{"data":{"kind":"file","path":"README.md","version_id":"u953m8v6yazuxsyrhebtfege","entry":{"name":"README.md","path":"README.md","is_directory":false,"size":6423,"modified_at":"2026-09-22T14:44:46.554000","content_hash":"fe88a7b2be99a6a519e1438c31c259454a6af6ce4707dfb547d90a481d545d81"},"entries":[],"content":"# Human Cancer Genetics & Precision Oncology Evaluation Environment (`human-cancer-genetics`)\n\n**Version:** V0.1.0  \n**Author:** jackbull  \n**Tags:** `cancer-genomics`, `precision-oncology`, `molecular-biology`, `genetics`, `verifiers`, `rl-environment`\n\n---\n\n## Overview\n\n`human-cancer-genetics` is a high-fidelity benchmark and reinforcement learning evaluation environment built on the `verifiers` framework for the **Prime Intellect Environments Hub**. It evaluates language models on expert-level molecular biology, somatic and germline cancer genetics, oncogenic driver pathways, and precision targeted therapeutics.\n\nModern precision oncology demands rigorous understanding of gene alteration mechanisms (activating mutations, chromosomal fusions, loss-of-function, epigenetic silencing) paired with pharmacological reasoning (mechanisms of action, resistance mutations, synthetic lethality). This environment probes these capabilities across 16 clinically curated scenarios.\n\n---\n\n## Benchmark Domains & Topics\n\nThe environment evaluates 16 structured challenges organized into 6 core pillars:\n\n### 1. Oncogenes & Targeted Therapeutics\n- **KRAS Codon 12 Mutations & Allele-Specific G12C Targeting (`cg-001`):** Impairment of GAP-mediated GTP hydrolysis, switch-II pocket covalent binding (sotorasib, adagrasib), and bypass resistance mechanisms.\n- **EGFR Exon 19 Deletion / L858R & T790M Gatekeeper (`cg-002`):** Altered ATP affinity ratio, steric clash of Met790 sidechain, and 3rd-generation mutant-selective Cys797 covalent inhibition (osimertinib).\n- **BRAF V600E & Paradoxical MAPK Activation (`cg-003`):** Autonomous monomeric activation loop mimicry, paradoxical CRAF transactivation in RAS-primed cells, and dual BRAF (dabrafenib) + MEK (trametinib) blockade.\n- **HER2 (ERBB2) Amplification & ADCs (`cg-004`):** Constitutive dimerization, HER3 p85 docking for PI3K activation, and bystander killing by membrane-permeable topoisomerase I inhibitors (T-DXd).\n\n### 2. Tumor Suppressor Genes & Knudson's Model\n- **TP53 Inactivation & Dominant-Negative Mutations (`cg-005`):** DNA-binding domain missense hotspots (R175, R248, R273), homotetramer poisoning, MDM2 feedback, and p21 vs PUMA/BAX transcriptional fate.\n- **RB1 & Knudson's Two-Hit Model (`cg-006`):** Familial vs sporadic retinoblastoma genetics, sequential phosphorylation by CDK4/6 and CDK2, E2F release, and CDK4/6 inhibitor requirements.\n- **PTEN Inactivation & PI3K/AKT Hyperactivation (`cg-007`):** 3-phosphatase dephosphorylation of PIP3 to PIP2, PH-domain membrane docking (PDK1/AKT Thr308 and Ser473), and downstream resistance to anti-EGFR therapies.\n- **APC Loss & Canonical Wnt/Beta-Catenin Signaling (`cg-008`):** Destruction complex (APC, Axin, GSK3B, CK1), beta-TrCP ubiquitination degron, TCF/LEF target transactivation (MYC, CCND1), and the Vogelstein adenoma-carcinoma sequence.\n\n### 3. DNA Damage Repair & Synthetic Lethality\n- **BRCA1/BRCA2 Deficiency & PARP Inhibitor Synthetic Lethality (`cg-009`):** 5'-to-3' end resection, PALB2 bridging, RAD51 nucleofilament assembly, replication fork collapse via trapped PARP-DNA complexes, and secondary reversion mutations.