AI molecular search engine paired with RNA sequencing analysis to develop potent and selective VEGFR-3 inhibitors.

3104 Background: Tumor angiogenesis (TA) is driven by several VEGF factors and corresponding receptors (VEGFRs). Targeting TA inhibits tumor growth, however there are no selective small molecule TA inhibitors on the market yet. Here we report development of a selective VEGFR-3 inhibitor with potential anti-tumor activity in multiple cancers. Methods: The Bioptic virtual screening pipeline employs two models. The first is a SMILES-based LLM fine-tuned on binding affinity data of contrastive molecular pairs. The screening was performed on the ultra-large 40-billion-compound virtual library Enamine REAL. For selectivity, the top-ranked molecules were re-scored using a secondary model, a GNN specifically designed for this task and trained to differentiate activity across similar kinases. The Oncobox algorithm was used to select cancer types with the highest sensitivity to VEGFR1-3 inhibitors based on VEGF(R)s expression and TA pathways activation. It was applied to RNA-seq profiles from TCGA (11428 profiles from 33 primary sites) and the internal relevant RWD cohort (1056 profiles from 89 cancer types). Results: Compounds with IC50 < 10 μM in Eurofins VEGFRs KinaseProfiler were considered active. Among 110 tested compounds, 1 was active against VEGFR-1, 1 - against VEGFR-2, and 4 - against VEGFR-3. One compound was active against all VEGFRs, and 3 - against a single VEGFR. One VEGFR-3 active showed > 45-fold selectivity against both VEGFR1 & 2, while no compound was specific to VEGFR-1 or VEGFR-2. All 3 selective VEGFR-3 inhibitors showed minimal activities ( < 50% at 10 μM) against the 12 off-target kinases including B-Raf, c-Raf, c-Kit, FGFR1, FGFR2, FGFR3, FGFR4, Flt3, Met, PDGFR-a, PDGFR-b, Ret. In order to select cancer types for further pre-clinical validation, Oncobox algorithm was used to simulate predicted efficiency of a VEGFR-3 inhibitor in multiple cancer types from TCGA and internal RWD cohort. The highest response rate is expected for papillary thyroid cancer, followed by clear-cell renal, pancreatic, ovarian cancers, and sarcomas. A selective VEGFR-3 inhibitor may be beneficial when compared to already developed pan-VEGFR inhibitors due to lower toxicity. Finally, identification of potential responders to selective anti-VEGFR-3 therapy via RNA-seq analysis may enable patient enrichment in further clinical trials and development of a companion diagnostic for the drug. Conclusions: Our results suggest that Bioptic’s molecular search engine significantly enhances identification of potent and selective inhibitors for a specific target, and, paired with the Oncobox algorithm, may facilitate development of novel anti-cancer drugs.