Natural Language Prompts Guide the Design of Novel Functional Protein Sequences.

Area of Science:

• Computational Biology
• Artificial Intelligence
• Protein Engineering

Background:

• Natural language interaction is driving innovations in text-guided generation across various domains.
• Designing functional proteins with desired attributes remains a significant challenge in biotechnology.

Purpose of the Study:

• To introduce the Biological Multi-Modal Model (BioM3), a novel framework for designing functional proteins using natural language prompts.
• To enable the conditional generation of novel protein sequences with specific attributes via text descriptions.

Main Methods:

• BioM3 integrates natural language and protein design through a three-stage process: joint embedding space alignment, text embedding refinement, and conditional sequence generation.
• Utilizes contrastive learning for representation alignment and a discrete autoregressive diffusion model for sequence generation.
• Employs in silico validation for various protein property predictions and in vitro/in vivo experimental tests on designed Src-homology 3 (SH3) domain proteins.

Main Results:

• BioM3 successfully synthesizes protein sequences based on detailed text annotations of structure, lineage, and function.
• In silico validation demonstrated strong performance in subcellular localization prediction, reaction classification, and homology detection.
• Experimental tests confirmed that BioM3-designed proteins exhibit native-like tertiary folds and wild-type functional levels.

Conclusions:

• BioM3 represents a significant advancement in protein design, enabling the creation of functional proteins guided by natural language.
• The framework achieves state-of-the-art performance in zero-shot prediction and homology detection tasks.
• BioM3 holds promise for accelerating protein engineering and the development of novel biotechnological applications.