Protein function-conditioned language models for variant effect prediction and controllable design.

Area of Science:

• Computational Biology
• Protein Engineering
• Bioinformatics

Background:

• Protein engineering aims to predict sequence variant effects and design functional proteins.
• Current methods often lack the ability to generalize across protein families or incorporate functional constraints effectively.

Purpose of the Study:

• To develop a unified framework for protein engineering that conditions generative models on molecular function semantics.
• To improve variant effect prediction and enable controllable protein sequence design.

Main Methods:

• Utilized Gene Ontology (GO) terms, including text and topology, to create function embeddings.
• Developed Func2Seq, a transformer-based model for P(seq|func), and Func2Prot, a joint sequence-structure model P(seq, str|func).
• Employed function-conditioned meta-pretraining for cross-family transfer before family-specific fine-tuning.

Main Results:

• Func2Seq demonstrated improved predictive performance on variant effect prediction benchmarks compared to existing methods.
• Meta-pretraining significantly enhanced performance, especially for small or low-diversity protein families.
• The study showed successful functional site-level constraint preservation in novel sequence design for the ARID domain.

Conclusions:

• Function-conditioned generative modeling provides a unified approach for protein engineering tasks.
• The proposed framework enhances prediction accuracy and enables controllable, function-guided protein design.
• Joint sequence-structure modeling offers advantages for specific applications, such as enzyme fitness studies.