Ensemble-conditioned protein sequence design with Caliby.

Area of Science:

• Computational biology
• Protein engineering
• Bioinformatics

Background:

• Structure-conditioned sequence design aims to create proteins with specific folds.
• Current deep learning models struggle with non-idealized backbones due to biases in training objectives.
• Native sequence recovery can introduce non-structural signals, limiting generalizability.

Purpose of the Study:

• To introduce Caliby, a novel Potts model-based sequence design method.
• To address limitations in current protein design models for non-idealized backbones.
• To improve the accuracy and scope of de novo protein design.

Main Methods:

• Caliby conditions sequence design on an ensemble of structures generated from a target backbone.
• Ensemble conditioning averages out biases towards native sequences.
• A variant of Caliby was trained on soluble proteins for specific binder design tasks.

Main Results:

• Caliby reduces native sequence recovery while improving AlphaFold2 self-consistency.
• Outperforms state-of-the-art models (ProteinMPNN, ChromaDesign) on native and de novo backbones.
• Identifies previously undesignable protein binders as designable using a soluble protein-trained variant.

Conclusions:

• Caliby expands the de novo design space beyond idealized backbones.
• Ensemble-conditioned design effectively mitigates biases in protein sequence generation.
• Highlights limitations in existing filtering pipelines for protein design.