Evaluating AlphaFold3 on D-Amino Acids
Project Description
In May 2024, DeepMind released AlphaFold3 [1], a revolutionary AI model for predicting protein structures and biomolecular interactions, building on its Nobel Prize-winning predecessor, AlphaFold2. Shortly after, MIT introduced Boltz1 [2], an open-source reproduction claiming enhanced capabilities. However, recent research by Childs et al. [3] revealed that AlphaFold3 performs poorly on proteins containing D-amino acids—mirror-image variants of natural amino acids with critical roles in antibiotics, bacteria, and neuroscience. This project aims to reproduce those findings and test whether Boltz1—which promises better handling of non-standard amino acids—actually improves upon AlphaFold3’s limitations. Additional experiments can be considered if time permits.
Supervisor
ZHANG Nevin Lianwen
Quota
2
Course type
UROP1100
UROP2100
UROP3100
UROP3200
UROP4100
Applicant's Roles
The student is expected to carry the experiments using SuperPOD under the guidance of a PhD student.
Applicant's Learning Objectives
Gain hands-on practice running cutting-edge models on GPUs, analyzing protein structures, and contributing to an open question in computational biology
Complexity of the project
Moderate