Competition Overview
Protein binding remains one of the most challenging aspects of de novo protein design. In recent years, generative AI algorithms have been proven to immensely boost the process of drug discovery and development across all computational approaches. During Winter 2024, Adaptyv Bio organized its second-round protein design competition, focusing on EGFR binding optimization. The competition drew significant participation, with 130 teams, including Vizuro, submitting 1,131 novel designs of which 400 underwent wet lab validation.
Our Strategic Approach
EGFR binding represents a well-established research area, with four FDA-approved monoclonal antibodies and six currently undergoing clinical trials. For this competition, we chose Cetuximab as our template for EGFR binding optimization. Our methodology incorporated cutting-edge generative AI algorithms to re-design the Cetuximab Fab heavy chain, coupled with deep learning-governed binding affinity validation. Following Adaptyv Bio’s guidelines, we implemented a rigorous novelty check (ensuring more than ten amino acid mutations from the existing protein database and patent database) and integrated competition evaluation metrics to select our final 10 designs. Using our de novo protein design protocol, we applied 8 distinct design strategies targeting different regions of Cetuximab for re-design. This enabled the generation and binding affinity estimation of 80,000 de novo proteins. Levering a single RTX 3090 (24 GB VRAM), our protocol identified the top 10 candidates in just 6 hours.
Competition Outcomes
Ranking: Our team placed 25th amongst 130 participating groups.
Notable Achievement: We secured the top potion in the “Best PAE” category with a score of 26.27. PAE (Predicted Aligned Error) from AlphaFold measures the global confidence score in predicting relative residual positions between amino acids pairs within the structure.
Critical Threshold: While the top 24 designs achieved EGFR binding affinities below 10-5 nM, our design was the “best non-binders”.
Reflection: It has been an invaluable experience in testing our computational approach against real-life challenges. Though we didn’t achieve our primary objective, the experience has motivated us to refine our techniques for future opportunities.
Key insights
Template Selection is Crucial: The winning team, “Cradle”, also utilized Cetuximab as their template, validating our fundamental approach while highlighting opportunities for optimization.
Binding Affinity Prediction Is Critical: As the competition winners claimed, generative AI designs coupled with machine learning-based binding affinity ranking yield the best results.
Generative AI sheds light on drug discovery: the competition’s winning design achieved better binding affinity than Cetuximab, a drug approved 20 years ago. This milestone marks the initiation of revolutionary AI-oriented drug discovery and drug design.
Vizuro could deliver de novo protein design within hours: our de novo protein protocol can redesign and estimate the binding affinity of 80,000 proteins in just 6 hours using a single RTX 3090.
About Apotek Bench
Apotek is Vizuro's comprehensive end-to-end SaaS platform tailored for life sciences. It streamlines drug discovery and research with features like an interactive omics data dashboard, a drug discovery pipeline for immuno-oncology, and multi-dimensional target ranking. Apotek also provides AI-powered bioinformatics support, a digital twin for biosimulation, and services for experimental research. Apotek accelerates scientific insights and innovations in healthcare and biotechnology.
Learn more about Apotek Bench: https://www.vizuro.com/apotek
Reach out to Vizuro team: https://www.vizuro.com/contact
Comments