FREE Reverse Engineering Self-Study Course HERE
Fusion Oncology is an open-source precision oncology platform (44 modules, 19,000+ lines, 528 tests) that fuses XGBoost drug-sensitivity models with DNABERT-2 genomic embeddings into a joint feature space, then routes predictions through ODE digital-twin simulation, two-compartment PK/PD, RL dose optimisation, GNN interaction scoring, and Bayesian uncertainty quantification to produce confidence-scored, AMP/ASCO/CAP-tiered companion diagnostic reports.
| Component | Method | What It Captures |
|---|---|---|
| Drug sensitivity | XGBoost on GDSC LN_IC50 features | Which genes best discriminate cancer types |
| Engineered | 10 row-level distributional features | Per-sample mean, std, skew, kurtosis, IQR, CV |
| Genomic context | DNABERT-2 768-dim embeddings, PCA to 50 | Noise-reduced sensitivity-weighted gene context |
| Fusion model | XGBoost on concatenated (N + 10 + 50) | Jointly learned drug + distribution + sequence |
The fusion model produces one unified set of CV metrics (Accuracy, Precision, Recall, F1, F2, ROC AUC). The Fusion Index (importance x instability x 1000) ranks targets that are both biologically important and structurally vulnerable.
- Pathway enrichment -- maps targets to PI3K-Akt, MAPK, p53, Wnt, Notch, cell-cycle, DNA-repair, apoptosis, angiogenesis, and immune-checkpoint pathways
- Drug-target mapping -- cross-references 20+ gene targets against approved oncology drugs (EGFR-Osimertinib, BRAF-Vemurafenib, KRAS-Sotorasib, etc.)
- Survival analysis -- Kaplan-Meier + log-rank stratification hooks
- Publication-quality figures -- bar charts, scatter plots, heatmaps, box plots
- Self-contained HTML report -- one-click shareable with collaborators
- Multi-omics integration -- MAF mutation parsing, copy-number alteration analysis, methylation profiling, and combined feature matrix construction
- Clinical evidence aggregation -- real-time queries to OpenTargets, CIViC, and ClinicalTrials.gov APIs with composite evidence scoring
- Synthetic lethality detection -- curated database of 24 SL pairs (BRCA1/2-PARP, RB1-Aurora kinase, etc.) plus expression-based anti-correlation screening
- Neoantigen prediction -- codon translation, mutant peptide generation, simplified MHC-I binding scoring for immunotherapy candidate ranking
- Resistance prediction -- 12-gene resistance mechanism database with risk scoring, drug-specific evasion strategies, and clinical counter-measures
- Network pharmacology -- drug-gene-pathway tripartite interaction graph with degree/betweenness centrality, polypharmacology scoring, and combination target identification
- CRISPR guide design -- PAM scanning on both strands, Doench-inspired on-target scoring, off-target heuristics, and exportable guide libraries
- Companion diagnostics -- patient-level mutation profiling, AMP/ASCO/CAP actionability tiering (Tiers I-IV), drug matching, and ranked treatment plans
- Digital twin simulation -- Gompertzian tumour growth ODE model with drug regimens (cycling/scheduling), immune dynamics, RECIST response classification, and regimen comparison
- GNN interaction learning -- graph neural network replacing static centrality with learned drug-gene-pathway embeddings via message passing
- RL treatment optimiser -- policy-gradient agent that learns optimal dosing schedules using the digital twin as a gymnasium-style environment
- Compartmental PK/PD -- two-compartment pharmacokinetics with Emax pharmacodynamics, replacing a flat efficacy constant in the digital twin
- Immune micro-environment model -- structured TIME compartment with CD8+ T cells, Tregs, MDSCs, NK cells, and cytokine-mediated dynamics
- Bayesian uncertainty quantification -- posterior credible intervals and calibrated confidence scores for clinical predictions
- SHAP interpretability -- gene-level and pathway-level explanations for XGBoost predictions
- Benchmark framework -- ablation studies and baseline comparisons (Random Forest, Logistic Regression, SVM, MLP) with paired t-tests
