TeBaAb: Text-Based Antigen-Conditioned Antibody Redesign via Directed Evolution

An earlier version of this work was presented at NeurIPS 2025 workshops as a non-archival presentation:

• AI for Science: https://openreview.net/pdf?id=Imw5NGMgje
• Machine Learning and the Physical Sciences: https://ml4physicalsciences.github.io/2025/files/NeurIPS_ML4PS_2025_107.pdf

Abstract

The design of antibodies with high affinity and specificity for target antigens is a cornerstone of therapeutic and diagnostic innovation. Traditional optimization strategies, such as phage or yeast display and directed evolution, remain resource-intensive and limited in their ability to integrate contextual information. Recent AI-driven approaches have accelerated protein engineering, but most rely exclusively on structured inputs, overlooking the potential of natural language as a flexible design interface. In this work, we introduce TeBaAb, a novel text-based antigen-conditioned framework for antibody redesign that combines generative modeling with iterative optimization inspired by directed evolution. TeBaAb integrates a Conditional Variational Autoencoder (CVAE) jointly conditioned on antigen sequences and textual descriptions of antibody properties, coupled with a two-stage binding affinity predictor and an iterative enrichment loop. To support this approach, we curated AbDes, a new dataset of 7,684 text–antibody–antigen pairs with accompanying structural and binding information.

In silico experimental evaluations demonstrate that TeBaAb improves the predicted binding affinity by an average of $10%$ compared to the original antibodies, while preserving structural confidence (RMSPE $<$ 1.0 Å) and generating sequences that are diverse and novel. By enabling text-conditioned antigen-specific antibody design, TeBaAb provides a promising new paradigm for accelerating therapeutic antibody discovery and expanding the antibody design space beyond traditional methods.

Table of Contents

• Installation
• Usage
  • Datasets
  • Configuration
  • Training
  • Antibody Design
• Evaluation

---

Installation

Requirements

• Python 3.10 or higher
• A virtual environment (e.g., venv or conda) is recommended
• Dependencies listed in requirements.txt

Steps

1. Clone the repository:

   git clone https://github.com/HySonLab/TeBaAb.git
   cd TeBaAb

2. Set up a virtual environment and install dependencies:

   conda env create -f environment.yml 

   or

   python -m venv venv
   source venv/bin/activate  # On Windows: venv\Scripts\activate
   pip install -r requirements.txt

3. Create the checkpoints directory:

   mkdir -p checkpoints

---

Usage

Datasets

• AbDes: Contains antibody-antigen sequences paired with text descriptions.
  • Download: Hugging Face: AbDes
  • Path: datasets/abdes

Ensure datasets are placed in the specified paths or update configuration files accordingly.

Configuration

Configuration is managed via YAML files in configs/:

• training.yaml: Defines architectures for protein/text encoders, CVAE, and fitness predictor.
• optimize.yaml: Configures directed evolution parameters (e.g., generations, mutation rate).

Preparing Data

Extract and preprocess the dataset:

cd datasets

python3 extract_embedding.py \
    --input_csv ./abdes/train.csv \
    --output_dir ./cvae \
    --output_prefix train \
    --modality all \
    --embedding_type pooler \
    --device cuda:0 \
    --esmc_cache /path/to/esmc_300m_2024_12_v0.pth

Training

1. Train the CVAE:

   ./scripts/run_train_cvae.sh

   • Trains on datasets/cvae.

2. Train the Oracle:

   python3 scripts/train_predictor.py

   • Trains on datasets/affinity.

Antibody Design

Generate and optimize protein sequences using directed evolution:

./scripts/optimize.py

---

Evaluation

Training Metrics

• CVAE: Reconstruction loss, KL divergence, validation loss.
• Oracle: Mean Squared Error (MSE) for fitness prediction.

Antibody Design Metrics

• Binding affinity scores from the oracle.
• Diversity, novelty.
• 3D Structure Error: ABodyBuilder2
• Antibody Developability: TAP

---

Please cite our work

@inproceedings{
nguyen2025tebaab,
title={TeBaAb: Text-Based Antigen-Conditioned Antibody Redesign via Directed Evolution},
author={Cuong Manh Nguyen and Huy-Hoang Do-Huu and Viet Thanh Duy Nguyen and Truong-Son Hy},
booktitle={NeurIPS 2025 AI for Science Workshop},
year={2025},
url={https://openreview.net/forum?id=Imw5NGMgje}
}