abdix

Deep learning models developed for the 2025 Antibody Developability Prediction Competition.

Overview

abdix provides a few models for predicting antibody developability properties directly from sequence data. It integrates multiple feature modalities, including protein language models, handcrafted biochemical features, and metadata about antibody subclasses.

The five target properties are:

• Polyreactivity
• Titer
• HIC
• Self-association
• Thermostability

Features

• Pre-trained protein language models (p-IgGen, ESM-C)
• Handcrafted sequence descriptors
• CDR-specific annotations
• Metadata encodings (IgG subtype, light-chain type)

Architectures

• Modality-specific projections into a shared representation space
• Transformer encoders for cross-modal reasoning
• Task-specific attention pooling with independent regression heads

Models

abdix includes three model variants designed to trade off accuracy and computational cost:

Model	Architecture	Parameters
regressor_pooled.py	Minimal MLP baseline	~100K
regressor_lite.py	Lightweight transformer	~1–2M
regressor.py	Full multi-modal transformer	~6–7M

Results

Training a single multi-task model on a small dataset (N ≈ 200) proved challenging. On the held-out test set, predictions for polyreactivity ranked among the top 10, while predictions for Titer, HIC, Self-association, and Thermostability showed negative correlations.

Usage Example

from abdix.model.regressor import AntibodyRegressor
from abdix.features.base import PIgGenFeaturizer, ESMCFeaturizer

# Initialize the model
model = AntibodyRegressor(
    featurizers={
        'piggen': PIgGenFeaturizer(),
        'esmc': ESMCFeaturizer(),
    },
    property_names=['Polyreactivity', 'Titer', 'HIC', 'Self-association', 'Tm'],
)

# Predict
predictions = model(batch)  # Shape: (batch_size, 5)

Datasets

• GDPa1

References

• Ginkgo Antibody Developability Benchmark
• ESM
• p-IgGen
• abdev-benchmark