Knowledge
Knowledge graph analysis
Learns the relationships among diseases, drugs and genes as a knowledge graph. A Graph Attention Autoencoder (GATE) searches for relationships that have not yet been reported.
Novelty
AI-Native Drug Exploration
From analyzing AI
to exploring AI.
We combine RePhaIND®, our AI drug discovery platform, with AI agents to find new indications for drug candidates whose development has been halted.
01Background
New indications for candidates that lie dormant.
High-quality drug candidates lie dormant in pharmaceutical companies, shelved by changes of strategy. Repurposing draws out their value through new indications.
Repurposing requires many kinds of analysis and cross-field investigation. AI now handles individual analyses. But once they are linked into an exploration, small deviations in premises and judgment at each step can accumulate, and the work can drift from its goal.
The name of a candidate alone does not support the next decision.
- Why this candidate?The reason for selecting or rejecting it is needed.
- Why would it work?An explanation of the mechanism is needed.
- Where is the evidence?Sources that can be traced are needed.
02Platform
RePhaIND®: layering different kinds of analysis to find candidates and support them.
RePhaIND® is an AI drug discovery platform developed by GEXVal. It combines knowledge graphs, real-world data and gene expression data to evaluate candidates in multiple layers.
Clinical
Real-world data analysis
Examines candidates against clinical data through disproportionality analysis of the FDA Adverse Event Reporting System (FAERS).
Robustness
Phenome
Gene expression and pathway analysis
Assesses biological plausibility through pathway analysis of gene expression data (LINCS).
Resolution
Integration
Expert information search with LLMs
Uses large language models to search the literature and specialist databases across sources.
Connection
Analyses used alongside
Cheminformatics
Binding partner prediction from compound structure
Predicts the proteins a compound may bind to from its structure, as a clue to new indications. It covers small molecules. Results are used to prioritize hypotheses and are evaluated together with the other analyses.
Public expression data
Target expression check in disease tissue
Checks, using public gene expression data, whether the target of a candidate is expressed in the tissues and cells relevant to the disease. Results serve as reference information for planning experiments.
03Exploring AI
Linking analyses into one exploration.
RePhaIND® and AI agents link the steps from candidate search to prioritization into a single flow.
- 1Candidate searchList candidates broadly
- 2Mechanistic hypothesesTurn why it may work into hypotheses
- 3Evidence gatheringCollect evidence with its sources
- 4PrioritizationNarrow down and keep the reasons
At the core: the tacit know-how of drug discovery
The tacit know-how of drug discovery, absent from the public record, is implemented as procedures the AI can execute and step-by-step criteria for judgment. Our aim is AI-native drug exploration: rebuilding the process around AI rather than handing it individual analyses.
04Research
Shown in a peer-reviewed paper, with COVID-19 as the example.
We built a framework that supports GATE-based candidate prediction with real-world data and gene expression data. It is joint research with the Research Center for Advanced Science and Technology at the University of Tokyo and the Faculty of Pharmacy at Kindai University.
16
Candidates identified by GATE
12 were already used or recognized in the context of COVID-19 or its symptoms
4
Drugs evaluated in multiple layers
Drugs with minimal prior COVID-19 associations, evaluated with two kinds of analysis
2
Prioritized candidates
Cilastatin and megestrol
62.7%
Share of approved indications absent from the knowledge graph that ranked within the top 100 of 4,314 drugs
Text embedding 43.7%node2vec 37.8%80.3% within the top 500
27,644
Nodes in the knowledge graph
Diseases 11,882Drugs 4,314Genes 11,448
Pharmaceuticals 2026, 19(10), 1522 Open Access
Multilayered Prioritization of Graph Attention Autoencoder–Derived Drug Repurposing Candidates for COVID-19
doi.org/10.3390/ph19101522The results of this study are hypotheses based on computational predictions and do not demonstrate efficacy. Experimental confirmation is needed.