2026/10/07 – Minhuan Li

From Prediction to Scientific Inference: Grounding Generative Models in Biological Experiments

Biological measurements and predictive models are both scaling rapidly, yet the interface between them remains a bottleneck. In structural biology, deposited structures have enabled major advances in AI but capture only part of the underlying experimental evidence. Connecting models more directly to measurements could better exploit information about conformational populations, experimental context, and uncertainty.

This talk presents experiment-grounded generative inference: combining learned priors with measurements to infer sample-specific biological states. I will develop these ideas primarily in structural biology, where the states of interest are biomolecular conformations. I will focus on ROCKET, which uses AlphaFold as a learned prior for inference from X-ray crystallography and cryo-EM measurements, and EmbedOpt, which steers conditional protein diffusion models by optimizing their conditioning embeddings at inference time. Both approaches use pretrained models as structured search spaces, allowing experimental evidence to guide inference beyond a model’s default prediction. I will also briefly discuss GOTO, an optimal-transport-based objective that provides informative optimization signals for joint inference of molecular states and cryo-EM imaging parameters.

I will close with an outlook on MIMIC and possible routes for making measurements first-class inputs and training signals in multimodal biomolecular models. The broader goal is to use learned priors to interpret experimental evidence—and ultimately use that evidence to improve the priors themselves.

References:
Golkar, Li, et al. — MIMIC (arXiv, 2026)
Fadini, Li et al. — AlphaFold as a prior (Nature Methods, 2026)
Li et al. — EmbedOpt (arXiv, 2026)
Woollard, Li et al. — GOTO (bioRxiv, 2025)