and Machine Learning. Listen on Raising Health: https://lnkd.in/g6QUm2-A

41

5 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

There’s a lot to unpack here, but the open ended questions/comments that are left in this piece alone make you think: - how much more data generation will this trigger? - if the results are “meh” with 60 patients, why would they be stellar in a larger cohort? How would you pick that cohort, since it’s a rare disease? And many many other questions… #ai #drugdiscovery #biologyishard

8

Trying to build AI into your products but worried about potential failure cases? There's a wide gap between a cool demo and something that's production ready. You can bridge that gap with a good data flywheel and solid eval sets. LLM applications need continuous evaluation, monitoring, and improvement to maintain performance over time. Here's a streetfighting cheatsheet for LLM evaluation: 1. Define clear, task-specific success metrics by examining actual LLM outputs, not just theorizing about potential failures. 2. Implement code-based, LLM-based and human judged eval metrics. The order of value is generally human > LLM > code. This also maps to how expensive the evals are to run. 3. Use binary (True/False) metrics when possible for easier alignment and consistency. 4. Validate inputs as well as outputs. 5. Keep metric implementations aligned by regularly sampling and labeling production data. 6. Use dynamic few-shot examples in prompts, retrieved based on input similarity and weighted by recency. 7. Manually iterate on prompts or pipelines based on metric scores and patterns in low-performing instances. 8. Bootstrap improvement by fixing low-scoring outputs and using them as few-shot examples. 9. Break down "good" outputs into multiple dimensions for more accurate labeling and targeted improvements. 10. Consider uncertainty quantification for LLMs to enhance alignment and prioritize data for human review. 11. If your LLM system has many points of call, graph it out. Whiteboarding your system and the potential failure points is very helpful. 12. Log all your data. Use tools like langfuse to make sure no production interaction goes to waste.

6

OMG you would think I would LOVE an announcement that Illumina and NVIDIA were working together but honestly this sounds like this morning's second case of Illumina having to give up a lot to get a little. Deal has two parts: (1) Illumina's "get" is to host a bunch of Nvidia tools for LLMs and data analytics in on the Illumina Connected Analytics platform. Also MONAI for spatial cell imaging workflows though not sure what they have to do with NGS. Anyway, this seems like a nice-to-have, maybe a convenience to customers - but I don't think it opens up any new functionality. Doesn't cost Nvidia anything, and they get to add another logo on their partners slide 🤷‍♂️ (2) Illumina (this is the "give" part) will be "enabling DRAGEN algorithms on NVIDIA GPUs". Not sure why this is good for Illumina, and the PR doesn't explain this benefit at all 🤯 DRAGEN is built on a hardware architecture - FPGA - that has nothing to do with GPU at all, in fact it's kind of the opposite. So Illumina is going to have to make a major investment in porting DRAGEN to Nividia GPU for no stated benefit to Illumina. But the benefit to Nvidia is clear: they get to claim yet another customer for their already vast GPU customer base, and they get to big to displace FPGAs in Illumina instruments. Must be missing something here 🤔🤔🤔 (Don't forget to check out our JMP podcast! https://lnkd.in/gU_qpY3S)

220

29 Comments

It pains me a little as their former communications lead to admit it, but this is the best story written on Recursion to-date. STAT reporter Allison DeAngelis nailed it; the early skepticism, Founder-CEO Chris Gibson's "intense" and indomitable spirit, and importantly, the differences between Recursion's first shots on goal and the rest of the pipeline to come. Excerpt: ---The market, which initially had stratospheric expectations for AI drug development, has become somewhat skeptical of this first generation of candidates following early failures at Exscientia and another firm called Benevolent AI. Powers, the ZS Associates partner, said that Recursion has likely been stressing in public statements that it has 10 data readouts planned for the next 18 months so that it isn’t judged too harshly on the first two or three. “Of course we’re hedging,” Gibson said. “I’m ready for the volatility that will inevitably come…. Nobody’s going to get 10 out of 10.”--- A must read for anyone in the #AI #pharma sector.

