Transparent, deterministic therapeutic-intent target triage for anti-PD-1-resistant melanoma.
TargetIntel-IO helps research teams distinguish possible therapeutic targets, biomarkers, resistance mechanisms, immune-context markers, tumor-intrinsic drivers, and poor direct targets while preserving the rules and evidence boundaries behind each result.
Research-use software only. It generates hypotheses and portfolio-triage outputs, not medical advice, treatment recommendations, validated therapeutic targets, qualified biomarkers, causal proof, or clinical-response predictions.
| Layer | Status | Purpose |
|---|---|---|
| Deterministic baseline | Available | Role classification and therapeutic-intent ranking |
| Common Evidence Layer | Complete | Typed source-linked evidence and read-only report decoration |
| Grounded evidence infrastructure | Complete | Audited extraction, review and grounded synthesis |
| Target feasibility | Complete | Offline modality-specific feasibility and coverage |
| Functional dependency | Research preview | DepMap/CRISPR evidence, closure, portable reporting and human review |
| Graph-native evidence | Research preview | Immutable evidence graph, deterministic retrieval and BYOG export |
| Single-cell / clinical context | Development preview | Precomputed descriptive summaries with explicit sample and patient boundaries |
The project has progressed through v0.1.3 deterministic therapeutic-intent baseline, v0.2.0 Common Evidence Layer, v0.3.0 grounded evidence and human-review infrastructure, v0.4.0 target feasibility, v0.5.0 DepMap/CRISPR functional-dependency architecture, and v0.6.0 graph-native evidence retrieval. A sanitized, portable DepMap Public 26Q1 aggregate bundle and versioned research-preview reports are published. Published reports retain the complete frozen productive baseline score and rank pairs for all three therapeutic intents; the reporting layer does not infer or recalculate them. The historical status text “real DepMap Public 26Q1 repository snapshot publication is pending Issue 512” does not describe this release.
Anti-PD-1-resistant melanoma is biologically heterogeneous. A relevant gene is not automatically a direct drug target: it may instead support biomarker research, describe resistance context, mark an immune compartment, or be unsuitable for direct intervention. TargetIntel-IO therefore produces separate therapeutic-intent rankings rather than one undifferentiated “best target” list.
flowchart TD
OT[Open Targets + curated rules] --> FT[Deterministic feature table]
FT --> RC[Stable role classification]
RC --> TR[Therapeutic-intent rankings]
TR --> OUT[Cards / HTML / figures]
EV[Reviewed evidence] --> RD[Read-only report decoration]
FE[Target feasibility] --> RD
DP[Validated local DepMap release] --> PS[Portable aggregate snapshot]
PS --> EC[Immutable DependencyReportEvidence]
EC --> RH[Optional Markdown / HTML rendering]
RH --> RD
TR --> RD
The deterministic baseline is authoritative. Optional layers decorate reports after ranking; they do not modify feature construction, scores, ranks, roles, target selection, or candidate activation.
The productive v0.1.3 baseline retrieves and annotates melanoma-associated targets, constructs a deterministic feature table, assigns a stable role, and ranks candidates for antibody/IO combination, resistance biomarker, and small-molecule intents. The original productive baseline remains 300 genes. Baseline scores and ranks remain unchanged by all optional reporting layers.
The internally curated 56-target benchmark measures implementation consistency and rule behavior. It is not an independent biological validation dataset and is not evidence of clinical performance.
The v0.2.0 Common Evidence Layer stores typed, source-linked observations separately from interpretation and provides read-only report decoration. v0.3.0 provides audited extraction, mandatory review, reviewed snapshots, and grounded synthesis infrastructure; v0.2.0 is infrastructure and report decoration, not clinical validation or a production LLM extractor. v0.4.0 feasibility reports modality-specific coverage, missingness, contradictions, tractability, clinical precedence, doability, and safety-data state without changing deterministic prioritization.
v0.5.0 is a research-preview architecture for optional, post-ranking functional-dependency reporting in melanoma anti-PD-1 resistance research. DepMap Public 26Q1 is the pinned functional-dependency release. The curated benchmark contains 56 targets, the discovery universe contains 331 unique identities, and the 18,531-gene DepMap universe is background only. That background is not a productive ranking: no new 18,531-gene productive ranking is generated.