\n- **DNA Mismatch Repair (dMMR), MSI-H & Immunotherapy (`cg-010`):** MutS/MutL heterodimer proofreading, polymerase slippage at microsatellites, frameshift neoantigen generation, and anti-PD-1 checkpoint sensitivity.\n- **ATM/ATR Checkpoint Signaling in Genomic Stability (`cg-011`):** Double-strand break sensing by MRN vs RPA-coated single-stranded DNA sensing, CHK1/2 cascades, CDC25 regulation, and synthetic lethality with ATR/WEE1 inhibitors.\n\n### 4. Chromosomal Alterations & Chimeric Fusions\n- **BCR-ABL1 t(9;22) & T315I Gatekeeper Mutation (`cg-012`):** Philadelphia chromosome reciprocal translocation, coiled-coil oligomerization domain, imatinib DFG-out binding, T315I steric clash, and allosteric myristoyl pocket targeting (asciminib) or ethynyl-linkage bypass (ponatinib).\n- **EML4-ALK Inversion & Solvent-Front Resistance (`cg-013`):** Paracentric inv(2)(p21p23), ligand-independent autophosphorylation, TKI evolution, and macrocyclic overcome of G1202R solvent-front resistance (lorlatinib).\n\n### 5. Epigenetics & Metabolic Reprogramming\n- **IDH1/IDH2 Neomorphic Oncometabolite & CIMP (`cg-014`):** Conversion of alpha-ketoglutarate to (R)-2-hydroxyglutarate (2-HG), competitive inhibition of TET2 and JmjC histone demethylases, CpG Island Methylator Phenotype (CIMP), and targeted IDH inhibitors (ivosidenib, vorasidenib).\n- **Promoter CpG Island Hypermethylation (`cg-015`):** De novo (DNMT3A/B) and maintenance (DNMT1) methylation, MeCP2/HDAC chromatin condensation, silencing of MLH1/CDKN2A/BRCA1, and irreversible covalent trapping by 5-azacitidine and decitabine.\n\n### 6. Somatic vs Germline Cancer Genetics\n- **Lynch Syndrome vs Sporadic MSI-H Differential (`cg-016`):** Autosomal dominant germline MMR / EPCAM deletions vs sporadic somatic MLH1 promoter hypermethylation, BRAF V600E exclusion rule, and cascade testing protocols.\n\n---\n\n## Verification Rubrics\n\nThe environment implements a 4-tier composite evaluation rubric:\n\n1. **Mechanistic Accuracy (`metric_mechanistic_accuracy`, Weight: 0.35):**\n   Evaluates whether the completion accurately identifies and details the biochemical, genetic, and structural mechanisms using canonical nomenclature.\n2. **Key Concepts Coverage (`metric_key_concepts_coverage`, Weight: 0.30):**\n   Verifies presence and coherent discussion of essential multi-step biological pathways and concepts.\n3. **Scientific Reasoning & Depth (`metric_scientific_reasoning_and_depth`, Weight: 0.20):**\n   Assesses causal reasoning markers, structural organization, and comprehensive mechanistic depth.\n4. **Biological Consistency (`metric_biological_consistency`, Weight: 0.15):**\n   Penalizes biological hallucinations, erroneous protein classifications (e.g. classifying TP53 or RB1 as oncogenes), or therapeutic contradictions.\n\n---\n\n## Quickstart & Verification\n\n### Local Installation\n```bash\ncd environments/human-cancer-genetics\npip install -e .\n```\n\n### Run Unit Tests\n```bash\npytest tests/ -v\n```\n\n### Load in Python\n```python\nfrom human_cancer_genetics import load_environment\n\nenv = load_environment()\nprint(f\"Loaded {type(env).__name__} with {len(env.dataset)} tasks.\")\n```\n\n### Run Evaluation with Prime CLI\n```bash\nprime eval run jackbull/human-cancer-genetics -m meta-llama/Llama-3.2-1B-Instruct\n```\n","encoding":"utf-8","truncated":false,"total_bytes":6423},"status":null}