ESM-MSR

LoRA adapters for ESM-MSR, a predictor of protein stability changes (ฮ”ฮ”G, stabilizing-positive) for single and multiple mutations, trained on ESM3-small-open (esm3-sm-open-v1). The code, inference instructions and benchmarks are in the GitHub repository (SKTeamLab/esm-msr).

Files

File Contents
lora_seed1.safetensors, lora_seed2.safetensors, lora_seed3.safetensors The three training seeds of the release recipe (FINAL_p0): WT adapter (rank 2, alpha 4) and MT adapter (rank 16, alpha 16), plus the four calibration scalars. float32, 588 tensors each.
hparams.yaml The training configuration shared by the three seeds; the loader reads it next to the weights. The seeds differ only in seed.
splits/splits_oct06_capped.pkl The training / validation / test split the seeds were trained on (--split splits_oct06_capped reads it from data/).
splits/splits_oct06_capped.csv The same split as text: the structure files in each partition.
splits/splits_oct06_capped_library_assignment.csv One row per MegaScale library: single and double counts, homology cluster, the reason for any exclusion, and the partition.

Each seed was trained on the homology-restricted MegaScale split (splits_oct06_capped), with both passes unmasked. The checkpoint kept for each seed is the epoch with the best validation val_unmasked_rho_combined_avg:

Seed Epoch Validation ฯ (combined) SHA-256 of the safetensors
1 6 0.8558 ec8ca9bfca71f3b8ffa071c8e8d800de59290307093fd5c9ec5ae091fc3dfc75
2 8 0.8552 2823187f44689024d2b2da277cada03ebea73b910c991d8a5162966c0607573e
3 6 0.8574 6fcb0c225127b49d924084f1c24751874d0a4bf33a4e36f37efede9998426dd2

The safetensors files were converted from the Lightning checkpoints with inference_scripts/convert_lora_checkpoint.py. Every tensor equals the checkpoint's bit for bit, and inference from either file gives identical predictions.

Use

huggingface-cli download sareeves96/esm-msr --local-dir LoRA_models/esm-msr
python src/esm_msr/inference.py --checkpoint_path LoRA_models/esm-msr/lora_seed1.safetensors \
    --pdb_file structure.pdb --chain A --mode singles --output_csv predictions.csv

Seeds can be scored separately and averaged. --skip_reverse reads the WT adapter alone (additive, much faster); without it, multi-mutants use the combined WT + MT prediction.

Earlier releases

The previous adapters, from the earlier training recipe on the legacy hyperopt split, are kept on the archive-2026-11 branch.

Downloads last month
-
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for sareeves96/esm-msr

Adapter
(1)
this model