QRxVision: Multimodal QR Encoding and Computer Vision for Domain-Specific Drug Discovery

Abstract

Computational approaches applied to the drug discovery pipeline are invaluable because of their overall speed or time to solution, sustainability, accessibility combined with ever improving accuracy. Furthermore, the high-throughput virtual screening supports practical sampling of the vast chemical space for target-relevant bioactive compounds that can be experimentally validated further downstream in the drug development process. The current paradigm being practiced in computer-aided drug screening strategies centers around structure-based and ligand-based approaches supported by a range of computational tools including de novo design, molecular dynamics-based approaches, and machine learning (ML). Overall, the latter approach has shown great promise generating models of high computational speed coupled with good accuracy. In this work, we developed QRxVision, a domain-specific multimodal ML model explicitly optimized for kinase target assignment using similarity-based profile matching and retrieval. QRxVision utilizes image vision from a pretrained VGG16 (Visual Geometry Group–16) model for deep spatial feature extraction and cosine similarity calculation. By encoding drug multimodality-associated parameters (MAP) into standardized QR code images, QRxVision can evaluate the similarity between test and reference compounds. Drug targets can then be assigned via transitive profile matching. Validation using kinase inhibitors shows that QRxVision is 71% successful at assigning viable small molecules within the Top-5 predictions (p-value < 0.05). Overall, we believe that with the appropriate algorithmic modifications and domain retargeting, the QRxVision framework can be applied to other target assignment tasks.