- Methodology formalisation -- formal mathematical specification and empirical validation of the architecture
- Real clinical validation -- cBioPortal TCGA data ingestion for open-access expression, mutation, and clinical outcome testing
- Domain adaptation -- cell-line to patient distribution alignment bridging the in-vitro / in-vivo gap
# Clone and install
git clone https://github.com/mytechnotalent/fusion_oncology.git
cd fusion_oncology
pip install -e ".[dev]"
# Run the full 7-step pipeline
fusion-oncology run
# Faster smoke test
fusion-oncology run --top-k 3 --fuzz-iterations 5
# Download and cache the data
fusion-oncology ingest
# Query clinical evidence for a gene
fusion-oncology evidence --gene EGFR
# Check resistance mechanisms
fusion-oncology resistance --gene BRAF
# Run a digital twin tumour simulation
fusion-oncology simulate --days 180
# Run companion diagnostics
fusion-oncology companion-dx --mutations "EGFR:T790M,BRAF:V600E" \
--cancer-type NSCLC
# Regenerate a report from saved results
fusion-oncology report results/fusion_results.csvRun on Kaggle with the GDSC dataset (1,000+ cancer cell lines):
# Upload kaggle_notebook.ipynb to Kaggle
# Add GDSC dataset from:
# kaggle.com/datasets/samiraalipour/genomics-of-drug-sensitivity-in-cancer-gdsc
# Then run the notebook or use CLI:
!pip install git+https://github.com/mytechnotalent/fusion_oncology.git
!fusion-oncology run --top-k 10 --output-dir /kaggle/working/resultsSee KAGGLE_GUIDE.md for complete GDSC setup and alternative datasets.
fusion_oncology/
pyproject.toml
Makefile
install.sh
docs/
architecture.md
src/fusion_oncology/
__init__.py
config.py # Central dataclass (29 fields)
cli.py # Click CLI (8 commands)
data/
ingestion.py # TCGA download + ZIP extraction
preprocessing.py # Variance filter, log-norm, PCA
cache.py # Filesystem artefact cache
multi_omics.py # MAF mutations, CNA, methylation
gdsc.py # GDSC dose-response loader
tcga.py # TCGA patient cohort loader
domain_adaptation.py # Cell-line to patient alignment
models/
xgboost_engine.py # XGBoost training + importance
dnabert_engine.py # DNABERT-2 sequence embedding
fusion.py # Multi-modal fusion pipeline
crispr.py # CRISPR guide design + scoring
companion_dx.py # AMP/ASCO/CAP companion Dx
digital_twin.py # Gompertzian tumour ODE model
pharmacokinetics.py # Two-compartment PK/PD model
rl_optimizer.py # RL dosing optimiser
analysis/
instability.py # Mutation fuzzing + cosine drift
pathway.py # KEGG/Reactome pathway lookup
drug_target.py # Drug-target annotation
survival.py # Kaplan-Meier survival hooks
clinical_evidence.py # OpenTargets + CIViC + CT.gov
synthetic_lethality.py # SL pair detection + screening
neoantigen.py # Peptide + MHC binding scoring
resistance.py # Resistance mechanism prediction
network_pharmacology.py # Drug-gene-pathway network
gnn_network.py # GNN message-passing embeddings
immune_model.py # TIME immune compartment
interpretability.py # SHAP model explanations
uncertainty.py # Bayesian confidence intervals
viz/
plots.py # Matplotlib/Seaborn figures
report.py # HTML report generator
validation/
benchmark.py # Ablation + baseline comparisons
methodology.py # Architecture formalisation
real_data.py # cBioPortal TCGA validation
utils/
bio.py # Entrez fetch, GC, CpG islands
log.py # Logging setup
tests/
conftest.py # Shared fixtures, OMP guard
test_benchmark.py
test_bio.py
test_cache.py
test_cli.py
test_clinical_evidence.py
test_companion_dx.py
test_config.py
test_crispr.py
test_digital_twin.py
test_dnabert_engine.py
test_domain_adaptation.py
test_drug_target.py
test_fusion_engine.py
test_gdsc.py
test_gnn_network.py
test_immune_model.py
test_ingestion.py
test_instability.py
test_interpretability.py
test_log.py
test_methodology.py
test_multi_omics.py
test_neoantigen.py
test_network_pharmacology.py
test_pathway.py
test_pharmacokinetics.py
test_plots.py
test_preprocessing.py
test_real_data.py
test_report.py
test_resistance.py
test_rl_optimizer.py
test_survival.py
test_synthetic_lethality.py
test_tcga.py
test_uncertainty.py
test_xgboost_engine.py
$ fusion-oncology --help
Usage: fusion-oncology [OPTIONS] COMMAND [ARGS]...
