AI models sample more protein folding stages, but miss folded endpoints

Across 20 human proteins, the authors report roughly tenfold broader coverage of a folding coordinate than varied-seed AlphaFold 3 predictions, while AlphaFold 3 better captures the fully folded neighborhood.

Independent

Lisheng Zhang · Qingsong Li · Zinan Xie · Yanting Yang · Qikui Yang · Peng Lu · +4 more

Research Digest·4 November 2026·2 min read

Illustration: AI models sample more protein folding stages, but miss folded endpoints

In one line

FoldScape predicts protein structures at any folding stage from sequence, with constant accuracy across the folding coordinate.

Read it as

Proteins fold from a loose chain into a precise native shape through many intermediate stages. Current AI predictors only output the final folded structure, missing these intermediates that are crucial for protein function. Simulating the folding process with computers is extremely slow and expensive.

FoldScape is a deep learning model trained on simulations of folding and unfolding. It takes the protein's amino acid sequence and a desired folding fraction Q, a number between 0 (unfolded) and 1 (folded), and generates a 3D structure at that stage.

Compared to AlphaFold 3, FoldScape covers about ten times more of the folding coordinate, and its prediction accuracy remains constant across all Q values, not just near the native state. This allows studying protein folding dynamics without costly simulations, aiding understanding of diseases and protein design.

How it works

stage by stage

RSS

The Daily Edition

The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.

Concepts

flow matching conformational ensemble folding reaction coordinate hierarchical distance regression sliding-window crop-and-superpose coarse-grained molecular dynamics