Terminal
Terminal/Skill
Helps reason through drug repurposing in life sciences: choosing a target or phenotype approach, weighing evidence, checking translatability, and ranking candidates.
By Amazon Web Services
Works with
npx skills add awslabs/hcls-agent-skills --skill drug-repurposingThis skill teaches the agent how to reason about drug repurposing — identifying approved or clinical-stage compounds that may treat a new indication. Repurposing succeeds when three conditions align: (1) a credible mechanistic link between the drug and the new disease, (2) sufficient evidence strength from orthogonal sources, and (3) a realistic translatability path given existing safety, PK, formulation, and IP data.
Use this skill to evaluate proposals, rank candidates, spot weak evidence, and guide the user toward a defensible repurposing strategy before committing to wet-lab or clinical work.
Invoke this skill when the user:
The agent should respond by:
Every repurposing effort starts from one of two directions. The choice determines the databases, methods, and evidence types involved.
| Approach | Starting point | Method | Key databases | Strength | Limitation |
|---|---|---|---|---|---|
| Target-based | Known disease target (gene/protein) | Find approved drugs that bind or modulate the target | DGIdb, ChEMBL, DrugBank, OpenTargets | Mechanistically interpretable; leverages structural biology | Assumes the target is causal and druggable |
| Phenotype-based | Disease molecular signature (transcriptomic, proteomic) | Find drugs whose signature opposes the disease signature | CMap (Connectivity Map), LINCS L1000 | Agnostic to mechanism; can discover unexpected connections | Signatures are noisy; opposing expression ≠ therapeutic effect |
Start
│
├── Do you have a validated disease target (gene/protein)?
│ ├── Yes → Target-based approach.
│ │ Query DGIdb, ChEMBL, DrugBank for drugs hitting this target.
│ │ Cross-reference with OpenTargets for target-disease association strength.
│ └── No → Continue.
│
├── Do you have a disease molecular signature (e.g., differentially expressed genes)?
│ ├── Yes → Phenotype-based approach.
│ │ Query CMap/L1000 for drugs with negatively correlated signatures.
│ │ Validate top hits against known biology.
│ └── No → Continue.
│
├── Do you have both a target and a signature?
│ ├── Yes → Hybrid: run both approaches independently, then intersect.
│ │ Candidates appearing in both lists have stronger support.
│ └── No → Insufficient starting data. Gather target evidence or generate
│ disease signatures before attempting repurposing.
Query multiple databases — no single source is sufficient.
| Database | Content | Use for |
|---|---|---|
| OpenTargets | Target-disease association scores (genetics, literature, pathways) | Validating target relevance; scoring association strength |
| DGIdb | Drug-gene interactions from 30+ sources | Finding approved drugs that interact with a target gene |
| ChEMBL | Bioactivity data (IC50, Ki, EC50) | Quantitative binding/activity evidence; SAR context |
| DrugBank | Drug targets, PK, indications, interactions | Drug metadata, safety profiles, formulations |
| CMap / LINCS L1000 | Expression signatures of ~20,000 compounds | Phenotype-based repurposing; signature reversal |
| TTD | Targets, drugs, clinical trial status | Cross-referencing target validation and development stage |
| STITCH | Chemical-protein interaction network | Network-level drug-target-disease relationships |
Not all evidence is equal. Rank candidates by the strongest evidence type supporting them.
| Tier | Evidence type | Description | Weight |
|---|---|---|---|
| 1 (strongest) | Genetic association via Mendelian randomization (MR) | Genetic variants that proxy drug action associate with disease risk; mimics a natural RCT | Highest — causal inference without confounding |
| 2 | Clinical observation | Case reports, retrospective cohorts, or claims data showing the drug affects the new indication in humans | High — real-world human evidence, but confounded |
| 3 | In vitro / in vivo experimental | Cell-based assays, animal models showing efficacy in the disease context | Moderate — mechanism confirmed but translation uncertain |
| 4 (weakest) | Computational prediction | Signature matching, network proximity, docking scores, AI predictions | Lowest — hypothesis-generating only; requires experimental validation |
Start
│
├── Is there Mendelian randomization evidence (Tier 1)?
│ ├── Yes → Strong foundation. Proceed to translatability assessment.
│ │ Still seek Tier 3 confirmation (mechanism in disease model).
│ └── No → Continue.
│
├── Is there clinical observation evidence (Tier 2)?
│ ├── Yes → Promising, but check for confounding.
│ │ Was the analysis adjusted? Was it an emulated target trial?
│ │ Seek Tier 3 confirmation.
│ └── No → Continue.
│
├── Is there experimental evidence (Tier 3)?
│ ├── Yes → Mechanistically supported.
│ │ Is the model relevant to the target disease?
│ │ Seek Tier 1 or Tier 2 to strengthen the case.
│ └── No → Continue.
│
├── Is there only computational evidence (Tier 4)?
│ └── Hypothesis only. Do not recommend for clinical development.
│ Design experiments to generate Tier 3 evidence first.
│
└── No evidence at all?
└── Not a repurposing candidate. Return to target/disease validation.
