Drug Repurposing and Combination Therapy
Repurposing and combination therapy sit between discovery and translation. They start with assets, mechanisms, signatures, or drug pairs that already have some evidence, then ask whether a new indication or regimen is biologically plausible. AI adds value when it narrows a search space and clarifies mechanism, not when it turns a computational match into a therapeutic claim.
This chapter gives you a framework for repurposing and combination-therapy claims. You will learn to:
- Distinguish transcriptional signature matching, knowledge-graph traversal, network proximity, and synergy prediction
- Read Connectivity Map/L1000, DeepSynergy, SynergyFinder, and COVID-19 repurposing examples at appropriate evidence weight
- Separate mechanism-of-action discovery from pure pattern-matching repurposing
- Evaluate combination claims against dose, schedule, toxicity, assay context, and validation design
- Treat Recursion and Insilico Medicine as industrial context unless specific claims have independent evidence
- Recognise the failure mode where a plausible computational match is treated as efficacy evidence
Repurposing approaches:
| Approach | Examples | What it supports | Main caution |
|---|---|---|---|
| Signature matching | Connectivity Map, L1000 | Mechanism and reversal hypotheses | Cell line, dose, and time point may not match disease biology |
| Knowledge graphs | Hetionet, PrimeKG, Open Targets | Target, disease, compound, and mechanism links | Edge type and source provenance determine credibility |
| Network medicine | Protein interaction and pathway neighborhoods | Mechanistic neighbourhoods | Proximity is not causality |
| Clinical and real-world evidence review | Prior trials, labels, safety data | Feasibility and safety context | New indication still requires efficacy evidence |
Combination-therapy tools and evidence:
| Resource | Main use | Verified source |
|---|---|---|
| DeepSynergy | Drug-combination synergy prediction | Preuer et al., 2017 |
| SynergyFinder 2.0 | Multi-drug synergy visual analytics | Ianevski et al., 2020 |
| Connectivity Map L1000 | Perturbational signature reference | Subramanian et al., 2017 |
| Drug-repurposing review | Progress, challenges, recommendations | Pushpakom et al., 2018 |
Three failures that look like success:
| Failure mode | Looks like | Actually means |
|---|---|---|
| Signature reversal overreach | Compound reverses disease signature | Disease model, cell context, dose, and endpoint remain unproven |
| Synergy-score overreach | Drug pair has high synergy | Toxicity, dose schedule, and tissue context may make regimen unusable |
| Repurposing shortcut myth | Existing drug means lower risk | New indication still needs benefit-risk evidence |
Introduction
Drug repurposing is attractive because it starts from existing pharmacology, known chemistry, prior safety data, or clinical experience. That starting point can reduce uncertainty, but it does not remove the need for efficacy evidence in the new context. Combination therapy is attractive because many diseases require multi-node intervention. That promise raises complexity: dose, schedule, toxicity, tissue context, resistance, and mechanism all matter.
The scientific discipline is to keep the claim small until validation expands it. A transcriptional match is a hypothesis. A graph path is a hypothesis. A synergy score is a hypothesis. A prior label is context, not proof for a new indication.
Demonstrated
Transcriptional Signature Matching
The Connectivity Map L1000 platform made perturbational signature matching a central repurposing method by profiling transcriptional responses to many perturbagens at scale (Subramanian et al., 2017). The logic is simple: if a disease state has a gene-expression signature, a compound that reverses or phenocopies that signature may be mechanistically relevant.
The demonstrated value is hypothesis generation. Signature matching can suggest mechanisms, compounds, and follow-up experiments. The main constraints are cell type, dose, exposure time, assay platform, disease model, and whether the signature is causal or downstream noise. Signature reversal is not efficacy.
Knowledge Graph and Network Repurposing
Biomedical knowledge graphs and networks connect drugs, targets, diseases, pathways, phenotypes, and evidence sources. Hetionet, PrimeKG, and Open Targets are covered in the knowledge-graph chapter because they form the evidence layer for repurposing workflows. Drug-repurposing reviews emphasise that successful repositioning depends on mechanism, safety, intellectual-property position, clinical feasibility, and evidence quality, not only computational ranking (Pushpakom et al., 2018).
