Illuminating the Druggable Human Proteome with an AI Protein Profiling Platform

Area of Science:

• Biochemistry
• Computational Biology
• Drug Discovery

Background:

• Proteomic approaches face limitations in coverage and data heterogeneity.
• Existing machine learning (ML) models struggle with structural dependencies and varied experimental labels.

Purpose of the Study:

• To develop a multimodal AI platform (AiPP) for predicting and characterizing protein ligand interaction sites directly from amino acid sequences.
• To create a comprehensive, proteome-wide atlas of ligandable sites to aid therapeutic discovery.

Main Methods:

• Utilized evolutionary-scale protein large language models (LLMs) to power the AiPP platform.
• Developed harmonized ML training sets from activity-based protein profiling (ABPP) and co-crystal structure databases.
• Implemented a LLM representation-based clustering framework to reconcile and augment experimental data.

Main Results:

• AiPP accurately predicts cysteine liganding events, recovering 80% from co-crystal structures when trained solely on ABPP data.
• Identified ligandable pockets in "undruggable" transcription factors and previously undetected sites in protein tyrosine phosphatases.
• Generated a proteome-wide covalent ligandability atlas, identifying novel sites in therapeutic targets like MC3R.

Conclusions:

• AiPP advances the creation of a ligandable proteome atlas, crucial for understanding protein function and accelerating drug development.
• The platform effectively identifies ligandable sites, including allosteric ones, in proteins previously missed by experimental methods.
• The LLM-based approach offers a broadly applicable method for interrogating large-scale, heterogeneous data in proteomics and ML model development.