Chemical language models for modified peptides

Protein language models (PLMs) achieve state-of-the-art performance in several protein and peptide related tasks such as property prediction and de novo design. However, because they are based on residue-level sequence information, they cannot account for common chemical modifications of peptides such as cyclization, incorporation of unnatural amino acids, backbone modifications and others. By representing the modified peptide as a string that encodes the explicit molecular graph, this project will leverage the significantly higher flexibility of Chemical Language Models (CLMs) for the prediction of properties of modified peptides.

- Familiarization with the state of the art in PLMs and CLMs
- Using the RDKit cheminformatics library to convert the chosen representations of modified peptides (e.g. in HELM or PDB format) into parsable molecular graphs
- Using pre-existing CLMs (e.g. ChemBERTa) to represent these molecular graphs
- Using the resulting embeddings as features for the prediction of properties on existing modified peptide benchmarks such as cycPepMPDB and benchmark it against the state of the art and compare advantages and disadvantages of CLMs in this context