---

Abstract

There is significant interest in targeting disease-causing proteins with small molecule inhibitors to restore healthy cellular states. The ability to accurately predict the binding affinity of small molecules to a protein target in silico enables the rapid identification of candidate inhibitors and facilitates the optimization of on-target potency. In this work, we present T-ALPHA, a novel deep learning model that enhances protein-ligand binding affinity prediction by integrating multimodal feature representations within a hierarchical transformer framework to capture information critical to accurately predicting binding affinity. T-ALPHA outperforms all existing models reported in the literature on multiple benchmarks designed to evaluate protein-ligand binding affinity scoring functions. Remarkably, T-ALPHA maintains state-of-the-art performance when utilizing predicted structures rather than crystal structures, a powerful capability in real-world drug discovery applications where experimentally determined structures are often unavailable or incomplete. Additionally, we present an uncertainty-aware self-learning method for protein-specific alignment that does not require additional experimental data, and demonstrate that it improves T-ALPHA’s ability to rank compounds by binding affinity to biologically significant targets such as the SARS-CoV-2 main protease and the epidermal growth factor receptor. To facilitate implementation of T-ALPHA and reproducibility of all results presented in this paper, we have made all of our software available at this repository.

For full details of T-ALPHA, please refer to the preprint on bioRxiv.

Table of Contents

1. Installation
2. Accessing Data Files
3. Running the Scripts
   • Training the Model
   • Performing Inference
   • Monte Carlo Dropout and Semi-Supervised Fine-Tuning
4. On-the-Fly Inference Notebook
5. Contact

Installation and Setup

1. Clone the Repository

git clone https://github.com/gregory-kyro/T-ALPHA.git
cd T-ALPHA

2. Set Up a Virtual Environment (Recommended)

python -m venv venv
source venv/bin/activate    # On Linux/Mac
venv\Scripts\activate       # On Windows

3. Install Dependencies

pip install -r requirements.txt

Accessing Data Files

The pre-trained model parameters and test datasets are hosted on Zenodo for easy access.

1. Download the model parameters and test files using the following links:

• Model Parameters: Download Here
• Test Files: Download Here

2. Place the downloaded files in the data/ directory of the repository:

T-ALPHA/
├── data/
│   ├── model_parameters.tar.gz
│   ├── test_files.tar.gz

3. Extract the files. For Linux/macOS, you can use:

tar -xvzf test_files.tar.gz
tar -xvzf model_parameters.tar.gz

Running the Scripts

This repository provides three primary scripts to train the model, perform inference, and run Monte Carlo Dropout for uncertainty estimation.

1. Training the Model

To train the model from scratch, run:

python scripts/train.py --train_set <path_to_train_hdf> \
                        --val_set <path_to_val_hdf> \
                        --checkpoint_dir <path_to_save_checkpoints> \
                        --save_dir <path_to_logs> \
                        --save_name <experiment_name>

2. Performing Inference

To evaluate the model on test data and compute key metrics, run:

python scripts/inference.py --ckpt_path <path_to_checkpoint> \
                            --test_set_path <path_to_test_hdf>

This will produce:

• A CSV file with predictions
• A scatter plot of predictions vs. true values

3. Estimating Uncertainty via Monte Carlo Dropout

To perform uncertainty estimation with Monte Carlo Dropout, run:

python scripts/mc_dropout.py --checkpoint <path_to_checkpoint.ckpt> \
                             --test_set <path_to_test_dataset> \
                             --output_results <path_to_save_results.csv> \
                             --output_weighted <path_to_save_weighted_results.csv>
                                

On-the-Fly Inference Notebook

The T-ALPHA Inference Notebook enables real-time protein-ligand binding affinity prediction. It supports:

• Sequence-Based Inputs: Protein sequence and ligand SMILES.
• Structural Inputs: Protein PDB and ligand SDF files.
• PDB ID-Based Inputs: Fetch protein sequence automatically using a PDB ID.

With minimal setup, you can quickly obtain pK_d predictions. Try it now here

Contact

For questions, issues, or collaborations, feel free to reach out:

• Name: Gregory W. Kyro
• Email: [email protected]