VALDO: Using Variational Autoencoders to Improve the Signal-to-Noise Ratio of Drug Fragment Screens

Abstract

Fragment-based drug discovery (FBDD) methods have become a popular way to identify starting points for small molecule modulators of protein activity. These methods typically screen libraries of hundreds or thousands of small fragments, using computational methods to identify the fragments that successfully bind to the protein. However, such computational methods are sensitive to heterogeneity among the individual samples, making it difficult to isolate the true signal of a binding event from noise.

To address this issue, we present VAE-Assisted Ligand Discovery (VALDO), a novel method to remove noise post-screening and identify ligand binding in drug fragment screens that supplements current FBDD methods. VALDO is a neural alternative to traditional approaches to detecting binding events based upon existing techniques in dimensionality reduction.

Specifically, VALDO trains a VAE to embed the inputted holo state into low-dimensional space before decoding it back into a high-dimensional dataset. VALDO relies on the assumption that the latent space will only be able to capture the general structure of the given sample, forgetting idiosyncratic variation, such as changes induced by a ligand, which is unique to the specific sample. This reconstructed state then acts as our estimate of the apo state, allowing us to construct difference maps useful for downstream tasks such as identifying ligand binding and other events of interest.

Through experimental results and comparison with current state-of-the-art methods, such as PanDDA and Cluster4x, we demonstrate that VALDO can identify binding events with a better signal-to-noise ratio than current methods, while also identifying meaningful non-binding changes in protein conformation.