Lessons from implementing AlphaFold3 in the wild
Tuesday November 12th, 4-5pm EST | Arda Goreci (Ligo Biosciences)
Abstract: AlphaFold3 introduced a number of architectural and training innovations that introduces new ligand and nucleic acid modeling capabilities useful for drug discovery. An open source implementation of the full model is currently not available. We attempted to replicate a minimally viable version of AF3 and identified a number of discrepancies between the deep learning literature and the AF3 pseudocode. Two of these severely compromised training dynamics, and we introduce minor changes that we believe reflect the actual implementation in AF3’s code.
Twitter/X Release Thread: https://x.com/ArdaGoreci/status/1830744265007480934
Github: https://github.com/Ligo-Biosciences/AlphaFold3/tree/main
Arda graduated from Cell & Systems Biology in Oxford, where he was a Google Cloud Research Innovator for research on deep learning-based proteomics. There, he met his two cofounders, and they went on to found Ligo Biosciences, a deep tech company working on enzyme design.