When a drug is proposed for a new indication, the agent must evaluate whether the drug's mechanism is relevant to the new disease.
Steps:
| MOA evaluation criterion | Pass | Fail |
|---|---|---|
| Drug target participates in disease pathway | Target in top 3 disease pathways (OpenTargets) | No pathway overlap |
| Directionality matches | Drug inhibits overactive target / activates deficient target | Drug action opposes therapeutic need |
| Activity at therapeutic concentrations | IC50/Ki < Cmax at approved dose | IC50 >> Cmax (cannot achieve target engagement) |
| Selectivity acceptable | < 10 off-targets at therapeutic concentration | Broad off-target activity likely to cause toxicity |
A mechanistically sound candidate may still fail if it cannot be practically developed. Evaluate these dimensions:
| Dimension | Favorable | Unfavorable |
|---|---|---|
| Safety profile | Approved drug with well-characterized AE profile; AEs acceptable for new indication severity | Narrow therapeutic index; serious AEs disproportionate to new indication |
| Pharmacokinetics | Drug reaches target tissue at therapeutic concentrations; known human PK | Poor bioavailability in target tissue; no human PK data for relevant route |
| Dose feasibility | Therapeutic effect achievable at or below approved dose | Required dose far exceeds approved dose (new toxicity territory) |
| Formulation | Existing formulation suitable for new indication (e.g., oral for chronic disease) | Requires new formulation (e.g., intrathecal for CNS disease; IV for outpatient) |
| Patent / exclusivity | Generic available or patent allows new-indication development | Active composition-of-matter patent blocks development; no method-of-use patent opportunity |
| Regulatory path | 505(b)(2) pathway available; existing safety data reduces trial burden | No regulatory shortcut; full NDA-equivalent development required |
When multiple drugs emerge from database queries, rank them systematically.
Scoring framework (assign 0–3 per dimension):
| Dimension | 0 | 1 | 2 | 3 |
|---|---|---|---|---|
| Evidence tier | Tier 4 only | Tier 3 only | Two tiers (e.g., 3+4) | Tier 1 or Tier 2 + Tier 3 |
| MOA relevance | No pathway overlap | Pathway overlap, wrong directionality | Correct directionality, weak activity | Correct directionality, strong activity (IC50 < Cmax) |
| Safety profile | Narrow TI, serious AEs | Moderate AEs | Well-tolerated, minor AEs | Excellent safety record for target population |
| PK/dose feasibility | IC50 >> Cmax | IC50 ~ 2–5× Cmax | IC50 ≤ Cmax at approved dose | IC50 << Cmax; wide therapeutic window |
| Regulatory/IP path | Blocked by patent; full NDA | Uncertain IP; standard pathway | 505(b)(2) possible; some existing data | Generic available; extensive safety database |
Total score interpretation:
Once a candidate is selected, the validation path depends on available evidence.
Start
│
├── Tier 1 (MR) + Tier 3 (experimental) evidence available?
│ └── Proceed to clinical validation.
│ Design: Phase II proof-of-concept trial (randomized, controlled).
│ Endpoint: disease-specific primary endpoint.
│ Leverage existing safety data to shorten Phase I or skip it (505(b)(2)).
│
├── Tier 3 (experimental) evidence only?
│ └── Strengthen with additional preclinical work.
│ - Confirm efficacy in second disease model.
│ - Demonstrate target engagement at achievable concentrations.
│ - Then proceed to clinical validation.
│
├── Tier 2 (clinical observation) only?
│ └── Formalize with retrospective study.
│ - Emulated target trial design on EHR/claims data.
│ - Adjust for confounders (propensity score, IV if available).
│ - If positive, proceed to prospective trial.
│
└── Tier 4 (computational) only?
└── Generate experimental evidence first.
- In vitro assay in disease-relevant cell line.
- Confirm target engagement.
- Do not proceed to clinical work on computational evidence alone.
MR provides the strongest non-trial evidence for repurposing by using genetic variants as instruments for drug action. The logic: if a genetic variant that mimics the drug's effect on the target also associates with the disease, the target is causally implicated.
MR quality checklist:
| Criterion | Requirement | Red flag |
|---|---|---|
| Instrument strength | F-statistic > 10 per SNP | F < 10 indicates weak instruments; results biased toward confounded observational estimate |
| Relevance | SNPs must associate with the target gene's expression or protein level (cis-pQTL or cis-eQTL) | Trans-pQTLs may act through intermediaries; less specific |
| Independence | No linkage disequilibrium between instruments | Correlated SNPs inflate apparent strength |
| Exclusion restriction | SNPs affect disease only through the target | Horizontal pleiotropy violates the core assumption |
| Pleiotropy tests | MR-Egger intercept non-significant; weighted median and MR-PRESSO consistent with IVW | Significant MR-Egger intercept = pleiotropy present |
| Replication | Consistent across ancestries and biobanks | Single-ancestry MR may reflect population-specific LD patterns |
Interpreting MR results for repurposing:
A systematic repurposing screen follows this order:
Step 1: Define the target or disease signature
│
├── Target-based: Identify gene symbol, UniProt ID, Ensembl ID.
