PhosFate: A Framework for Phosphate Binding Site Prediction Based on ESM2 Protein Embeddings

PhosFate is a machine learning framework that uses ESM2 protein language model embeddings to predict and classify anion binding sites, focusing on phosphate recognition and providing data-driven insights for biomolecular design and nutrient recovery. Maintained by ChowdhuryRatul, this project was released on GitHub on May 27, 2026 (link: https://github.com/ChowdhuryRatul/PhosFate). Its core value lies in combining large-scale protein language models to address the limitations of traditional methods.

Protein-anion interactions are critical in biological systems. Phosphate is involved in processes such as energy metabolism and signal transduction. Accurately predicting its binding sites is of great significance for understanding functions, designing biomolecules, and nutrient recovery. Traditional experimental methods (X-ray crystallography, nuclear magnetic resonance) are costly and time-consuming; computational methods face challenges like protein structure diversity and complex microenvironments of binding sites. The rise of protein language models provides a solution.

The core innovation of PhosFate is using ESM2 protein language model embeddings from Meta AI, combined with downstream classification models to achieve high-precision prediction. The framework components include: the Scripts directory (data processing and training scripts), the Utils directory (auxiliary tools), backend (backend inference service), frontend (user interface), and phosfate_inference_code.ipynb (inference example notebook).

Advantages of ESM2 embeddings: 1. Encodes evolutionary constraint information, linking structure and function; 2. Context-aware, capturing local patterns; 3. Strong generalization ability, suitable for multiple protein families. Prediction workflow: Input protein sequence → ESM2 generates residue embeddings → Classification model processing → Output binding site probability scores.

Biomolecular design: Guides rational design of phosphate-binding proteins, optimizes enzyme activity, and develops biosensors; Nutrient recovery: Optimizes phosphate capture proteins for wastewater, designs bio-extraction systems for low-grade phosphate ores, and supports agricultural conversion of organic waste phosphorus cycling.

Framework features: Modular architecture for easy expansion and maintenance, front-end and back-end separation supporting web deployment and API calls, and complete inference examples to lower the entry barrier. Usage: Quickly set up via a conda environment (environment.yml), open-source under the MIT license, and easy to get started with via Jupyter Notebook.

PhosFate is a successful practice of applying protein language models to bioinformatics, combining ESM2's representation capabilities with classification models to provide an efficient and accurate solution for anion binding site prediction. In the future, with model development and data accumulation, similar tools are expected to improve accuracy and applicability, opening up new possibilities for life science research and biotechnology.