IC-225ESM3 (1.4B) performs zero-shot protein conformation generation with competitive quality across BPTI dynamics, conformation changing pairs, and intrinsically disordered proteins

Jiarui Lu, Xiaoyin Chen, Stephen Zhewen Lu, Chence Shi, Hongyu Guo, Yoshua Bengio, Jian Tang

SourceStructure Language Models for Protein Conformation Generation

The paper evaluates the pre-trained ESM3 1.4B model in zero-shot mode (no fine-tuning) for protein conformation generation using iterative decoding. On BPTI, ESM3 achieves JS-PWD 0.406, JS-TIC 0.445, validity 0.800, and TM-ens 0.842, competitive with diffusion-based baselines. On conformation changing pairs, it achieves global residue flexibility correlation of 0.312 (apo/holo) and 0.388 (fold-switch). On 114 IDP targets from PED, it achieves MAE of 6.606/4.301 (pairwise distance) and 0.249/0.174 (contact map), the best among all methods tested.

Evidence
correlational
Key metric
BPTI: JS-PWD 0.406, JS-TIC 0.445, JS-RG 0.561, validity 0.800, TM-ens 0.842, RMSD-ens 1.450; Apo/holo: resflex r (gl.) 0.312, TM-ens 0.839/0.876; Fold-switch: resflex r (gl.) 0.388, TM-ens 0.627/0.717; IDP: pairwise distance MAE 6.606/4.301, radius of gyration 4.346/2.329, contact map 0.249/0.174
Caveat
ESM3 uses the DVAE structure tokenizer from Hayes et al. (2024) which has non-negligible reconstruction error (RMSD > 0.5A) for a fraction of fold-switching and IDP targets, limiting the quality of the latent representation.
Model
ESM3
Datasets
Apo/Holo Pairs [eval], Fold-Switching (Chakravarty & Porter 2022) [eval], PED [eval]
Related work
AlphaFlow [compared-to], EigenFold [compared-to], STR2STR [compared-to], ESMFlow [compared-to]
Related findings
IC-226, IC-227, IC-228
Extraction
automatic-extraction