Scouting Frontiers in AI for Biology: Dynamics, Diffusion, and Design, with Amelie Schreiber

Episode Summary

Executive Summary: The episode surveys rapid advances in AI for biology, emphasizing that the frontier has shifted from static protein structure prediction toward modeling dynamics, ensembles, and multi-step design workflows. Amelie Schreiber argues that specialized models—AlphaFold 3, ESM3, RFdiffusion, peptide and antibody generators, enzyme pocket models, and MDGen—are increasingly being chained together by agents to create more capable, higher-throughput protein engineering and drug discovery pipelines.

Main Topics: From structure prediction to biological design workflows (Priority: 5/5): The conversation frames the field as moving beyond single-model structure prediction toward end-to-end workflows that design, validate, and optimize biomolecules for specific tasks. AlphaFold 3 and interaction mapping (Priority: 5/5): AlphaFold 3 is discussed as a major step because it models complexes involving proteins, DNA, RNA, small molecules, and ions, enabling richer interaction-network inference, though licensing limits commercial use. Dynamics as the new frontier (Priority: 5/5): Schreiber argues that static structures are insufficient for catalysis and many functional problems; conformational ensembles, side-chain motion, and molecular trajectories are becoming central. Specialized generative models for peptides, antibodies, and enzymes (Priority: 4/5): New models such as PepFlow, GOAB, and EnzymeFlow are presented as task-specific generators for difficult biomolecular design problems, especially short/disordered peptides and catalytic pockets. MDGen and trajectory generation (Priority: 5/5): MDGen is highlighted as a potentially transformative model for generating molecular dynamics trajectories, upsampling, interpolating between states, and inpainting missing trajectory segments much faster than traditional MD. Workflow orchestration and agentic design (Priority: 4/5): The future is described as agent-driven systems that chain multiple models, tune hyperparameters, generate hypotheses, and iteratively improve outputs, turning biomolecule design into a more automated pipeline. Open source vs closed ecosystems (Priority: 3/5): Much of the most capable tooling remains closed or fragmented; open-source efforts are still assembling models manually, while companies and some NVIDIA API offerings are pushing integrated workflows.

Key Arguments: Protein engineering is shifting from a single-model paradigm to chained workflows where structure prediction, sequence design, dynamics modeling, and validation each have distinct roles. AlphaFold 3 is powerful for interaction prediction because it incorporates multiple biomolecule types, but commercial restrictions limit its usefulness in drug discovery pipelines. Structure remains important for protein-protein interaction screening; sequence alone is often insufficient to capture geometry-dependent binding. The biggest unmet need is dynamics: many enzyme and binding problems require conformational ensembles, not just static structures. MDGen could become highly impactful because it may approximate molecular dynamics much faster while adding useful capabilities like trajectory interpolation and inpainting. Peptide and antibody design are converging problems, especially for disordered regions and CDR loops, suggesting shared model architectures and training data could improve both. Agentic orchestration of multiple models is likely to be the next major leap because it can automate hypothesis generation, design iteration, and workflow tuning. Wet-lab validation remains essential; strong in-silico metrics are not enough to make these models truly useful in practice.

Data Points: Time since previous deep dive: about six months - The guest and host compare the field’s progress since their last conversation. Protein-protein interaction screening scale: 10,000 to 100,000 candidates - Binder design workflows can generate and screen very large numbers of candidates. Human proteome pairwise interactions: all possible pairwise interactions - A stripped-down RosettaFold model was used to predict pairwise interactions across the human proteome. MSA computational bottleneck: MSA generation is the slow step - Schreiber emphasizes that multiple-sequence-alignment construction is the main latency source in some pipelines. Radius relevant to catalysis: about 20 angstroms - Dynamics around catalytic sites within this distance are described as important for enzyme function. EnzymeFlow output: de novo catalytic pockets - A new flow-matching model generates catalytic pockets for enzyme design. PepFlow sequence length: less than 30 residues - The peptide models focus on short peptides, which are often disordered and difficult to model. Model categories in ESM3 function vocabulary: a few hundred tokens - Schreiber criticizes the finite function vocabulary as too limiting compared with open-vocabulary approaches. MDGen speed: around 1,000x faster than molecular dynamics - The guest describes the potential throughput advantage of trajectory generation models. AlphaFold 2/3 comparison: as good as AlphaFold 2 at structure prediction - ESM3 is described as matching AlphaFold 2-level structure prediction in this discussion. AlphaProteo success rate: 80-something percent on one target; 9% on another - An example of highly target-dependent success rates in binder design. AUC metric reference: area under the precision-recall curve - Used to evaluate protein-protein interaction prediction quality. Nobel Prize context: 5 laureates won Nobels in large part because of AI techniques - The host notes a broader Nobel signal for AI across physics and chemistry.

Pivotal Quotes: "If you can scale that process and have an agent drive a big complicated workflow and solve a task, and you can just churn out molecules, that changes things a lot." — Amelie Schreiber: On the long-term importance of agentic, multi-model biomolecule design workflows "The frontier is dynamics versus statics." — Amelie Schreiber: On what remains unsolved in protein and enzyme design "I think the measure of success is a well-designed molecule. And so, like, my success measure would be measured in the wet lab, not on the computer screen." — Amelie Schreiber: On how to evaluate AI biology systems in practice

Implications: The field is moving toward automated, high-throughput biomolecular design pipelines that could accelerate therapeutics and industrial enzymes. Near-term wins will likely come from agentic workflows plus better dynamics models, but wet-lab validation and open tooling remain key bottlenecks.