Functional-dependency evidence can report gene-effect evidence, dependency-probability evidence, context-versus-reference comparisons, selectivity, and bounded integration-rank observations. It is deliberately distinct from clinical anti-PD-1 response evidence. DepMap cell-line dependency is not clinical anti-PD-1 response evidence, and absence of tumor-cell dependency does not invalidate an immune target.
Implemented components include deterministic local ingestion; immutable file and schema contracts; functional-dependency profiles; benchmark and coverage evaluation; a bounded integration overlay; release closure and reproducibility checks; portable aggregate snapshot export; immutable report evidence contracts; deterministic Markdown and HTML rendering; and regression and isolation gates. Production activation remains disabled, no approved authorization is emitted, and human review remains mandatory.
The optional graph plane converts validated reviewed evidence into immutable
entities and assertions, supports deterministic entity resolution, focused
retrieval, path inspection, communities and provider-neutral GraphRAG
bring-your-own-graph tables. EvidenceItem and ReviewedEvidenceSnapshot remain
the scientific source of truth. Graph retrieval can restrict evidence for the
existing grounded writer, but every selected identifier must resolve exactly to
reviewed evidence; graph-only summary identifiers fail closed rather than being
silently omitted.
Precomputed single-cell and clinical-response summaries can also be represented
as development-only graph context. These adapters do not load raw matrices, run
differential expression, treat cells as patients, predict response, calculate
scores, modify ranks or activate candidates. The graph snapshot remains the
canonical source even when exported to another GraphRAG engine. See
docs/GRAPH_RAG_ARCHITECTURE.md and the
v0.6.0 specification.
git clone https://github.com/rsolerortuno/TargetIntel-IO.git
cd TargetIntel-IO
conda env create -f environment.yml
conda activate targetintelgit clone https://github.com/rsolerortuno/TargetIntel-IO.git
cd TargetIntel-IO
python -m venv .venv
source .venv/bin/activate
python -m pip install --upgrade pip
python -m pip install -e ".[dev,graph]"The normal deterministic workflow does not require a local DepMap release.
targetintel run
targetintel run --validate
targetintel run --refresh
targetintel run --depmap-snapshot data/releases/depmap/DepMap_Public_26Q1
targetintel run --help
targetintel graph demo --output results/graph-demo
targetintel graph validate-bundle results/graph-demo
targetintel graph export-graphrag results/graph-demo --output results/byog
python -m pytest -qThe productive workflow writes a deterministic feature table, therapeutic-intent ranked targets, Markdown cards, HTML reports and figures. Optional reviewed evidence, feasibility, and functional-dependency sections decorate matching reports only. A portable aggregate DepMap Public 26Q1 snapshot is a separate research-preview artifact and never changes scores, ranks, roles, or activation.
Versioned examples include HTML reports, figures, the benchmark snapshot, and sensitivity outputs.
Single-cell and clinical-response summaries remain development-only inputs for the planned v0.7.0 clinical-response research model.
TargetIntel-IO uses deterministic rule application and tie-breaking, versioned configuration and benchmark material, immutable evidence contracts, and offline regression tests. Run targetintel run --validate for the existing workflow validation and python -m pytest -q for the test suite. The 56-target benchmark remains an internal consistency check, not independent validation or clinical performance evidence.
The v0.6.0 graph release has a separate productive-isolation audit: clean baseline and integrated runs produced 56/56 byte-identical artifacts.
The current 56-target benchmark is an internal implementation benchmark:
- Open Targets retrieval coverage is 25/56 (44.6%).
