Computational tools to address challenging targets in drug discovery: target-focused chemical libraries and structure-based machine learning.

Abstract

There are many human drug targets without any known small molecule inhibitors, and a lot of these challenging targets play a crucial role in important disease-relevant processes. RNA-binding proteins (RBPs) are some of the examples of these kind of targets. While RBPs play a crucial role in countless cellular processes, including post-transcriptional regulation of genes, efforts directed at finding small molecule inhibitors for these targets have been largely unsuccessful. For this reason, I have focused my PhD studies in developing computational methods that will allow us to rapidly and robustly identify small-molecule inhibitors of RBPs. From an in-silico standpoint, the scoring functions that power most computational structure-based drug discovery are limited by high false positive rates. To address this challenge, I built the first false-positive-aware machine learning scoring function (vScreenML). vScreenML demonstrated a significant improvement in the false positive rate over the current state-of-the-art classical and machine learning scoring functions—both in retrospective and prospective evaluations. More broadly, existing virtual screening approaches were also built to suit traditional drug targets. By contrast, RBPs have unique structural features that differentiate them from most of these types of target classes: they have a large, shallow interface that more polar than most traditional drug targets. Since RBPs are structurally distinct, this may explain why traditional methods have struggled to find chemical matter to address these. To tackle this challenge, we built the first fully automated RBP pharmacophore extractor that identifies “hotspots” on the RNA that contribute extensively to the binding affinity of the protein-RNA interaction; these hotspots are then used as template for pharmacophoric virtual screening. This tool also powers the first PDB-wide pharmacophore analysis and selectivity profiling of RBPs and was instrumental to our success in designing the first series of rationally designed inhibitors of Musashi proteins. Finally, the compounds that comprise typical screening libraries are also biased towards the types of chemical space that are appropriate for traditional drug targets. To address this, I developed a method for building target-focused libraries of synthetically accessible compounds and applied this to build a collection of compounds enriched in likely Musashi inhibitors. To test the utility of this library, I synthesized and tested some of the top-scoring hits and confirmed that we had identified new Musashi inhibitors from this library. Looking ahead, I envision that these three tools will collectively enable development of better inhibitors targeting Musashi, and entirely new inhibitors of other RBPs.