Biotrainer

Biological prediction models made simple.

Overview

Biotrainer is an open-source framework that simplifies machine learning model development for protein analysis. It provides the following modules:

• Training: Easy to use training and inference pipelines for protein feature prediction
• Embedding: Built-in support for protein language models (ProtT5, ESM-2, ESM-C, ...) and baselines (blosum62, one_hot_encoding, ...)
• BioEngineer: Zero-shot predictions and mutation generation of protein sequences
• AutoEval: Automated evaluation of protein language models on curated downstream tasks

Quick Start

1. Installation

Install using pip:

pip install biotrainer

Manual installation using uv:

# First, install uv if you haven't already:
pip install uv

# Create and activate a virtual environment
uv venv
source .venv/bin/activate  # On Unix/macOS
# OR
.venv\Scripts\activate  # On Windows

# Basic installation
uv pip install -e .

# Installing with jupyter notebook support:
uv pip install -e ".[jupyter]"

# Installing with onnxruntime support (for onnx embedders and inference):
uv pip install -e ".[onnx-cpu]"    # CPU version
uv pip install -e ".[onnx-gpu]"    # CUDA version
uv pip install -e ".[onnx-mac]"    # CoreML version (for Apple Silicon)

# You can also combine extras:
uv pip install -e ".[jupyter,onnx-cpu]"

# For Windows users with CUDA support:
# Visit https://pytorch.org/get-started/locally/ and follow GPU-specific installation, e.g.:
pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118

2. Basic Usage

# Training
biotrainer train --config examples/sequence_to_class/config.yml

# Inference after Training
python3
>>> from biotrainer.training import BiotrainerModel
>>> model = BiotrainerModel.from_training_result("output/out.yml")
>>> predictions = model.predict("SEQUENCE")

# AutoEval
autoeval_report = (AutoEval(embedder_name="facebook/esm2_t6_8M_UR50D", development_mode=True).
                    pbc_supervised_contact().
                    pbc_zeroshot_contact().
                    pbc_supervised().
                    pgym(zero_shot_method=ZeroShotMethod.MASKED_MARGINALS).
                    run())
autoeval_report.summary(development_mode=False)

3. Quick Start Datasets

• Subcellular Localization Prediction
  • Protocol: sequence_to_class/residues_to_class
  • Citations and Download
• Secondary Structure Prediction
  • Protocol: residue_to_class
  • Citations and Download

Features

Supported Training Protocols

• Residue-level classification (residue_to_class)
• Residue-level regression (residue_to_value) [BETA]
• Sequence-level classification (sequence_to_class)
• Sequence-level regression (sequence_to_value)
• Residues-level classification (residues_to_class, like sequence_to_class with per-residue embeddings)
• Residues-level regression (residues_to_value, like sequence_to_value with per-residue embeddings)

Autoeval

The biotrainer autoeval module allows automatical evaluation of a protein language model on downstream tasks. You can find public results (wip!) on the autoeval dashboard and compare them to your own. Learn more in the autoeval examples.

Documentation

Tutorials

• First Steps Guide
• Interactive Training Tutorial
• Config Options Overview
• Biocentral Web Interface

Detailed Guides

• Data Standards
• Configuration Options
• Troubleshooting

Example Training Configuration

protocol: residue_to_class
input_file: input.fasta
model_choice: CNN
optimizer_choice: adam
learning_rate: 1e-3
loss_choice: cross_entropy_loss
use_class_weights: True
num_epochs: 200
batch_size: 128
embedder_name: Rostlab/prot_t5_xl_uniref50

Docker Support

# Run using pre-built image
docker run --gpus all --rm \
    -v "$(pwd)/examples/docker":/mnt \
    -u $(id -u ${USER}):$(id -g ${USER}) \
    ghcr.io/sacdallago/biotrainer:latest /mnt/config.yml

More information on running docker with gpus: Nvidia container toolkit

Getting Help

• Check our Troubleshooting Guide
• Create an issue
• Visit biocentral.cloud

Citation

@inproceedings{
sanchez2022standards,
title={Standards, tooling and benchmarks to probe representation learning on proteins},
author={Joaquin Gomez Sanchez and Sebastian Franz and Michael Heinzinger and Burkhard Rost and Christian Dallago},
booktitle={NeurIPS 2022 Workshop on Learning Meaningful Representations of Life},
year={2022},
url={https://openreview.net/forum?id=adODyN-eeJ8}
}