Nanofold is an innovative machine learning model designed for predicting protein structures, building upon the frameworks established in the AlphaFold 2 and AlphaFold 3 research papers. Tailored to run efficiently on mid-tier GPUs, Nanofold significantly streamlines the training process while maintaining high accuracy in protein structure predictions.

Key Features

• Enhanced Efficiency: Utilizing the architecture of AlphaFold 3, this model is more efficient and focuses solely on monomer protein chains, thereby minimizing the training data requirements.
• Memory Optimization: Implements gradient checkpointing to significantly reduce GPU memory usage.
• Advanced Data Handling: Stores input features in Apache Arrow IPC format, allowing for the processing of datasets larger than available RAM.
• Speed Improvements: Incorporates torch.compile for JIT compilation, enhancing training speed.
• Integration with MLFlow: Monitors training metrics and manages model checkpoints efficiently.
• Space-Saving Compression: Compresses datasets using sparse matrices, conserving valuable disk space.
• Docker Support: Provides Docker images facilitating both training and data processing pipelines.
• Continuous Integration: Utilizes GitHub Actions for automated testing.

Implementation Overview

Data Processing Pipeline

The data processing pipeline, accessible via nanofold/preprocess/__main__.py, is designed to:

1. Parse mmCIF files from the Protein Data Bank to extract protein chain details, including residue sequences and atomic coordinates.
2. Search genetic databases such as small BFD and Uniclust30 for proteins with similar sequences, enabling the creation of multiple sequence alignments (MSA).
3. Locate structurally similar proteins (templates) using MSA data from the PDB70 database.
4. Prepare and store all features in an Arrow IPC file for use in the training pipeline.

Training Pipeline

The training component implements core algorithms as per the AlphaFold Supplementary Information, making modifications such as:

• Focusing exclusively on individual protein chains, ignoring complex interactions with ligands or multi-chain structures.
• Omitting specific auxiliary metrics trained in AlphaFold 3 to simplify the model's focus.

Further Exploration

For comprehensive project documentation, visit ogchen.github.io/nanofold and explore the associated blog post for deeper insights into the development and potential applications of this pioneering model.