Monte-Carlo dropout applied only at inference improves zero-shot protein fitness predictions for ESM2 models on a subset of ProteinGym, with dropout 0.1 working across model sizes.
Proteingym: Large-scale benchmarks for protein fitness prediction and design
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Protein Language Model Zero-Shot Fitness Predictions are Improved by Inference-only Dropout
Monte-Carlo dropout applied only at inference improves zero-shot protein fitness predictions for ESM2 models on a subset of ProteinGym, with dropout 0.1 working across model sizes.