IC-228ESM3 (1.4B) fails to capture MD ensemble statistics on the ATLAS benchmark, achieving pairwise RMSD correlation of only 0.08

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

SourceStructure Language Models for Protein Conformation Generation

On the ATLAS MD ensemble test set (100ns all-atom simulations, 1390 monomeric targets), ESM3 in iterative decoding mode achieves very low correlations with ground-truth MD statistics: pairwise RMSD r=0.08, global RMSF r=0.19, per-target RMSF r=0.67, RMWD 7.27, and PC similarity 22%. The authors attribute this to ATLAS's short 100ns timescale rarely capturing substantial conformational changes, and to the categorical nature of structure tokens making it difficult to represent subtle structural changes. AlphaFlow-MD achieves substantially higher correlations (pairwise RMSD r=0.48, global RMSF r=0.60).

Evidence
correlational
Key metric
ESM3 (id): pairwise RMSD r 0.08, global RMSF r 0.19, per-target RMSF r 0.67, RMWD 7.27, RMWD trans 5.22, RMWD var 4.35, MD PCA W2 2.06, joint PCA W2 5.97, PC sim 0.5% 22, weak contacts J 0.45, transient contacts J 0.26; AlphaFlow-MD: pairwise RMSD r 0.48, global RMSF r 0.60, per-target RMSF r 0.85
Caveat
The authors note the 100ns simulation timescale rarely captures slower, substantial conformational changes, which particularly challenges tokenized structure approaches; this may be a limitation of the benchmark rather than the model.
Model
ESM3
Concepts
Failure mode
Datasets
ATLAS [eval]
Related work
AlphaFlow [compared-to]
Related findings
IC-225, IC-226, IC-227
Extraction
automatic-extraction