Identifying binding sites similarity between unrelated targets across the proteome is valuable to drug design in a range of applications spanning polypharmacology, off-target detection and ligand design, but still is a challenge to state of the art binding site comparison methods, notably regarding local similarities arising from pocket microenvironment. We therefore developed ProCare, an open source computational approach to compare protein pockets with a 3D point cloud registration algorithm. In computer vision, point cloud registration is a fundamental problem of finding the best transformation (rotation, translation, scaling) to align two clouds of points. A protein pocket is here represented as an ensemble of 3D atomic coordinates annotated with microenvironment specific pharmacophoric properties. Following the characterization of each point with a hybrid shape-chemical descriptor (c-FPFH), two pockets are aligned by superimposing their corresponding points sharing the most similar patterns. The alignments are then evaluated by estimating the proportion of matched points sharing the same pharmacophoric properties. Out of a large-scale comparison where ~33,000 subpockets from different proteins in the sc-PDB database were compared to the pocket of the trimeric Tumor Necrosis Factor alpha (TNF-α), ProCare suggested a similarity with the non-nucleoside binding site of HIV-1 reverse transcriptase (HIVRT). Extensive binding site comparisons using different structures of TNF-α and HIVRT reinforced that similarity hypothesis, which was later confirmed by microscale thermophoresis assay where two out of three tested HIVRT FDA-approved drugs were found to bind to TNF-α (KD in the 20-40 µM range). Remarkably, other binding site comparison methods as well as ligand 2D fingerprints and 3D shape methods were not able to detect that similarity. ProCare allows local comparison of protein pockets of different sizes while yielding visually interpretable results. The method is ideally suited to identify local and unobvious similarities among totally unrelated targets and appears as a promising idea generator for fragment-based ligand design, able to pick starting fragments at a proteomic scale, not necessarily influenced by existing ligand or cavity neighborhoods.

Unexpected similarity between TNF-α and HIV-1 reverse transcriptase binding sites revealed by a 3D computer vision-inspired method

ACS Fall 2021

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.