Fusion Oncology - Precision cancer therapeutics platform
Commands:
clear-cache Delete all locally cached artefacts
companion-dx Run companion diagnostic analysis on patient mutations
evidence Query clinical evidence databases for genes
ingest Download and cache the TCGA Pan-Cancer dataset
report Regenerate HTML report from saved results
resistance Predict resistance mechanisms for genes
run Run the full 7-step fusion analysis pipeline
simulate Run digital twin tumour growth simulation
Query clinical evidence from OpenTargets, CIViC, and ClinicalTrials.gov.
$ fusion-oncology evidence EGFR BRAFClinical Evidence for EGFR:
OpenTargets: 85.0% confidence
ClinicalTrials.gov: 20 trials
CIViC: Data unavailable
Clinical Evidence for BRAF:
OpenTargets: 87.7% confidence
ClinicalTrials.gov: 15 trials
CIViC: Data unavailable
Composite Evidence Scores:
EGFR: 85.0% (high actionability)
BRAF: 87.7% (high actionability)
Predict resistance mechanisms and evasion strategies for drug targets.
$ fusion-oncology resistance EGFR ALKResistance Profile for EGFR:
Mechanisms:
1. T790M Gatekeeper Mutation
Risk: HIGH | Clinical: Very common in osimertinib resistance
Strategy: Switch to 3rd-gen TKI or 4th-gen EGFR-selective
2. MET Amplification
Risk: MEDIUM | Clinical: Bypass signaling in 5-20% of cases
Strategy: Combine EGFR-TKI with MET inhibitor
3. HER2 Amplification
Risk: MEDIUM | Clinical: Emerging bypass in 5-12%
Strategy: Dual EGFR-HER2 blockade
Resistance Profile for ALK:
Mechanisms:
1. ALK G1202R Gatekeeper Mutation
Risk: HIGH | Clinical: Most common in lorlatinib resistance
Strategy: Investigational 4th-gen ALK inhibitor or TPX-0131
2. ALK L1196M Mutation
Risk: HIGH | Clinical: Sensitive to lorlatinib
Strategy: Switch to 3rd-gen lorlatinib
Simulate tumour growth dynamics with drug regimens using a Gompertzian ODE model.
$ fusion-oncology simulate --drug Osimertinib --efficacy 0.15 --days 90Digital Twin Tumour Simulation
Configuration:
Initial Volume: 10000.0 mm3
Simulation Duration: 90 days
Drug: Osimertinib
Efficacy: 0.15/day
Final Results (Day 90):
Tumour Volume: < 0.001 mm3
Reduction: 100.0%
RECIST Response: CR (Complete Response)
Trajectory saved to: digital_twin_trajectory.csv
Options: --days, --drug, --efficacy, --days-on, --days-off,
--initial-volume, --growth-rate, --carrying-capacity, --output
Generate treatment recommendations for patient mutation profiles.