│
└── Phenotype-based: Generate or obtain disease DEG list
(log2FC > 1, FDR < 0.05 from case-control transcriptomics).
Step 2: Query primary databases
│
├── Target-based:
│ ├── OpenTargets → target-disease association score (require > 0.5).
│ ├── DGIdb → list of drugs with known interactions for the target.
│ └── ChEMBL → bioactivity data (IC50, Ki) for each drug-target pair.
│
└── Phenotype-based:
└── CMap/LINCS L1000 → query disease signature → rank drugs by
negative connectivity score (require < -90).
Step 3: Cross-reference and filter
│
├── For each candidate drug:
│ ├── DrugBank → current indications, safety profile, PK parameters.
│ ├── ClinicalTrials.gov → prior trials for this drug + new indication.
│ └── ChEMBL → confirm activity at achievable concentrations (IC50 < Cmax).
│
└── Remove candidates with:
├── IC50 > 10× Cmax (cannot achieve target engagement).
├── Prior Phase II/III failure for the same indication.
├── Unacceptable safety profile for indication severity.
└── Blocked IP (active composition-of-matter patent, no generic).
Step 4: Rank surviving candidates
│
└── Apply the scoring framework (Section 6) across all dimensions.
Prioritize candidates scoring ≥ 12/15.
Step 5: Validate top candidates
│
└── Follow the study design decision tree (Section 7).
Network-based approaches complement direct target queries by considering the drug's effect in the context of protein-protein interaction (PPI) networks and disease modules.
When network pharmacology adds value:
When network pharmacology is insufficient:
Key metrics:
| Metric | Interpretation | Threshold |
|---|---|---|
| Network proximity (z-score) | How close drug targets are to disease genes in PPI network | z < -2 (significantly closer than random) |
| Overlap coefficient | Fraction of drug targets that are disease genes | > 0 required; > 0.1 is notable |
| Separation score (SAB) | Topological separation between drug and disease modules | SAB < 0 indicates overlap |
Rules:
Wrong: Treating a negative CMap connectivity score as proof of therapeutic efficacy Right: Treat CMap results as hypothesis-generating (Tier 4 evidence) and validate with experimental or clinical data Why: Cell line transcriptomic signatures do not recapitulate in vivo disease biology — opposing expression does not equal therapeutic effect
Wrong: Using a DGIdb "interacts with" hit without checking whether the drug inhibits or activates the target Right: Always verify interaction directionality (inhibitor vs activator) and confirm it matches the therapeutic need Why: An inhibitor of a target that needs activation will worsen the disease
Wrong: Treating any ChEMBL binding hit as evidence of therapeutic activity Right: Always compare IC50/Ki to Cmax at the approved dose — require IC50 < Cmax for serious consideration Why: A drug may bind a target at concentrations far above what is achievable in vivo, making the interaction clinically irrelevant
Wrong: Basing a repurposing decision on a single database query Right: Cross-reference at least two independent databases (e.g., DGIdb + ChEMBL, or OpenTargets + DrugBank) Why: Each database has biases and gaps — DGIdb may miss recent data, ChEMBL lacks clinical context, DrugBank has limited disease associations
Wrong: Proposing a drug with serious side effects for a mild condition based on mechanistic support alone Right: Evaluate the risk-benefit ratio relative to indication severity before advancing any candidate Why: The risk-benefit test fails regardless of mechanistic support when side effects are disproportionate to the condition
Wrong: Investing effort in a scientifically sound candidate without checking patent status Right: Check patent landscape early — look for composition-of-matter blocks and method-of-use opportunities Why: IP barriers can prevent development regardless of scientific merit, wasting time and resources
Wrong: Assuming the approved dose will achieve target engagement for the new indication Right: Check PK/PD modeling for the new target — the required exposure may differ from the original indication Why: The approved dose was optimized for the original target; the new target may require different concentrations
Wrong: Presenting network proximity (drug-target-disease PPI distance) as validation evidence Right: Treat network-based predictions as Tier 4 (hypothesis-generating) and require experimental confirmation Why: Computational network proximity is insufficient for clinical development decisions
Wrong: Proposing a drug for a tissue it cannot reach (e.g., non-BBB-penetrant drug for a CNS disease) Right: Verify tissue-specific distribution — confirm the drug reaches the disease-relevant tissue at therapeutic concentrations Why: A validated target in the disease tissue is irrelevant if the drug cannot access that tissue
Wrong: Advancing a repurposing candidate without searching ClinicalTrials.gov for prior attempts Right: Always check for prior failed trials for the same drug-indication pair before investing resources Why: A Phase II failure for the same indication is a strong negative signal unless the failure was due to trial design rather than efficacy
Wrong: Citing a single underpowered Mendelian randomization study as definitive Tier 1 evidence Right: Require valid instruments (F-statistic > 10), no horizontal pleiotropy (MR-Egger intercept), and replication across ancestries Why: A single underpowered MR study without sensitivity analyses does not meet the Tier 1 evidence standard