A 12-step small-molecule target-discovery pipeline, developed and validated on glioblastoma (GBM).
A 12-step computational pipeline that nominates novel, druggable small-molecule inhibition targets in glioblastoma (GBM) by integrating a genetic-dependency core (DepMap CRISPR) with a therapeutic-window filter (single-cell tumor-vs-microenvironment, normal-tissue safety, tumor enrichment, tractability), then applying LLM-selected category weights and an LLM nomination ensemble — with built-in ablations that test whether the nominations are driven by evidence or by gene-name recognition.
The pipeline screens 17,916 genes → 283 glioma-selective candidates → a ranked shortlist. Its distinctive, deliberately provocative lead is KIF2C, whose glioma dependency tracks the recurrent 9p21/CDKN2A deletion as a synthetic-lethal relationship — a novel, testable, and as-yet-unvalidated hypothesis.
Single model throughout (no cross-model consensus). Feature definitions were specified by the scientist; the pipeline executes scoring, weighting, and nomination. These are computational hypotheses, not validated targets — see Limitations.
| Genes screened (DepMap 24Q4) | 17,916 |
| Glioma-selective candidates | 283 (236 pan-glioma-dependent, 47 subtype-dependent) |
| Top-ranked candidate | KIF2C (novel tier; 9p21/CDKN2A synthetic-lethality lead) |
| Robust novel picks (survive gene-masking) | VRK1 (masked nom. freq 0.99), ELAVL1 (0.92) |
| Name-inflated (collapse when masked) | KIF18A (1.00 → 0.00), CDK7 (1.00 → 0.03) |
| Positive controls recovered | 4/5 in the top 12 — PDGFRA #2, MET #4, VRK1 #7, FGFR1 #12 (PTPN11 #74, correctly down-weighted as common-essential) |
The gene-masked ablation (step 11a) is the methodological centerpiece: re-running the LLM nomination on the same evidence with gene identities replaced by opaque codes separates candidates whose support is real from those riding name recognition.
See docs/FINAL_REPORT.md for the full scientific write-up and
docs/METHODS.md for complete methods, versions, and thresholds.
| Step | Script | What it does |
|---|---|---|
| 0 | step00_fetch_data.py |
Download all external inputs (DepMap, GBmap, GTEx, Lake, Xena) |
| 1 | step01_dependency.py |
Glioma-selective dependency: Cohen's d + Welch t + BH-FDR over CRISPR gene effect → 283 candidates |
| 3 | step03_synthetic_lethality.py |
Genotype-stratified SL vs 9 recurrent GBM lesions + expression-context paralog scan |
| 4 | step04_tumor_vs_tme.py |
Pseudobulk GBmap single-cell: malignant-vs-microenvironment specificity |
| 5 | step05_therapeutic_window.py |
Graded safety over GTEx + brain subregions + Lake normal-brain neurons |
| 6 | step06_tumor_enrichment.py |
TCGA-GBM vs GTEx-brain log2FC + prevalence (UCSC Xena Toil recompute) |
| 7 | step07_tractability.py |
Open Targets / ChEMBL / DGIdb / UniProt → graded tractability + a separate novelty axis |
| 8 | step08_feature_matrix.py |
Assemble the 283 × 12 feature matrix into 7 evidence categories; normalize |
| 9 | step09_weight_selection.py |
LLM distributes 100 points across the 7 categories (100 expert + 100 non-expert runs) |
| 10 | step10_nomination.py |
LLM nominates the strongest candidates (100 runs); assemble the master 283-gene ranking |
| 11 | step11_ablations.py |
Ablations: composite-only vs +LLM, expert vs non-expert, weight jitter, control recovery |
| 11a | step11a_gene_masked_ablation.py |
Blinded re-nomination on anonymized candidates — evidence vs name recognition |
| 11b | step11b_kif2c.py |
KIF2C 9p21/CDKN2A SL deep-dive: genotype-selective dependency, mechanism, novelty ledger |
(Step 2 was merged into step 1 during the original analysis; the numbering is kept for
traceability with docs/METHODS.md.)
Run order and dependencies are encoded in src/run_all.py.
