YC Root Access

The Future of AI Molecular Discovery

During last month’s NeurIPS 2025 conference, YC’s Ankit Gupta sat down with Ellen Zhong, an assistant professor of computer science at Princeton, to discuss how machine learning is reshaping structural biology. They explore how proteins aren’t static structures but dynamic molecular machines, and how techniques like cryo-electron microscopy combined with ML are revealing protein motion beyond traditional structure prediction. The conversation also dives into inverse problems, noisy experimental data, and what’s next for AI-driven scientific discovery. Chapters 00:11 — Introduction 00:55 — From Supercomputers to Cryo-EM 02:43 — The Rise of Cryo-EM 03:30 — Proteins as Dynamic Systems 04:30 — Inverse Problems in Biology 05:31 — Lessons from DeepMind, Industry and Academia 07:35 — Why Protein Dynamics Remain Unsolved 08:29 — Collaborating with Experimental Scientists 09:28 — What’s Overhyped and Underhyped in AI-Driven Biology 10:51 — The Future of AI-Driven Biology Apply to Y Combinator: https://www.ycombinator.com/apply Work at a startup: https://www.ycombinator.com/jobs

AGAnkit GuptahostEZEllen Zhongguest

Jan 24, 2026·11m·Watch on YouTube ↗

📖 CHAPTERS

1. 1

   NeurIPS meetup: Ellen Zhong’s lab focus—protein dynamics from cryo-EM and small-molecule structure elucidation

   The conversation opens at a NeurIPS afterparty, introducing Ellen Zhong (Princeton) and her group’s agenda in molecular machine learning. She frames the lab around scientific discovery problems: inferring protein dynamics from cryo-EM and extending similar ideas to small-molecule structure elucidation.

2. 2

   Career path: D. E. Shaw supercomputers and molecular dynamics as the starting point

   Zhong describes entering protein-structure work somewhat accidentally through D. E. Shaw Research. That experience grounded her in computationally intensive molecular dynamics (MD) simulations and the culture of high-performance computing for protein folding-related questions.

3. 3

   Switch to experimental structure: discovering cryo-EM during an MIT PhD exploration phase

   During her PhD at MIT’s Computational & Systems Biology program, Zhong explored diverse biological measurement modalities before landing on cryo-EM. Cryo-EM appealed because it can ground questions of motion and structure in real experimental data rather than purely simulation outputs.

4. 4

   Why cryo-EM ‘took off’: a 2012–2013 inflection similar to deep learning’s rise

   Zhong connects cryo-EM’s recent success to a technological inflection point—improvements in instruments and imaging that enabled atomic-resolution structures. That new data quality created a rich computational challenge: reconstructing 3D structure (and motion) from extremely noisy measurements.

5. 5

   Proteins aren’t static: ensembles, conformations, and molecular machines

   The discussion shifts from structure as a single snapshot to proteins as dynamic systems. Cryo-EM images capture ensembles of particles in different states, so the central inference task becomes recovering multiple conformations and their relationships—key to understanding function.

6. 6

   Machine learning as inverse-problem solving: from noisy 2D projections to 3D distributions

   Zhong explains where ML fits: cryo-EM analysis is fundamentally an inverse problem. The measurements are incomplete (noisy 2D projections), so learning requires physics-informed models that combine evidence across many images to infer 3D structures and distributions over states.

7. 7

   Lessons from D. E. Shaw, DeepMind (AlphaFold2 era), and academia: objectives, rigor, and problem structure

   Reflecting on her time across industry and academia, Zhong highlights recurring themes: prioritize the scientific question, enforce reproducibility, and define problems cleanly when possible. She contrasts crisp objective-driven settings (e.g., AlphaFold-style tasks) with messier design problems where validation is harder.

8. 8

   Why protein dynamics remains unsolved—and why academia is a good home for it

   Zhong argues that while static sequence-to-structure prediction has made major strides, protein dynamics lacks a general, practical description and remains a long-horizon research challenge. She positions academia as enabling deeper collaboration and exploratory work that blends computation with new science questions.

9. 9

   Collaboration model: pairing ML with structural biology and chemistry to turn data into discoveries

   Zhong emphasizes that impactful progress comes from close partnerships with experimentalists who generate data and define what constitutes a real discovery. Her lab aims to build methods that automate analysis pipelines and extract additional information from complex experimental measurements.

10. 10

    What’s overhyped vs. underhyped: folding ‘solved’ for static structures, but biology is far bigger

    In the closing segment, Zhong distinguishes between the success of static structure prediction and the broader landscape of unsolved problems. She notes that many proteins are massive multi-component machines and that large regions of structural and functional space remain poorly characterized.

11. 11

    Future of AI-driven biology: new experimental technologies + ML to bridge molecular biology and health

    Zhong forecasts that major advances will require not only better models but also new experimental modalities that expand what data exists. With current AI largely trained on existing datasets, she argues that bridging molecular understanding to human health will depend on tighter ML–experiment co-design.