Recently, significant effort has been given to developing proteochemometric (PCM) compound-kinase bioactivity prediction models that account for the complex interplay between chemical and protein factors affecting binding. The choice of features is critical to the performance of such models. While compound descriptors have been extensively studied in this context, protein features are typically based on a simple amino acid (AA) sequence representation. Lately, protein embeddings learned using neural networks trained on millions of AA sequences have shown promise, though their efficacy for tasks involving specific classes of proteins like kinases is still unclear, nor is it known whether they adequately capture 3D information about the binding pocket.

Here, we develop a kinase featurization scheme that augments AA sequence-based descriptors with 3D features that capture structural characteristics of the ATP binding pocket. Since many kinases do not yet have experimentally determined structures, we analyze structures predicted by AlphaFold2 (AF2) and use them to derive physicochemical and spatial properties of kinase binding pockets using KiSSim. We then train pairwise kernel regression models with kinase ATP binding pocket kernels based on (i) Smith-Waterman alignment of AA sequences, (ii) embeddings from a recent ProtBERT model, and (iii) 3D structure-based KiSSim descriptors (Fig. A). All models employ an ECFP-based compound kernel.

Models were trained using ~170K kinome-wide bioactivities (pKd, pKi, pIC50) curated from publicly available databases, and evaluated on a challenging test set of ~15K compound-kinase pairs. We compare performance of the individual models, as well as ensembles of models using different protein features. We find that an ensemble of models based on the ProtBERT embeddings and our 3D descriptors performs best (Fig. B), indicating that these two feature types provide distinct information. These results demonstrate the utility of AF2-generated structures for downstream machine learning tasks, and highlight the importance of 3D kinase features in modeling compound-kinase bioactivity.

Balaguru Ravikumar

Appearances