omrikais/alphafold3-mac
Run AlphaFold 3 protein structure prediction natively on Apple Silicon Macs (M1/M2/M3/M4). MLX inference, web UI, restraint-guided docking.
What's novel
Run AlphaFold 3 protein structure prediction natively on Apple Silicon Macs (M1/M2/M3/M4). MLX inference, web UI, restraint-guided docking.
Code Analysis
7 files read · 3 roundsA complete reimplementation of AlphaFold 3's protein structure prediction model in Apple's MLX framework, enabling native inference on Apple Silicon Macs with a full data pipeline, web UI, CLI, and restraint-guided docking.
Strengths
Genuinely deep implementation of the full AlphaFold 3 architecture (Evoformer, diffusion head, confidence head, attention, geometry) in MLX with careful JAX-to-MLX weight mapping, proper numerical parity considerations (float32 casting, mask handling), and a complete surrounding ecosystem (web UI, REST API, CLI, MSA caching, restraint system).
Weaknesses
The weight name mapping table in model.py is a large static dictionary that could be fragile across AF3 versions, and the lazy _build_conditioning pattern in the diffusion head (rebuilding layers on first call) is a code smell that could mask dimension mismatches. Some files are very large (model.py at 90KB, runner.py at 40KB) suggesting opportunities for further decomposition.
Score Breakdown
Signal breakdown
Innovation
Craft
Traction
Scope
Evidence
Commits
25
Contributors
2
Files
485
Active weeks
5
Repository
Language
Python
Stars
2
Forks
0
License
NOASSERTION