ManyFold: an efficient and flexible library for training and validating protein folding models.

Area of Science:

• Computational biology
• Structural biology
• Deep learning

Background:

• Protein structure prediction is crucial for understanding biological function.
• Existing deep learning models like AlphaFold and OpenFold have advanced the field.
• There is a need for flexible and trainable libraries to facilitate further research.

Purpose of the Study:

• Introduce ManyFold, a novel deep learning library for protein structure prediction.
• Enable support for diverse input modalities, including MSAs and pLM embeddings.
• Facilitate the training and fine-tuning of new protein structure prediction models.

Main Methods:

• Developed ManyFold using Jax for efficient distributed computation.
• Implemented support for multiple sequence alignments (MSAs) and protein language model (pLM) embeddings.
• Enabled inference of existing models like AlphaFold and OpenFold within the library.

Main Results:

• Trained a proof-of-concept pLM-based model, pLMFold, from scratch.
• pLMFold achieved reasonable protein structure prediction results.
• Demonstrated reduced computational overhead compared to AlphaFold.

Conclusions:

• ManyFold provides a flexible and trainable platform for protein structure prediction.
• The library supports diverse inputs and existing models, accelerating research.
• pLMFold shows promise for efficient protein structure prediction with deep learning.