A Computer Vision Approach to Align and Compare Protein Cavities: Application to Fragment-Based Drug Design

Abstract

Identifying local similarities in ligand binding sites from distant proteins is still a major hurdle to rational drug design approaches. We herewith present a novel method, borrowed from computer vision, particularly adapted to mine fragment subpockets and compare them to entire ligand-binding sites.Pockets are represented by pharmacophore-annotated point clouds mimicking ideal ligands or fragments. Point cloud registration is used to find the transformation enabling an optimal overlap of points sharing similar topological neighborhoods and pharmacophoric features. Importantly, the local environment of randomly sampled cavity points is used to generate a preliminary alignment that is next refined by an iterative closest point algorithm. A scoring function has been tuned to quantify the degree of shape and pharmacophoric overlap, and assess pocket pairwise similarity. The method (ProCare) was first calibrated on a large set of known similar and dissimilar druggable cavities, shown to be insensitive to moderate variations in atomic coordinates, and then applied to the specific problem of comparing fragment subpockets with entire cavities. Starting from a unique set of cavity points, point cloud registration outperformed a state-of-the-art computational method (shape-based similarity search) in detecting local similarities between fragment subsites and entire cavities. A collection of 33,953 subpockets annotated with their bound fragments was screened for local similarity to cavities from three recently described protein X-ray structures. ProCare was able to detect local similarities between remote pockets and transfer the corresponding fragments to the query cavity space. Interestingly, fragments selected from totally unrelated proteins, nicely overlap substructures of the masked original ligand co-crystallized with the target query, thereby proposing an automated first step to an automated fragment-based design approach targeting ligand-orphan cavities.

Unexpected similarity between HIV-1 reverse transcriptase and tumor necrosis factor revealed by binding site image processing

“…Entire cavities ("cavity_all" output) were calculated for TNF-α structures whereas only cavity points closer than 4.0 Å from any fragmented ligand heavy atom ("cavity_4" output) were considered for HIV-1 RT binding sites. VolSite cavity points were directly used for point cloud registration and determination of colored fast point feature histograms (c-FPFH) as previously described [17]. Finally, the respective set of c-FPFH descriptors for the two cavities were aligned and compared to each other using a RANSAC algorithm [19,20] with default parameters [17].…”

“…Finally, the respective set of c-FPFH descriptors for the two cavities were aligned and compared to each other using a RANSAC algorithm [19,20] with default parameters [17]. Alignments results were scored with the default ProCare scoring function [17] which evaluates with a Tversky metric the proportion of aligned points of the same pharmacophoric features. In agreement with our previous study, a similarity ProCare score of 0.47 (p-value of 0.05) was used as threshold to distinguish similar from dissimilar binding sites.…”

“…ProCare. ProCare [17] v.0.1.1 pairwise comparison were performed on cavities computed with the VolSite module [21] in IChem v5.2.9 [40]. Entire cavities ("cavity_all" output) were calculated for TNF-α structures whereas only cavity points closer than 4.0 Å from any fragmented ligand heavy atom ("cavity_4" output) were considered for HIV-1 RT binding sites.…”

“…Since ProCare uses local descriptors to compare and align binding subpockets, the method is particularly suited to fragment-based design strategies aimed at positioning fragments in subpockets of any druggable cavity. While validating the method by focused benchmarking studies [17], we noticed some unexpected local similarity between subpockets from two unrelated proteins: the asymmetric human tumor necrosis factor alpha (TNF-α) trimer [22] and the human immunodeficiency virus type 1 reverse transcriptase (HIV-1 RT) [23]. To exclude potential artifacts or biases and provide a strong statistical support to this initial prediction, we here systematically compared the inner cavity of three inhibitor-bound TNF-α trimer structures with 122 non-nucleoside inhibitor-bound HIV-1 RT X-ray structures.…”

Untitled

“…The fragment based multi-target QSAR method/computational fragment-based drug design is employed to obtain the most appropriate fragments as structural alerts performing anti-cancer activity. This method is also used to develop innovative molecular entities that are expected, by this model, for being potentially potent and versatile anti-cancer agents (Eguida & Rognan, 2020). The methods most frequently used include.…”