flowchart LR
A[Open Targets + curated rules] --> B[Deterministic features]
B --> C[Stable role]
C --> D[Three intent rankings]
D --> E[Cards, reports and figures]
F[Reviewed evidence] --> G[Read-only report layers]
H[DepMap 26Q1] --> G
G --> E
TargetIntel-IO
Explainable, therapeutic-intent-aware target triage
A deterministic research platform for classifying and ranking candidate targets in anti-PD-1-resistant melanoma.
v0.6.0 · Research preview View repository
The problem
A gene can be highly relevant to disease and still be the wrong direct therapeutic target. In anti-PD-1-resistant melanoma, the same candidate might instead be a resistance biomarker, an immune-context marker, a tumour-intrinsic dependency or a mechanism with poor tractability.
TargetIntel-IO does not compress those possibilities into one universal score. It asks a more useful question: valuable for what therapeutic purpose?
What the tool does
The platform retrieves and structures evidence, assigns stable biological roles, and produces separate rankings for:
- Antibody / IO combination strategies;
- Resistance biomarker strategies;
- Tumour-intrinsic small-molecule strategies.
The deterministic baseline remains authoritative. Optional evidence, feasibility and functional-dependency layers can decorate a report, but they cannot silently change baseline features, roles, scores or ranks.
Why this design matters
| Candidate pattern | More defensible interpretation |
|---|---|
| Immune checkpoint with combination rationale | Antibody / IO-combination candidate |
| Loss associated with resistance | Biomarker or mechanistic marker |
| Selective tumour-cell dependency | Tumour-intrinsic intervention candidate |
| Broad essentiality or poor modality fit | Biologically relevant, but weak direct target |
This separation makes a result easier to challenge, reproduce and discuss with experimental teams.
Current release
TargetIntel-IO has progressed through five released layers:
- v0.1.3: deterministic role classification and intent-specific ranking;
- v0.2.0: typed, source-linked Common Evidence Layer;
- v0.3.0: audited extraction, review and grounded-synthesis infrastructure;
- v0.4.0: modality-specific target feasibility;
- v0.5.0: optional DepMap/CRISPR functional-dependency architecture and portable reports;
- v0.6.0: graph-native evidence—an immutable evidence graph with deterministic local retrieval and a provider-neutral, GraphRAG-compatible export. The decision plane remains authoritative: graph output cannot promote a candidate into the productive ranking without an explicit validation and authorization gate.
The productive baseline contains 300 genes. The current discovery universe contains 331 unique identities. The full 18,531-gene DepMap background is context only—not a new productive ranking.
Validation without overclaiming
The internal 56-target benchmark measures implementation consistency:
| Metric | Current snapshot |
|---|---|
| Open Targets retrieval coverage | 25 / 56 (44.6%) |
| Stable-role accuracy on covered targets | 100.0% |
| Strict primary-intent accuracy | 91.1% |
| Acceptable-intent accuracy | 100.0% |
A 42-scenario local sensitivity analysis produced a minimum Spearman rank correlation of 0.8762. These results test deterministic behaviour around curated rules. They are not independent biological or clinical validation.
Outputs
The CLI generates:
- deterministic feature tables and three ranked target lists;
- target cards and HTML reports;
- score and rank-shift figures;
- benchmark and sensitivity snapshots;
- optional reviewed-evidence and feasibility sections;
- optional portable DepMap/CRISPR reports.
Scientific boundary
TargetIntel-IO is research-use software. It does not recommend treatment, establish causality, validate a therapeutic target or biomarker, or predict patient response. The immediate future direction is external validation with public clinical response cohorts and single-cell/spatial context while preserving the same evidence boundaries.
Explore the implementation, reports and versioned examples.