AI-Powered Target Identification & Deconvolution

In a peer-reviewed study, Plex's AI platform identified novel drug targets from phenotypic screening data — predictions that were then validated in biochemical assays and a mouse model of liver disease, reducing hepatic steatosis by 68%.

The Challenge: Target Deconvolution in Phenotypic Drug Discovery

Phenotypic screening identifies compounds that work in disease-relevant cell models — but it doesn't tell you why they work. Target deconvolution (identifying the molecular targets of a hit compound) traditionally requires months of pull-down assays, affinity chromatography, and genetic knockdown experiments. This bottleneck delays the transition from hit to lead and limits the ability to optimize compounds rationally.

How Plex Solved It

⛶Figure 4 from Hoegeman et al. (2025). Chain of evidence showing selected targets linking α-terthienyl to DPP-IV and HSD17β1 via structural similarity analysis and binding database cross-referencing.

Researchers at the University of Michigan used Plex's AI platform for target deconvolution of their lead compound, α-terthienyl, identified from a high-content phenotypic screen for MASLD (fatty liver disease). Plex integrated the compound's chemical structure, quantitative phenotypic fingerprint, and biological knowledge graphs from 60+ public databases to generate focal graphs — knowledge subgraphs analyzed using centrality algorithms like PageRank — that ranked protein target hypotheses by probability.

What Plex Identified

• DPP-IV (dipeptidyl peptidase 4) — a validated metabolic target, confirmed by biochemical assay with IC50 of 1.47 μM
• HSD17β13 (17-beta hydroxysteroid dehydrogenase 13) — a genetically validated MASLD target, confirmed by enzymatic inhibition assay
• Dual-target polypharmacology — Plex predicted the compound would act through both targets simultaneously, a mechanism validated in vivo
• 8 total target hypotheses ranked by confidence, with the top predictions confirmed experimentally

Validated Results

The Plex-predicted targets were validated across multiple experimental approaches: molecular docking confirmed binding affinity, biochemical assays confirmed enzymatic inhibition, and an in vivo mouse model demonstrated 68% reduction in hepatic steatosis, near-complete resolution of microvesicular steatosis, significantly improved glucose tolerance, and normalized liver biomarkers (ALT/AST). The compound showed an EC50 of 106 nM with no cytotoxicity up to 15 μM.

⛶Figure 3 from Hoegeman et al. (2025). H&E-stained liver sections of (A) vehicle control, (B) healthy control, and (C) α-terthienyl treated mice at 5 mg/kg showing resolution of steatosis. (D) Improved glucose tolerance and decreased liver weight in treated animals.

Key Capabilities Demonstrated

• AI-driven target deconvolution from phenotypic screening data — no prior target hypothesis required
• Multi-modal integration of chemical structures, phenotypic fingerprints, and biological knowledge graphs
• Focal Graph analysis using PageRank-style algorithms to rank target hypotheses by probability
• Rapid identification of polypharmacological mechanisms that would take months to discover experimentally
• Predictions validated by independent biochemical assays and in vivo disease models

⛶Figure 5 from Hoegeman et al. (2025). (A) Biochemical inhibition of DPP-IV. (B) In cellulo activity of HSD17β13. (C–D) Molecular docking of α-terthienyl into DPP-IV and HSD17β13 binding pockets confirming predicted binding interactions.