- The covered benchmark produced 100.0% stable-role accuracy.
- Strict primary-intent accuracy is 91.1%.
- Acceptable-intent accuracy is 100.0%.
These results measure implementation consistency, not independent biological accuracy. Complete TargetIntel evaluation coverage does not mean that Open Targets independently recovered every target. The benchmark is internally curated rather than derived from an independent, blinded, prospective, or clinical validation dataset.
The current local sensitivity snapshot evaluates 42 scenarios in which individual scoring weights are perturbed and renormalized.
Worst-case top-5 retention was: antibody/IO 100%, biomarker 100%, small-molecule 80%.
Worst-case top-10 retention was: antibody/IO 90%, biomarker 100%, small-molecule 90%.
Worst-case top-20 retention was: antibody/IO 100%, biomarker 95%, small-molecule 100%.
The minimum observed Spearman rank correlation was 0.8762.
The maximum absolute primary-intent accuracy change was 5.36 percentage points.
The maximum absolute acceptable-intent accuracy change was 3.57 percentage points.
The maximum absolute cross-intent-specificity change was 5.66 percentage points.
These observations describe local robustness around the configured deterministic rules. They do not establish that the rankings are independent of weight selection or biologically validated.
No external patient-level responder/non-responder cohort is currently used to validate the target rankings.
Public associations, internal benchmark agreement, sensitivity analysis, feasibility annotations, reviewed evidence, and DepMap functional-dependency observations do not establish causality, clinical utility, therapeutic efficacy, or response prediction. All generated hypotheses require independent experimental, translational, and clinical validation.
The optional grounded-evidence workflow is documented in examples/llm/README.md, with release boundaries in docs/releases/v0.3.0.md.
This layer requires mandatory human review and does not alter deterministic scores, rankings, or role classification.
TargetIntel-IO does not make treatment recommendations, validate targets or biomarkers, establish causality, or predict patient response. Missing evidence is not negative evidence. DepMap cell-line profiles do not reproduce the complete tumor microenvironment, and broad dependency can reflect general essentiality. The real 26Q1 portable aggregate publication is available under data/releases/depmap/DepMap_Public_26Q1, with reports under examples/html_reports/depmap_26q1 and examples/target_cards/depmap_26q1. Future research directions include single-cell/spatial integration, clinical-response research models, and knowledge-graph expansion.
targetintel/ Reusable package and CLI
targetintel/functional_dependency/ Optional portable reporting architecture
targetintel/graph/ Optional immutable graph and retrieval plane
configs/ Disease, resistance, benchmark and scoring rules
tests/ Unit, integration, regression and isolation tests
examples/ Versioned example outputs
docs/ Specifications, roadmap and release evidence
data/ and results/ Local caches and generated outputs; normally uncommitted
- v0.2.0 Common Evidence Layer
- v0.3.0 Grounded Literature Copilot and provider-agnostic LLM integration
- v0.4.0 Target feasibility and expanded Open Targets integration
- v0.5.0 DepMap/CRISPR functional dependency
- v0.6.0 Single-cell and spatial context plus graph-ready summary contracts
- v0.6.1 Immutable evidence-graph foundation and deterministic queries
- v0.6.2 Provider-neutral GraphRAG BYOG research preview
- v0.7.0 Clinical-response research model — evidence graph with sealed development/holdout scope
- v0.7.1 Clinical multihop and contradiction-preserving GraphRAG research queries
- v0.8.0 De novo target discovery and knowledge graph — explicit algorithms with popularity and source ablations
- v0.9.0 Multitumor transfer and external validation
- v1.0.0 Multitumor target-intelligence platform — validated graph-native release target
Soler Ortuño R. TargetIntel-IO: Explainable therapeutic-intent-aware target
intelligence for anti-PD-1-resistant melanoma.
Rafael Soler Ortuño, PhD
Released under the MIT License.