Leveraging the AI protein structure prediction revolution for biomedical research

Anastassis Perrakis

Keywords: protein structure prediction

TL;DR: Concepts and tools that enable biomediucal research to more effectively use the AI revolution in protein prediction

Abstract:

The advent of artificial intelligence (AI) in protein structure prediction, exemplified by tools like AlphaFold3 (AF3), has revolutionized biomedical research by enabling high-accuracy modeling of protein structures and interactions. This paper explores concepts and tools that harness this AI-driven revolution to advance biomedical discovery, focusing on approaches that integrate AI-based structure prediction with functional genomics, volumetric deep learning, and traditional bioinformatics to address key challenges in protein science.

AlphaFill leverages the DeepMind-EBI database of predicted models and the Protein Data Bank of experimental structures to incorporate functional small molecules into predicted protein structures, enhancing their functional annotation. Similarly, LegoFill, a deep neural network model, predicts metal ion, nucleotide binding and other common physiological ligand moiety biding sites using volumetric protein representations. These tools have been validated on both experimental and computationally predicted structures, with applications to novel protein families.

AlphaBridge focuses on using protein structure prediction in the context of multi-component complexes, refining them filtering and validating structures, improving reliability. AlphaBridge has been used for new discoveries, e.g. for finding new interactors of the condensin complex, or understanding interactions below the cell surface. It has also been used to filter proteomics data, by e.g. evaluating possible readers of complex patterns of ADP ribosylation in human proteins. developing. The derived TCR-bridge algorithm for the evaluation of structure prediction models in immune complex recognition, has the long term goal to predict best strategies for T-cell therapies targeting specific neo-antigens.

A new contribution is our effort for predicting multi-protein complexes, which integrates public functional genomics screen data and computational interactomes, with clustering methodologies and multi-protein AF3 structure predictions coupled to Alpha-Bridge filtering, yielding approximately 400 high-confidence complex structures. Based on this work we also demonstrate the use of large language models (LLMs) with structured prompts to accelerate hypothesis generation for uncharacterized complexes, enhancing early-stage research efficiency.

These tools and data, including structure predictions, are accessible through comprehensive online resources, fostering community-driven discovery. By bridging AI-driven structural biology with functional insights, our approaches empower researchers to uncover novel protein interactions, annotate functions, and accelerate drug discovery, highlighting the transformative potential of AI in biomedical research.

Submission Number: 36