14

Just before American Society of Clinical Oncology (ASCO) next week, Tempus AI have announced the have filed for an #IPO with the SEC this week! 🩸Tempus, who were founded back in 2015 are 'looking' to raise $100 million in public offering. 📖 If you dont know the story surrounding Tempus, then it is well worth a read. Founded, by Eric Lefkofsky (who is also the Groupon founder) after his wife was diagnosed with #breastcancer, he had the mission to deliver personalized patient care to patients through collection of molecular, clinical and genomic data by utilizing AI. 🚀Milestones so far include; - Funding of over $1 billion to date. - Investors include Google, Novo Holdings & Franklin Templeton! - Over 450 unique data collections across 2000 healthcare institutions. - Over 95% of pharmaceutical companies use Tempus offerings. - Over 7000 physicians use their tech which equates to more than 65% of academic medical centers in the US. - Over 2300 employees! - Valuation of over $8 billion! 💭Personally, I have never recruited for Tempus AI but, having spent the past 6 years learning more about their journey, it is hard not to root for them during this next phase. 🍿The market competitors will be watching closely no doubt... #cancer

19

This article is inspired by comments from Tony Czarnik and Stefan Ivanov regarding my previous blog post at https://lnkd.in/gFPhXfKW. To clarify, I am writing from the perspective of an AI believer or practitioner, acknowledging the critical viewpoint that claims "AI has shown little in terms of actual achievements in drug discovery." However, I believe it is not sufficient to simply agree with this stance without further elaboration. My focus here is on higher-level concepts rather than practical examples. Here are three key takeaways: 1. Need for Comprehensive Data Collection The current drug discovery process often lacks systematic documentation and relies heavily on subjective reporting, which hampers the ability to draw robust conclusions. Scientists, particularly chemists and biologists, should prioritize automating the collection of extensive data and metadata to create a rich database for analysis. 2. Integration of Mathematical and Data-Driven Approaches Two primary strategies for improving drug discovery are mathematical modeling and data-driven methods. While traditional scientific practices lean toward mathematical modeling for deeper understanding, advancements in AI highlight the effectiveness of data-driven approaches in addressing existing challenges and providing new insights into complex biological processes. 3. Transformative Role of AI AI technology plays a crucial role in drug discovery by enabling scientists to embrace data-driven methodologies that enhance problem-solving capabilities and deepen understanding of biological systems. Combining both mathematical models and data-driven techniques can lead to a more robust framework for innovation in the pharmaceutical industry.

17

21 Comments

What's exciting to me about drug repurposing is that it closes the feedback loop for validating AI-driven predictions. In this clip from the Cure. AI-Drug Discovery event, I explain that for-profit companies need to navigate supporting an AI platform to generate predictions and the lengthy process of traditional drug development. They must continue to support their AI functions without knowing if they have created something of value. In drug repurposing, however, we can validate these AI predictions quickly, directly feeding the outcomes back into the platform. This immediate validation is crucial, ensuring that our AI functions are constantly learning and improving, ultimately leading to more effective treatments for patients. #DrugRepurposing #AI #biotech #pharma

85

4 Comments

One of the most interesting features of AI today is that the barrier to building an initial GenAI product is incredibly low. But that means lots of noise. With the fall life sciences conference season here, I put together the questions sponsors can ask of vendors and their own internal teams to filter down who actually has or will have a successful AI product to allow teams to more efficiently test and progress on their AI goals. Check out the comments for a link to the post!

23

1 Comment

Scaling R&D workflows is tough in early-stage biotech, especially when data handling isn’t built for growth. You start from "quick and dirty" solutions when so much depends on speed. But when that small minority of exploratory work turns into something that's worth turning into a pipeline? That's when you have to pay a massive work debt on formalizing something that was barely annotated (and sometimes barely functional). Because we never know what the "right answer" will be when we start exploring, I think the best compromise to go from fast to robust is to use a system that automatically tracks metadata from the beginning. It should depend on as little human intervention as possible, but allow you to back-search through your exploratory attempts and make sense of it with that metadata. This process of exploring-then-pruning reminds me of slime mold maze solving (https://lnkd.in/e-2EXUkE), and what signals it takes to efficiently search the solution space.

6

I'm sharing with you why long reads have propelled genetic sequencing and precision medicine into a new direction. Remember when we relied purely on Excel to organize and analyze sequencing data (some companies still do this)? And now we sprinkle in AI, making long-read sequencing analysis even more efficient. Here is more of what you need to know 🔊 https://lnkd.in/dgWRbjqF #longreadsequencing #genetics #geneticists

17

1 Comment