# 1. Environment (conda or pip)
conda env create -f environment.yml && conda activate gbm-target-discovery
# or: python -m venv .venv && source .venv/bin/activate && pip install -r requirements.txt
# 2. Reproduce the published results offline (cached mode — no downloads, no API keys)
python src/run_all.py
# Inspect / run individual steps
python src/run_all.py --list
python src/step09_weight_selection.py
python src/run_all.py --only 09 10 11aEvery script is runnable standalone (python src/stepNN_*.py) and writes its outputs
to results/stepNN_*/.
The pipeline has exactly two stages that are not pure computation — connector queries
(host.mcp) and LLM calls (host.llm). Both are cache-first:
- cached (default) — reproduces the published results from the committed
data/cache/(connector responses) andresults/(including the exact 100-run LLM outputs). No network, no API key. Fully offline and deterministic. - live (
GBM_LIVE=1) — re-downloads raw inputs and re-queries the live APIs. The LLM steps need theanthropicpackage andANTHROPIC_API_KEY. Because those steps are stochastic, a live re-run draws fresh samples and will not reproduce the committed values bit-for-bit.
export ANTHROPIC_API_KEY=sk-... # live LLM steps only
GBM_LIVE=1 python src/run_all.pyPortability shim. The project was originally executed inside Claude Science, where a
hostobject providedhost.llm/host.mcp.src/adapters.pyre-implements that surface against the public Anthropic API so the code runs unchanged outside the platform; connector-dependent steps fall back to the committed cache.
AI-Driven-Therapeutic-Target-Discovery/
├── src/ # 16 runnable scripts + run_all.py
│ ├── common.py # paths, cache-first run-mode helpers
│ ├── adapters.py # portable host.llm / host.mcp shim
│ ├── figstyle.py # shared figure styling
│ ├── step00_fetch_data.py … step11b_kif2c.py
│ └── run_all.py # orchestrator (--from / --only / --list)
├── data/
│ ├── raw/ # large external inputs (git-ignored; fetched by step00)
│ └── cache/ # small connector/LLM responses (committed → offline repro)
├── results/ # per-step outputs: tables, figures, step reports
│ └── step01_dependency/ … step11b_kif2c/
├── docs/ # METHODS.md, FINAL_REPORT.md, summary, figure captions
├── skill/gbm-target-discovery/ # Claude Science skill wrapping the pipeline
├── requirements.txt · environment.yml · LICENSE · .gitignore
| Source | Version / access | Committed? |
|---|---|---|
| DepMap | 24Q4 Public (figshare 27993248): CRISPR/Chronos gene effect, 72 glioma vs 1,106 non-glioma lines | raw (fetched) |
| GBmap single-cell | CELLxGENE Census core atlas, 338,564 cells | raw (fetched, ~8.1 GB) |
| GTEx | v8, 54 tissues, bulk median TPM | cache (committed) |
| Lake normal brain | GEO GSE97930 snDrop-seq (frontal/visual cortex + cerebellar hem) | raw (fetched) |
| TCGA-GBM | UCSC Xena Toil recompute, 153 GBM vs 1,152 GTEx brain | cache/raw |
| Open Targets, ChEMBL, DGIdb, UniProt | tractability, mechanism, drug-gene, protein class | cache (committed) |
Raw inputs total ~8.6 GB and are not committed (see data/raw/README.md);
step00_fetch_data.py downloads them from source. The ~3.2 MB of connector/LLM caches
are committed so the API-dependent steps reproduce offline.
- Hypotheses, not validated targets. Every nomination is computational. The KIF2C–9p21 synthetic-lethality lead in particular is novel and experimentally unvalidated — it is offered as a testable hypothesis, not a conclusion.
- Single model, no consensus. Weights and nominations come from one model family; the expert/non-expert and gene-masked ablations quantify the framing/name-recognition effects but do not remove them.
- Small-molecule lane only. Tumor-suppressor and CAR-T / surface-antigen targets are out of scope by design; a gene flagged as an SL-partner of a lost suppressor is an annotation, not a small-molecule nomination.
- LLM steps are stochastic. Exact reproduction requires the committed cache; a live re-run will differ at the margin.
The pipeline code is MIT-licensed (LICENSE). The external datasets carry
their own licenses and terms of use — cite the original sources (DepMap, GBmap/CELLxGENE,
GTEx, Lake et al. 2018, TCGA/Xena, Open Targets, ChEMBL, DGIdb, UniProt) when using their
data. Versions and accessions are in docs/METHODS.md.