Chemical Descriptors and Deep Learning Embeddings for Scoring de novo Peptide Designs

Peptides, a versatile class of molecules, bridge the gap between small molecules and biologics, but their clinical application demands an efficient scoring system that takes into account several crucial factors such as target binding affinity, stability, membrane permeability, aggregation propensity, and non-fouling behavior. In this study, researchers compared two methods for evaluating peptide designs: traditional chemical descriptors and contemporary deep learning embeddings.

By collating nine public datasets covering five developability traits and four binding-affinity endpoints, the team discovered that chemical descriptors alone are sufficient to predict developability task labels with remarkable accuracy. This finding is noteworthy because chemical descriptors require significantly less computational power and are more interpretable than complex deep learning models.

Consequently, classical machine learning algorithms trained on these simple descriptors frequently match or surpass the performance of sophisticated deep learning architectures, making them a highly efficient and interpretable alternative for high-throughput scoring.

The researchers also examined the impact of different representations on scoring binding affinity. Among the tested pair representations, Boltz-2 pair representations emerged as the most informative; however, the model's predictive power was adversely affected by an unexpected bias stemming from the molecular weight of the peptides.

Ultimately, this comprehensive evaluation sheds light on the ongoing significance of interpretable chemical descriptors in tandem with deep learning techniques for predicting both peptide developability and binding affinity, facilitating the selection and scoring of de novo peptide designs.