Model card: Apache-2.0, GitHub URL tsubo-lab

ef965ba verified 6 days ago

2.39 kB

license: apache-2.0
library_name: peft
base_model:
  - biohub/ESMC-6B
  - westlake-repl/SaProt_1.3B_AFDB_OMG_NCBI
tags:
  - protein
  - protein-language-model
  - lora
  - expression-yield
  - nanobody

monoyield-predictor: trained weights

LoRA adapters and regression heads that predict the monomeric expression yield in E. coli (log10 mg/L) of a protein sequence. Companion weights of the paper [title - to be filled] and the code repository https://github.com/tsubo-lab/monoyield-predictor.

The files contain only the trained LoRA matrices and the regression head (7-12 MB per fold). The frozen base models are downloaded from Hugging Face when the code runs.

folder	model in the paper	base model	input	target
`esmc_monoyield/`	ESMC-MonoYield	`biohub/ESMC-6B`	amino-acid sequence	log10 monomer yield (mg/L), de novo + DMS
`saprot_lora/`	SaProt LoRA	`westlake-repl/SaProt_1.3B_AFDB_OMG_NCBI`	amino acid + Foldseek 3Di	same
`esmc_nb_monoyield/`	ESMC-Nb-MonoYield	`biohub/ESMC-6B`	nanobody sequence	log10 monomer yield (mg/L)
`esmc_nb_mdratio/`	ESMC-Nb-MDRatio	`biohub/ESMC-6B`	nanobody sequence	log10 monomer/dimer ratio

Each folder holds five fold checkpoints, <model>_1.pt ... <model>_5.pt; predictions are the mean of the five.

Usage

git clone https://github.com/tsubo-lab/monoyield-predictor
cd monoyield-predictor
pip install -e .
HF_REPO=Tsuboyama-Lab/monoyield-predictor bash download_weights.sh weights
python scripts/predict_esmc_mlp_head.py --external_csv examples/example_main_library.csv \
    --model_dir weights/esmc_monoyield --model_name biohub/ESMC-6B \
    --model_revision 45b0fa5d7fb06faefbd5e3b89bdcef35d564e79a --dtype float32 --batch_size 1 --out_csv predictions.csv

See the code repository for all models, input formats and training commands. ESMC-6B in fp32 needs a GPU with about 48 GB of memory (--dtype bfloat16 fits 24 GB but should be validated).

License and terms

The adapters are released under the Apache License 2.0 (Copyright 2026 Kotaro Tsuboyama and Selena Setsu). They are derivatives of the base models and must be used under the terms of those models: ESMC-6B (MIT with third-party notices and the Biohub Acceptable Use Policy, https://biohub.org/acceptable-use-policy/) and SaProt (MIT).

Citation

[to be filled]