CHAPTERS
- 0:11 – 0:50
NeurIPS chat setup: ML meets structural biology
Ankit Gupta introduces Ellen Zhong at a NeurIPS afterparty conversation focused on AI and structural biology. Ellen briefly frames her lab’s core interests: protein dynamics from cryo-EM and small-molecule structure elucidation.
- •Context: YC Root Access interview at NeurIPS
- •Ellen Zhong’s role at Princeton and research focus
- •Molecular machine learning as a tool for scientific discovery
- •Two main domains: cryo-EM protein dynamics and small-molecule structure elucidation
- 0:50 – 1:04
From D. E. Shaw supercomputers to experimental structure discovery
Ellen describes how she entered protein structure research through work at D. E. Shaw Research, where large-scale molecular dynamics (MD) simulations and specialized hardware were central. She explains why she later sought new biological problems beyond simulation-driven approaches.
- •Early exposure to protein folding via MD simulations at D. E. Shaw Research
- •Simulation-heavy paradigm and its strengths/limits
- •Motivation to expand beyond MD into other areas of biology
- •Transition mindset: learning “something new” in biology enabled by computation/AI
- 1:04 – 2:36
Discovering cryo-EM during her MIT PhD exploration
During her MIT PhD, Ellen explored multiple computational biology areas before landing on cryo-EM. Cryo-EM appealed because it uses real experimental measurements, offering a different path to studying molecular motion than purely simulated trajectories.
- •PhD program: computational systems biology at MIT
- •Exploration across neuroscience, mass spec, RNA before cryo-EM
- •Cryo-EM as experimental “pictures” of proteins for 3D structure
- •Why experimental data felt compelling vs simulation-only validation challenges
- 2:36 – 3:24
The rise of cryo-EM and its parallels to deep learning’s inflection point
Ellen explains that cryo-EM’s rapid progress mirrors deep learning’s ‘things suddenly worked’ era around 2012–2013. Improvements in imaging hardware enabled atomic-resolution structures and created new computational reconstruction challenges.
- •Hardware/measurement breakthroughs improved electron microscope images
- •Atomic-resolution structural determination became more feasible
- •Cryo-EM progress timeline compared to deep learning’s breakthrough period
- •New computational focus: reconstruction from noisy measurements
- 3:24 – 4:25
Proteins as dynamic machines, not static structures
The discussion shifts from single structures to ensembles and motion. Ellen emphasizes that proteins ‘jiggle’ and change conformations to perform function, and that capturing these motions is key to understanding biology.
- •Cryo-EM images are snapshots of an ensemble, not a single pose
- •Inference goal: recover multiple conformations/movies from heterogeneous data
- •Structural biology moving from static views to dynamic understanding
- •Function arises from motion in molecular machines
- 4:25 – 4:57
Inverse problems in biology: learning 3D structure from noisy 2D projections
Ellen connects cryo-EM analysis to machine learning through inverse problems: reconstructing missing, latent 3D structure from incomplete, noisy 2D measurements. Her lab uses physics-inspired ML to learn distributions over structures rather than a single answer.
- •Core framing: inverse problems with incomplete experimental measurements
- •Input: noisy 2D projection images; output: 3D coordinates and dynamics
- •Physics-inspired ML models for reconstruction and inference
- •Learning complex distributions of structures from imaging data
- 4:57 – 7:03
Lessons from D. E. Shaw, DeepMind/AlphaFold, and academia
Ellen contrasts research cultures across industry and academia and what she imported into her lab’s approach. She highlights reproducibility from D. E. Shaw, objective-driven problem framing from DeepMind, and the open-ended, collaborative nature of academia.
- •D. E. Shaw influence: rigor and reproducibility
- •DeepMind influence: crisp problem definitions and optimization objectives
- •Experience during AlphaFold2’s release and its unique environment
- •Academic challenge: framing messier problems (e.g., design) and validation
- 7:03 – 8:15
Why protein dynamics remains unsolved—and why it motivates academic research
Ellen argues that while static structure prediction has advanced, protein dynamics still lacks a general description and reliable modeling framework. She views progress as requiring more than ML alone, motivating long-term academic work and cross-disciplinary collaboration.
- •Static fold prediction ≠ full understanding of protein behavior
- •Dynamics lacks a unified representation and general solutions
- •ML alone is insufficient; scientific progress needs broader approaches
- •Academia enables long-horizon exploration and diverse collaborations
- 8:15 – 8:54
Collaborating with experimentalists (and chemists) to turn methods into discovery
Ellen explains how her group works directly with experimental scientists who generate data and define what counts as a meaningful biological or chemical discovery. The computational goal is to automate processing and extract more information from complex measurements.
- •Collaboration model: experimentalists provide data + domain expertise
- •Focus on discovery-oriented outcomes, not just benchmarks
- •Method development for automating pipelines and revealing hidden signal
- •Expansion beyond cryo-EM structural biology into chemistry collaborations
- 8:54 – 11:52
What’s overhyped vs underhyped in AI biology—and what comes next
Ellen distinguishes ‘solved’ static sequence-to-structure prediction from the much larger frontier involving massive complexes, unknown structural space, and dynamics. She predicts major progress will require new experimental technologies plus ML that can leverage new data to bridge molecular biology to health outcomes.
- •Static sequence → structure may be largely solved; dynamics and complexes are not
- •Many proteins are huge multi-component machines with complex motions
- •Models have absorbed existing data; next leaps need new data sources/tech
- •Bridging molecular mechanisms to human health will demand experiment–ML co-design