$ cat > patient_mutations.json << 'EOF'
{
"patient_id": "PT-2024-001",
"cancer_type": "Non-Small Cell Lung Cancer",
"mutations": [
{"gene": "EGFR", "variant": "L858R", "vaf": 0.42},
{"gene": "TP53", "variant": "R273H", "vaf": 0.38},
{"gene": "KRAS", "variant": "G12C", "vaf": 0.15}
]
}
EOF
$ fusion-oncology companion-dx patient_mutations.json=== COMPANION DIAGNOSTIC REPORT ===
Patient: PT-2024-001
Cancer Type: Non-Small Cell Lung Cancer
Detected Mutations:
EGFR L858R (VAF: 42.0%) - Tier I
TP53 R273H (VAF: 38.0%) - Tier III
KRAS G12C (VAF: 15.0%) - Tier I
Treatment Recommendations (Ranked by Confidence):
1. Osimertinib (EGFR L858R) -- 90% confidence
AMP/ASCO/CAP Tier: I
Evidence: FDA-approved for EGFR-mutant NSCLC
2. Sotorasib (KRAS G12C) -- 78% confidence
AMP/ASCO/CAP Tier: I
Evidence: FDA-approved for KRAS G12C NSCLC
3. Afatinib (EGFR L858R) -- 72% confidence
AMP/ASCO/CAP Tier: I
Resistance Risk Alerts:
EGFR T790M Gatekeeper Mutation (HIGH risk)
MET Amplification (MEDIUM risk)
Synthetic Lethality Opportunities:
CHK1 inhibition (TP53 R273H synthetic lethal)
WEE1 inhibition (TP53 R273H synthetic lethal)
Clinical Summary:
High-confidence actionable targets: 2 (EGFR, KRAS)
FDA-approved matched therapies: 6
Primary recommendation: Osimertinib monotherapy
# Full analysis
$ fusion-oncology run
# Fast smoke test
$ fusion-oncology run --top-k 3 --fuzz-iterations 5
# High-resolution analysis
$ fusion-oncology run --top-k 20 --fuzz-iterations 50 --xgb-trees 200Options: --top-k, --fuzz-iterations, --xgb-trees, --xgb-depth,
--output-dir, --log-level, --skip-plots, --skip-report
$ fusion-oncology ingestDownloads the TCGA Pan-Cancer dataset (801 samples x 20,531 genes) from the UCI Machine Learning Repository and caches it locally.
$ fusion-oncology report results/fusion_results.csvRegenerates the HTML report from previously saved results.
$ fusion-oncology clear-cacheClears all locally cached data and model artefacts.
GDSC Drug Sensitivity Data
|
+---------------+---------------+
| |
Step 1: XGBoost Baseline Step 2: DNABERT-2
(multi-class on LN_IC50) (768-dim gene embeddings)
| |
Top-K gene importances Embedding matrix (K x 768)
| |
+--------> Step 3: Fusion <-----+
Weight embeddings by
drug sensitivity per
cell line -> 768-dim
context vector
|
Concatenate: original
features + 10 engineered
+ 50 PCA DNABERT-2 dims
|
Step 4: Fusion XGBoost
5-fold x 3-repeat CV
-> unified metrics
|
+-------+-------+-------+
| | | |
Step 5: Step 6: Step 7: Advanced:
Pathway Drug Resist. GNN, PK/PD,
Enrich. Annot. RL, Immune,
Twin, CDx
# Install dev dependencies
make dev
# Run tests (528 tests, ~45 s)
make test
# Lint
make lint
# Type-check
make typecheck
# Coverage report
make test-cov| Package | Purpose |
|---|---|
xgboost |
Gradient-boosted gene importance |
transformers + torch |
DNABERT-2 sequence embeddings |
biopython |
NCBI Entrez sequence retrieval |
scikit-learn |
Preprocessing, CV, cosine similarity |
pandas + numpy |
Data manipulation |
scipy |
Statistical tests (Spearman, log-rank) |
matplotlib + seaborn |
Visualisation |
lifelines |
Kaplan-Meier survival analysis |
requests |
API queries (OpenTargets, CIViC, CT.gov) |
pyarrow |
Parquet file I/O for caching |
optuna |
Hyperparameter tuning (RL optimiser) |
shap |
Model interpretability explanations |
click + rich |
CLI interface |
TCGA Pan-Cancer (PANCAN) HiSeq RNA-Seq from the UCI Machine Learning Repository: https://archive.ics.uci.edu/dataset/401/gene+expression+cancer+rna+seq
- 801 samples x 20,531 genes
- 5 cancer types: BRCA, KIRC, COAD, LUAD, PRAD
MIT LICENSE