ManyFold

ManyFold provides deep-learning-based protein structure prediction that supports Multiple Sequence Alignments (MSAs) and protein language model (pLM) embeddings to enable flexible inference and model training for structural biology applications.

Key Features:

• Dual Input Support: Supports models that accept Multiple Sequence Alignments (MSAs) and Protein Language Model (pLM) embeddings as inputs.
• Model Inference Flexibility: Facilitates inference using established models such as AlphaFold and OpenFold.
• Full Trainability: Permits fine-tuning of existing models and training new models from scratch.
• Efficient Distributed Operations: Implemented in Jax and supports efficient batched operations in distributed settings.
• pLMFold proof-of-concept: Provides the pLMFold model trained from scratch on pLM embeddings, demonstrating reasonable predictive results with reduced computational overhead relative to AlphaFold.

Scientific Applications:

• Exploration of novel protein structures: Enables prediction and investigation of previously uncharacterized protein folds using MSA or pLM-derived inputs.
• Functional inference: Supports studies aiming to relate predicted structures to protein function.
• Method development in computational biology: Serves as a platform for developing and evaluating deep-learning approaches to protein structure prediction.

Methodology:

Includes the pLMFold proof-of-concept model trained from scratch using protein language model (pLM) embeddings, demonstrating reduced computational overhead compared to AlphaFold while delivering reasonable predictive performance.