EvoDiff introduces a novel diffusion framework for controlling protein generation. It combines evolutionary data with diffusion models to create high-fidelity protein sequences. The framework addresses limitations of structure-based models by generating proteins with disordered regions and enabling functional scaffold design. EvoDiff's sequence-based approach offers a new way to design proteins beyond the traditional structure-function paradigm.
EvoDiff's key strength lies in its ability to generate structurally plausible protein sequences, even those with disordered regions, which are challenging for traditional structure-based methods. The framework supports both unconditional and conditional generation, including evolution-guided protein design and scaffolding of functional motifs. It offers a flexible architecture adaptable to different protein design tasks and provides a comprehensive set of pretrained models.
- Sequence Generation: Generates novel protein sequences from masked or uniformly sampled amino acids.
- MSA Generation: Creates evolutionary alignments (MSAs) for sequence comparison and analysis.
- Conditional Design: Allows guiding sequence generation based on evolutionary information or functional motifs.
- Flexible Architectures: Provides both autoregressive and D3PM-based diffusion models.
- Extensive Models: Offers a range of pretrained models with different parameter sizes and corruption schemes.
- Data Support: Utilizes datasets like UniRef50 and OpenFold for training and evaluation.
- RLAR integration: Offers baseline LRA models for comparison with OADM and D3PM.
EvoDiff is a research project with active development, featuring a complete codebase and documentation. The project has received significant community interest, evidenced by its star and fork counts on GitHub. Documentation includes installation guides, examples, and instructions for loading pretrained models. Regular updates and contributions indicate ongoing maintenance and development.
EvoDiff benefits researchers and developers involved in protein engineering, drug discovery, and computational biology. It provides a powerful tool for designing novel proteins with desired properties, including those with intrinsically disordered regions or functional motifs. EvoDiff offers a valuable alternative to traditional protein design approaches, providing both a flexible and scalable platform for sequence generation.