Jeff Munzar

Over the past year, I’ve been collecting papers with surprising results about the success of machine learning in biology, ones that run against the grain of popular conceptions, that throw into question whether our models are learning biology at all. These are critical results for scientific and technological progress. These results are reminiscent of the hoary computational chemistry wisdom: It’s easy to build models that are effective locally but not globally. When claims of excellent global performance are made they reflect small training sets, bad testing hygiene, or highly biased evaluation data Perhaps it is not Machine Learning that matters, but instead Active Learning. Operationalizing Design-Build-Test cycles requires neither omniscient nor perfectly accurate models. All it requires is that the representations used, and the models built on top of them, are faithful enough that we can make local predictions from their results that are approximately correctly ordered. If anything, these assumptions are weaker than we have traditionally thought. What do we need? “Frictionless Reproducibility.” The ability to rapidly and intelligently analyze results, pick the next set of candidates to test, and then to assay them. The transformation of processes designed for solo scientists into active-learning driven industrial discovery will demand a new set of tools and systems, most of which do not yet exist. Lots more below: https://lnkd.in/e-xX5vi7 #ai #biotech #ml #biopharma #datascience

77

8 Comments

CRISPR and AI are redefining gene editing. Profluent and Stanford University are leading the charge with their latest breakthrough - AI-designed CRISPR models. No more time-consuming search for organisms. No more slow, resource-draining methods. Welcome to the era of AI-driven CRISPR research. AI models, steeped in protein and genomic sequences, can decode genetic patterns with ease. They create CRISPR designs that surpass traditional tools, offering precise edits and significantly reduced side effects. Meet OpenCRISPR-1. An AI-designed, error-minimizing, gene-editing powerhouse. Better yet, it's open-source. Now everyone can explore this technology, free of charge. The future of precision medicine is here. https://lnkd.in/dxt7vJBg #genomesequencing #biomarkers #aiindrugdiscovery

6

Lately, I've been experimenting and building data pipelines for Future Therapeutics and realized how little we talk about data pipelines in cheminformatics. It's a domain where, at any given time, we handle large volumes of data from multiple sources. I'm sharing a special collection of articles on Cheminformatics Data Pipelining, designed to raise awareness and enhance skills in this critical area. Part 1 of 8 is published. Have a look and feel free to share your feedback. Also, let me know if I miss something specific for cheminformatics. I would also love to talk to those actively involved in data pipelining projects to hear your thoughts and approaches. #cheminformatics #drugdiscovery #datapipelines https://lnkd.in/gDa7G2_z

22

I literally come from every panel, key notes, investor conversations and meeting I am invited to where heads of AI engineering to fresh interns start with “LLMs” in hand and mindlessly chant “We need more data, we need more data.” It is quite perplexing to see how the majority is in sync and part of this echo chamber. Here is a two part series about why we just need 'high-variance' small data and not big data. #ArtificialIntelligence #AI #BigData

5

The Precipex paper of the week this week really takes me back. We would have been all over this back at Tripos (now Certara) when we were doing protein ligand docking with genetic algorithms. Super cool idea and execution - definitely worth a look. It's also another example of what we keep seeing in this space: generative AI opening up whole new programming paradigms. Even if you don't care a whit about drug discovery, the approach presented here could be applied anywhere you have an unknown or intractable fitness landscape to optimize.

1

It's been another exciting week for AI in biotech! 🚀 EvolutionaryScale's ESM3 model, simulating 500M years of evolution, is transforming protein design. This joins other recent news on David Baker's Institute for Protein Design raising a whopping $1 billion to further its work. AlphaFold 3 is also pushing the boundaries of protein folding predictions. Still, we must remember that AI is not a magic solution to solve real use cases without having a relevant and efficient feedback loop. True innovation requires mastering both AI tools and hands-on lab skills. A good example is Enzymit which uses tools like ESM and AlphaFold 3 while maintaining a robust experimental feedback loop 💡. This is critical for ensuring success within the short time frames in which startups operate. #AI #Biotech #Innovation

2

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

"Don't work on LLM. This is in the hands of large companies, there's nothing you can bring to the table." -- Unfortunately, I believe that LeCun is correct.. The omniscient nature of LLMs makes this reality extremely concerning if you take this to one rational (potential) conclusion: large tech companies own the big LLMs and 0% chance of any small outfit entering the space given the cost to train a competitive LLM ---> LLMs can be applied to any sector/subject domain ---> a handful of big tech companies eventually dominate all sectors Just a thought-experiment and I tend to be much more optimistic than most, but I have been concerned about the potential for this for quite a while.. Curious as to what you guys think? Piyush Mathur MD,FCCM,FASA,FAMIA Kavya Pearlman ⚠️ Safety First ⚠️ Alfonso Limon, Ph.D. Anthony Chang, MD, MBA, MPH, MS https://lnkd.in/eYBKG2M6

