Revisiting Deep Learning DTA Methods from the Dataset Viewpoint
Year: 2026
Type: preprint
Abstract: In this work, we present a study revisiting deep learning drug-target affinity models, which are otherwise used as benchmark methods in the literature to compare more complex and heavy-weighted architectures. We focus on training and testing these models using pKd values from BindingDB, a dataset that presents binding affinity information of a high variety of ligans and proteins. These activity data need a proper pre-treatment and curation prior to actually using them for production tasks, a factor that is often overlooked in the literature. We show that a combination of a thorough curation of the dataset and a proper hyperparameter tuning of slight modifications of the revisited models (e.g, by using pretrained protein embeddings to describe the proteins) provides beyond state of the art performances at a much lower computational cost, with respect to more demanding deep learning models. We study the distributions of the protein and ligand learned (or eventually, pre-trained) embeddings, to observe how they classify the proteins depending on the (super)families they belong to (for the former), and how they distribute in the embedding space depending on the values of their binding affinities, with respect to a given protein (for the latter). We take advantage of the graph representations of the ligands to further approach explainability at an atom-wise level, by computing the directional derivative of the model with respect to each of the atoms of a ligand. This method, unlike other derivative-based approaches such as Grad-CAM and variants, turns out to be extremely useful for regression tasks of binding affinities, as this provides direct means of assessing which atoms contribute to an increase/decrease of the binding affinity, depending on both the sign of the directional derivative, and on its absolute value-an indication of how steep the function is with respect to each atom.
Source: ChemRxiv
Institution: Ri.MED
Language: English
Subfield: Drug Discovery
Domain: Life Sciences
Open Access status: green
OpenAlex ID: W7202076359