IC-227MSA-based conformation generation methods (AlphaFlow, MSA-subsampling) outperform sequence-based methods (EigenFold, STR2STR, ESMFlow) on conformation changing pair generation

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

SourceStructure Language Models for Protein Conformation Generation

On both the apo/holo (90 pairs) and fold-switching (77 pairs) benchmarks, MSA-based methods achieve higher global residue flexibility correlations than sequence-based methods. For apo/holo, MSA-subsampling achieves 0.398 and AlphaFlow 0.455, versus EigenFold 0.126, STR2STR 0.174/0.148, and ESMFlow 0.416. For fold-switching, MSA-subsampling achieves 0.350 and AlphaFlow 0.385, versus EigenFold 0.225, STR2STR 0.161/0.111, and ESMFlow 0.269. The authors conclude this highlights the importance of MSA for generating stable conformation changing protein targets.

Evidence

correlational

Key metric

Apo/holo resflex r (gl.): MSA-subs 0.398, AlphaFlow 0.455, EigenFold 0.126, STR2STR 0.174/0.148, ESMFlow 0.416; Fold-switch resflex r (gl.): MSA-subs 0.350, AlphaFlow 0.385, EigenFold 0.225, STR2STR 0.161/0.111, ESMFlow 0.269

Caveat

ESMFlow (0.416 apo/holo) narrows the gap with MSA-based methods, and the comparison is on only 5 structures per target (few-shot ensemble).

Model

AlphaFlow, AlphaFold2, EigenFold, STR2STR, ESMFlow

Datasets

Apo/Holo Pairs [eval], Fold-Switching (Chakravarty & Porter 2022) [eval]

Related findings

IC-225, IC-226, IC-228

Extraction

automatic-extraction