Language models can identify enzymatic binding sites in protein sequences.

Area of Science:

• Computational biology
• Bioinformatics
• Structural biology

Background:

• Recent advances in language modeling have significantly impacted the analysis of sequential data in scientific domains.
• Language architectures, initially prominent in natural language processing, are increasingly applied to model proteins and chemical processes.
• These models have shown an ability to elucidate structural relationships from sequential data, even revealing three-dimensional structural features.

Purpose of the Study:

• To investigate the capability of unsupervised language model architectures in analyzing bio-catalyzed chemical reactions.
• To determine if language representations of reactions can capture signals related to substrate-binding site atomic interactions.
• To assess the potential for identifying three-dimensional binding site positions in novel protein sequences.

Main Methods:

• Utilized a language model architecture applied to a language representation of bio-catalyzed chemical reactions.
• Employed a language representation incorporating reaction-simplified molecular-input line-entry system (SMILES) for substrates and products.
• Included amino acid sequence information for the enzyme in the language representation.
• Evaluated the model's performance in an unsupervised manner.

Main Results:

• The unsupervised language model successfully captured signals at the base of substrate-binding site atomic interactions.
• The approach demonstrated the ability to identify three-dimensional binding site positions in unknown protein sequences.
• Achieved 52.13% recovery of the binding site when compared to co-crystallized substrate-enzyme structures.
• Significantly outperformed existing attention-based models in this task.

Conclusions:

• Unsupervised language models can effectively analyze sequential data from bio-catalyzed reactions to predict protein structural features.
• The proposed method offers a novel way to identify substrate-binding sites in proteins without prior supervision.
• This approach holds promise for advancing protein structure prediction and drug discovery.