Graph methods help identify neighbourhoods, missing edges, and convergent evidence. They also make weak evidence look strong if all edge types are flattened. A co-mention edge, a genetic association, a pathway membership, a trial mention, and a causal perturbation should never receive the same interpretation.
Combination Synergy Prediction
Combination therapy requires more than ranking single agents. Synergy depends on dose, sequence, timing, cell state, genotype, tissue context, and toxicity. DeepSynergy predicted anti-cancer drug synergy using deep learning on drug and cell-line features (Preuer et al., 2017). SynergyFinder 2.0 provides visual analytics for multi-drug combination synergies (Ianevski et al., 2020).
These tools support screening design and hypothesis prioritisation. They do not establish a usable regimen. A combination that works in a cell line may fail because the exposure is not achievable, toxicity overlaps, schedule is incompatible, or the relevant tissue state is absent.
COVID-19 as a Repurposing Stress Test
COVID-19 showed both the speed and limits of repurposing. A SARS-CoV-2 protein-interaction map identified potential drug targets and repurposing candidates early in the pandemic (Gordon et al., 2020). Large randomized trials then filtered candidate therapies. The WHO Solidarity interim results found no clear mortality benefit for several repurposed antivirals in hospitalized patients (WHO Solidarity Trial Consortium, 2021), while the RECOVERY trial established a mortality benefit for dexamethasone in hospitalized patients requiring oxygen or ventilation (RECOVERY Collaborative Group, 2021).
The lesson is not that repurposing failed. The lesson is that fast computational and mechanistic hypotheses need fast, rigorous clinical filtering. Repurposing speed must be paired with evidence speed.
Industrial Platform Context
Recursion and Insilico Medicine illustrate industrial platform approaches that combine high-throughput biology, computational modelling, and therapeutic development (Recursion, 2026; Insilico Medicine, 2026). In this chapter they matter as platform context, not as proof of any particular repurposing or combination claim.
For diligence, the useful questions are specific: which asset, which indication, which mechanism, which evidence, which assay, which clinical stage, and which decision was changed by the computational method?
Theoretical
Mechanism-First Repurposing
The strongest repurposing logic starts with mechanism. If a compound has known target engagement, known pharmacology, known human exposure, and a disease biology rationale, AI can help prioritise context. The remaining challenge is proving that the mechanism matters in the new disease and that the dose required is tolerable.
Multi-Omic Combination Design
Combination design may improve when transcriptomics, proteomics, CRISPR screens, single-cell perturbation data, pathway models, and pharmacology are considered together. The theoretical value is high because combinations often aim to block compensation, resistance, or parallel pathways.
The hard part is validation. A multi-omic rationale can still fail if the drug exposure is wrong, the tissue context is absent, or toxicity dominates benefit.
Portfolio-Level Repurposing Intelligence
Companies with large internal datasets can search across failed assets, negative assays, historical screens, safety packages, and disease programs. That institutional memory may be more valuable than public-data ranking. The theoretical advantage comes from knowing what failed, not only what was published.
Beyond Current Capabilities
Efficacy from Computational Match Alone
No computational match establishes therapeutic efficacy. Signature reversal, graph proximity, structural plausibility, and synergy scores remain upstream evidence until tested in the right biological and clinical context.
General Combination Prediction Across Diseases
Combination effects do not transfer cleanly across diseases, genotypes, tissues, immune contexts, or dose schedules. General combination prediction remains beyond current capabilities for decision-grade therapeutic selection.
Repurposing Without Commercial and Regulatory Reality
A repurposing hypothesis may be scientifically plausible and commercially unusable. Patent position, formulation, dose, route, manufacturing, exclusivity, safety monitoring, and trial feasibility determine whether a program can proceed.
Practice Notes
Write the claim in one sentence: new indication, old asset, mechanism, population, dose logic, and evidence source.
Separate evidence types. Do not merge signature reversal, graph proximity, prior human exposure, and trial evidence into a single confidence score without showing the components.
Run orthogonal validation before advancing a program. Use disease-relevant assays, dose-response work, target engagement, combination matrices, and toxicity checks.
For combinations, document dose, schedule, order, tissue context, and toxicity. A synergy score without those details is not a regimen.
Treat company platform claims as leads for due diligence. Ask for asset-level evidence, not platform-level language.