monoyield-predictor

monoyield-predictor

Predicts how much of a protein sequence E. coli will make as a single-unit monomer, in log10 mg/L. Trained LoRA adapters and regression heads for ESMC-6B and SaProt, for protein researchers.

Made by Tsuboyama-Lab on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: monoyield-predictor holds the trained weights of a model that predicts the monomeric expression yield of a protein sequence in E. coli, on a log10 mg/L scale. The files are LoRA adapters and regression heads, 7 to 12 MB per fold. The frozen base models are not included, since the code pulls them from Hugging Face at run time.

Who it suits: Protein researchers who want to estimate expression yield before running an experiment. The weights come with a bioRxiv preprint and a code repository.

Industry | Who uses it | How they use it | Example
Biotech | Protein engineers | Rank designed sequences by expected E. coli yield | Score a CSV of candidate sequences
Academic labs | Protein researchers | Check a predictor against their own measurements | Compare predicted and measured yield
Antibody work | Nanobody developers | Predict monomer yield or the monomer to dimer ratio for nanobodies | Score a set of nanobody sequences
Research tools | Computational biologists | Use the adapters inside their own pipelines | Call the prediction script on new data

People without a large GPU, or who do not work with protein sequences, can skip it.

Where to use it:

• A Python environment on Windows, macOS or Linux
• A computer with a GPU of about 48 GB for fp32, or about 24 GB with bfloat16
• The companion code repository, which holds the scripts and example files

Use case | Who | What they get back
De novo and mutation scans | Protein engineers | A predicted log10 monomer yield for each sequence
Nanobody yield | Nanobody developers | A predicted log10 monomer yield for each nanobody
Nanobody dimer ratio | Nanobody developers | A predicted log10 monomer to dimer ratio

How to set it up:

• Python and pip
• A GPU with enough memory: about 48 GB for fp32, and about 24 GB for bfloat16, which the card says should be validated
• The companion code repository, which the card names

1. Install the Hugging Face hub tools and download these weights if you want a local copy.
2. Get the companion code repository and move into its folder.
3. Install it in editable mode.
4. Run its download script to fetch the weights.

pip install -U huggingface_hub
hf download Tsuboyama-Lab/monoyield-predictor

cd monoyield-predictor
pip install -e .
HF_REPO=Tsuboyama-Lab/monoyield-predictor bash download_weights.sh weights

How to use it: Run the prediction script on a CSV of sequences, with a weights folder and the base model it was trained on. Each folder holds five fold checkpoints, and a prediction is the mean of the five. The code repository lists all models, input formats and training commands.

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

Tool | What it does
esmc_monoyield | ESMC-6B adapter that reads an amino acid sequence and predicts log10 monomer yield
saprot_lora | SaProt adapter that reads a sequence plus Foldseek 3Di and predicts the same yield
esmc_nb_monoyield | ESMC-6B adapter for nanobody sequences, predicting log10 monomer yield
esmc_nb_mdratio | ESMC-6B adapter for nanobody sequences, predicting the log10 monomer to dimer ratio

Example: You point the script at the example library file and the esmc_monoyield folder. It scores each sequence with the ESMC-6B model and the five fold checkpoints, then writes the predictions to predictions.csv.

For newcomers and experts: This is research software for protein work, not a point and click app. A careful first step is reading the card's file table, then running the example CSV once you have a GPU with enough memory.

Under the hood, each adapter adds low-rank matrices and a regression head to a frozen base model. Four variants cover de novo and mutation scan yield, a structure-aware SaProt version and two nanobody targets. Averaging five folds smooths out the variation between checkpoints.

Sequences go in as a CSV file
The frozen base model reads each sequence
The adapter and head give a score per fold
The five folds are averaged
Predictions are written to a CSV file

Good to know:

• Hardware: Full fp32 precision with ESMC-6B calls for a GPU holding roughly 48 GB. With bfloat16 it fits in 24 GB, though the card says to check its results first.
• Base terms: The adapters are derivatives of ESMC-6B and SaProt, so you must follow those models' terms too.
• ESMC-6B terms: an MIT license plus third-party notices, and the acceptable use policy from Biohub. The SaProt model uses the MIT license.
• License: Apache 2.0 for the adapters.
• Cost to run: none named. You need a GPU with enough memory, and no paid service appears in the source.
• Data it reaches: Your sequence CSV, the base model files from Hugging Face and the companion code repository.
• Not for: The source names no other forbidden uses. Follow the Biohub acceptable use policy that comes with ESMC-6B.

On Hugging Face, 13 November 2026

• Library: peft
• Downloads in the last 30 days: 0
• Likes: 3
• License: apache-2.0
• Based on: biohub/ESMC-6B
• Tags: peft, protein, protein-language-model, lora, expression-yield, nanobody

Links

• Model: https://huggingface.co/Tsuboyama-Lab/monoyield-predictor
• Files: https://huggingface.co/Tsuboyama-Lab/monoyield-predictor/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download Tsuboyama-Lab/monoyield-predictor

Then follow the usage in its model card. Nothing in this kit runs by itself.