PeptideCLM-2 chemical language model scales SMILES-based representation for therapeutic peptide engineering

Therapeutic peptides occupy a crucial middle ground in drug discovery, offering high target specificity and chemical diversity, yet they face a significant computational blind spot. Traditional protein models are limited to natural amino acids, while small molecule chemical models struggle with the larger, polymer-like sequences of peptides. This computational disconnect forces researchers to rely on static chemical descriptors or complex, custom-tailored multiembedding pipelines, hindering efficient exploration of the vast peptide chemical space and slowing the development of new peptide therapeutics with improved pharmacokinetic and pharmacodynamic properties.

Researchers developed PeptideCLM-2, a suite of chemical language models designed to natively represent complex peptide chemistry. The model was trained on an extensive dataset comprising over 100 million molecules, utilizing SMILES-based representations to capture intricate chemical details. This approach aimed to bridge the gap between existing protein and small molecule computational models. The performance of PeptideCLM-2 was rigorously benchmarked against prior computational methods for predicting critical development endpoints, including membrane diffusion, biological function, and half-life, demonstrating its utility in various aspects of peptide drug discovery.

PeptideCLM-2 demonstrated strong performance across all evaluated development endpoints, significantly outperforming prior methods in representing complex peptide chemistry. The model's ability to natively process peptide sequences via SMILES-based representations effectively addresses the long-standing computational disconnect in the field. This advancement allows for a more nuanced capture of subtle chemical details that were previously missed by static descriptors or cumbersome multiembedding pipelines. The benchmarking results confirmed its superior predictive capabilities for crucial parameters such as membrane diffusion, biological function, and half-life, which are vital for successful therapeutic development. This indicates a substantial improvement in the computational toolkit available for designing and optimizing peptide drugs.

PeptideCLM-2 expands the available toolkit of machine learning models for therapeutic peptides, showing strong performance versus prior methods for predicting development end points including membrane diffusion, biological function, and half-life.