28

15 Comments

Let’s talk about “Software-Defined Biology”. I dislike the terms “AI Drug Discovery” and “techbio” since I think they’re imprecise and missing what makes the current shift in the way we do science so important — the software. We wrote a whole blog post on this: https://lnkd.in/gEnFiJ6u We argue that the core differentiation of the current class of modern biotech companies is their desire and willingness to write down their scientific processes in code. AI/ML is part of that, sure, but only part. These companies have a fundamental mindset shift: gone are the days of software as a mere follow-on to scientific processes. Science as an artisanal pastime is now being transformed into a programmable and scalable endeavor. Skeptics may argue that software has been used in drug discovery for decades. However, software-defined biology goes beyond traditional bioinformatics and computational chemistry, embracing a new role for software as a central component of the scientific process. Software-Defined Biology involves defining data models in code, capturing structured data and metadata, and introducing automated systems like machine learning models for decision-making. It shifts the focus to building systems that make the best molecules Two major trends driving this: first, the increase in useful AI models. Second, and less discussed, the increasing number of complex multi-part modalities. RNA + LNPs, ADCs, PROTACs, Molecular Glues, etc — all increasingly popular modalities Increased complexity is best managed in code. As more biotech companies embrace software-defined biology and prove their successes in the clinic, this approach will become increasingly prevalent. Companies who have embraced this shift are on the cutting edge and have historically had to build all of this software themselves. It’s still early, but we’re seeing the rise in specialized tools and platforms to support this next generation of biotech companies Interested in talking about the implications of this shift? Message me or say hi by emailing hello (at) sphinxbio (dot) com

63

2 Comments

If you use LLMs in your cancer research, check out my recent seminar all about how data quality impacts LLM performance. A reliable foundation that is well annotated and accessible to an LLM plays a major role in the value of its results. I share examples of how LLM-powered AI agents query across three versions of the same gene expression corpus with differing results, including: --Unstructured data from the public repository Gene Expression Omnibus. --Structured data from the Crowd Extracted Expression of Differential Signatures project (tool developed by the Ma’ayan Lab at the Icahn School of Medicine at Mount Sinai). --Clean, linked, and harmonized data. Watch it here: https://lnkd.in/gbPB5CRv #LLM #Data #Cancer #Research #Science #Innovation

52

We really like this kind of work at PostEra. Much of our core thesis in founding the company was that Design algorithms by themselves, or Synthesis algorithms by themselves, were not nearly as meaningful for Drug Discovery as methods that unified the two. It's not about always making easy molecules. And it's not about always making hard molecules. What we find matters is being able to quantify upfront the synthetic investment relative to the potential utility of molecules you're evaluating. This philosophy has really been at the heart of Proton, our AI platform, and how we do med chem at PostEra. Really cool work from Jenna Fromer and Connor W. Coley. ➡ Paper: https://lnkd.in/eZe-zGVX 💰 No Paywall Paper: https://lnkd.in/exe3tZJr

89

3 Comments

Belated post, but my recent talk on how we're leveraging #Metaflow for disease trends forecasting at a lab in Santa Clara University is now featured in the #Metaflow Stories playlist by Outerbounds! #Metaflow is an open-source framework that makes it easy to build and manage #AI projects. Thanks Shri Javadekar and Outerbounds for the platform, and to everyone who participated in the discussion! The full presentation is available here https://lnkd.in/dmmTdnHZ On a side note, I'm open for Summer & Fall internship opportunities in ML and LLM space. If you've something in mind, hit me up! #ML #MLOps #MLEngineer

48

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

This paper is inspired! Much of the current AI weather model research is focused on more accurate and faster versions of what we already have. That's important, but the most exciting advances will come from asking "what *new capabilities* can AI models provide?" I've heard the phrase "AI models will allow us to ask new science questions" before. These folks actually did it and found something fascinating. [Side note: the approach they use find a more accurate initial state is the same one we use to assimilate near real time streamflow observations into HydroForecast at Upstream Tech. Cool!]

35

3 Comments

Fascinating paper (preprint only) that digs into the weaknesses in basic reasoning of LLMs. I love the simple problems the researchers used: "Alice has N brothers and she also has M sisters. How many sisters does Alice’s brother have?" From the abstract ... We demonstrate here a dramatic breakdown of function and reasoning capabilities of state-of-the-art models trained at the largest available scales which claim strong function, using a simple, short, conventional common sense problem formulated in concise natural language, easily solvable by humans. The breakdown is dramatic, as models also express strong overconfidence in their wrong solutions, while providing often non-sensical "reasoning"-like explanations akin to confabulations to justify and backup the validity of their clearly failed responses, making them sound plausible. Various standard interventions in an attempt to get the right solution, like various type of enhanced prompting, or urging the models to reconsider the wrong solutions again by multi step re-evaluation, fail. We take these initial observations to the scientific and technological community to stimulate urgent re-assessment of the claimed capabilities of current generation of LLMs, Such re-assessment also requires common action to create standardized benchmarks that would allow proper detection of such basic reasoning deficits that obviously manage to remain undiscovered by current state-of-the-art evaluation procedures and benchmarks.

23

2 Comments

A design pattern in LLM workflow is to use LLMs to augment each step along the build workflow so that human efforts can be statistically amplified. Areas where we can use LLMs for amplification: human-in-the-loop synthetic training data generation, LLM as a judge, eval data generation, RLAIF, etc. In this pattern, the role of human shift from creators to reviewers. And with ongoing human review, the LLMs are self-improving overtime (using in-context learning) to become better human amplifiers. Who are building self-improving AI systems to support this pattern of ongoing human intent clarification & amplification?

21

To date, Erum Azeez Khan and I have published more than 80 episodes dissecting all things biotech on the #GrowEverything podcast. We’ve covered everything from biomaterials revolutionizing everyday products to microbes paving the way for interstellar exploration. Whether you're a seasoned biotech pro or a curious newcomer, there's always something new to learn. If you’re a synthetic biology enthusiast, we’ve selected some episodes just for you: Forget jargon, Ginkgo Bioworks, Inc.'s Cas Smith makes synthetic biology (synbio) fun and easy to understand. We explore how innovative synbio solutions are transforming multiple industries. Listen here: https://lnkd.in/ewPnMtkX What’s the latest in the world of synbio? Paul Freemont of Imperial College London tracked synbio’s progress and impact over the past five years. We talked about how universities are making biotech knowledge accessible worldwide, how rural and urban areas are collaborating on revolutionary manufacturing systems, and the surprising role of art and design in fueling bio breakthroughs. Listen here: https://lnkd.in/eGHFNAG5 How is synthetic biology shaping the future? In our conversation with SynBioBeta founder John Cumbers, we discuss bacteria for tasks like cleaning teeth, explore subscription-based revenue models, and the importance of innovation in a field threatened by monopolies. Listen here: https://lnkd.in/eZ7z9SiT When you listen, be sure to hit the subscribe button to join our growing community of bio-enthusiasts! Let’s Grow Everything!

38

5 Comments

"The hype will play out in the long term," James Fraser said. "The hard part of drug discovery (is) it's a pretty long business. You can't take the wet lab part out of it yet. Stuff needs to play out in the clinic." Ron Leuty takes an in-depth look at Profluent and how #AI has become a force in drug development (biologics & small molecule). We are seeing this play out in applications to the QB3 mentoring program. It is striking how many new companies are touting that their use of AI/ML provides a competitive advantage. It's almost more remarkable if a drug discovery company is NOT using AI. And it is very hard to judge the longer-term commercial prospects of a new company whose approach relies on AI/ML. Obviously the quality of the data the models are using is paramount. Anyone with thoughts about how to assess these companies, please weigh in! https://lnkd.in/gaVr3CPj

18

GENE EDITOR CREATED BY #GENERATIVEAI 𝗙𝗿𝗼𝗺 𝗧𝗲𝘅𝘁 𝘁𝗼 𝗩𝗶𝗱𝗲𝗼 𝗮𝗻𝗱 𝗻𝗼𝘄 𝗗𝗡𝗔, 𝗚𝗲𝗻𝗔𝗜 𝗰𝗼𝗻𝘁𝗶𝗻𝘂𝗲𝘀 𝘁𝗼 𝗰𝗿𝗲𝗮𝘁𝗲... Generative AI, known for creating poetry, programs, and imagery, is now used to generate blueprints for biological mechanisms to edit DNA, paving the way for more precise treatments for illnesses and diseases. Profluent, a Berkeley startup, introduced new AI technology to create gene editors by analyzing extensive biological data, including the microscopic mechanisms scientists use to edit human DNA. These gene editors leverage Nobel Prize-winning CRISPR technology, which has revolutionized the study and treatment of hereditary conditions like sickle cell anemia and blindness. Profluent's AI-generated gene editor, OpenCRISPR-1, is open-sourced, allowing individuals, academic labs and companies to experiment with the technology for free. Though Profluent is open sourcing the gene editors generated by its A.I. technology, it is not open sourcing the A.I. technology itself. While these innovations could transform healthcare, they haven't undergone clinical trials, and ethical and safety concerns, particularly about human gene editing, remain. If we remember, already the emergence of "home scientists" using CRISPR has raised concerns about unsupervised gene editing. 🔗 Check full article on The New York Times, link in the comments. 📽 Video Credit: Joseph Gallagher / Profluent Bio : A time-lapse of human cells edited by OpenCRISPR; source: The New York Times. #ArtificialIntelligence #transformers

8

1 Comment