evo2-e2e: Production Metastasis Interception Pipeline

Brain Metastasis → Breast Metastasis → Any Cancer

Production implementation of the Target-Lock / Assassin Score framework from the Intercepting metastasis: 8-step CRISPR design via multi-modal foundation models manuscript.

Architecture

evo2-e2e/
├── core/                      # Modality-agnostic scoring engines
│   ├── target_lock/           # Target-Lock scorer (Evo2 + Enformer)
│   ├── assassin_score/        # Assassin composite ranking
│   ├── evo2_client/           # Modal Evo2 service (crispro-evo2-v9, A100)
│   ├── enformer_client/       # Modal Enformer service
│   └── acmg/                  # ACMG/AMP VUS classifier (anti-hallucination)
│
├── brain_met/                 # Brain metastasis module
│   ├── targets/universe.py    # 19 positives + 9 hard negatives, 7 BrM steps
│   ├── guides/                # CRISPR guide design for BrM targets
│   ├── structural/            # AlphaFold3 structural validation
│   └── pipeline/run_brm_pipeline.py  # End-to-end pipeline
│
├── breast_met/                # Breast metastasis (extend universe.py pattern)
│   ├── targets/               # AURORA + TCGA-BRCA gene universe
│   └── guides/
│
├── shared/
│   ├── gene_coords/           # hg38 TSS coordinates (Ensembl validated)
│   ├── ground_truth/          # metastasis_rules_v1.0.1.json
│   └── calibration/           # Gene-specific Evo2 calibration snapshots
│
├── tests/
│   ├── stress/test_antihallucination.py  # 17-test suite, ClinVar ground truth
│   ├── integration/           # End-to-end pipeline tests
│   └── unit/                  # Component tests
│
├── scripts/
│   └── reproduce_all.sh       # One-command reproduction
│
└── data/
    ├── brain_met/             # Real data: GSE237446, GSE205033, MSK-MET, AURORA
    └── validation/            # live_variant_scores.json (Evo2 A100 scores)

Quick Start

# Install dependencies
pip install -r requirements.txt

# Run full BrM pipeline (fast mode: no Enformer, 2 Evo2 windows)
python brain_met/pipeline/run_brm_pipeline.py --fast

# Run full pipeline with Enformer
python brain_met/pipeline/run_brm_pipeline.py

# Run stress tests (anti-hallucination)
python tests/stress/test_antihallucination.py --fast

# Score specific step
python brain_met/pipeline/run_brm_pipeline.py --step bbb_transit

Deployed Services (Modal, crispro-test workspace)

Service	App	GPU	Status
Evo2 1B base	crispro-evo2-v9	A100	✅ LIVE
Enformer	crispro-enformer	T4	✅ LIVE
BrM ICL adapter	brm-icl-v1	CPU	✅ LIVE

Live Results (2026-03-28)

BrM Clinical Variants — Evo2 Conditional LL (A100, hg38, 8192bp):
  PIK3CA p.H1047R   delta_ll = -0.615  (top penalized — PI3K kinase domain)
  TP53   p.R175H    delta_ll = -0.418  (BrM hotspot, 2× enriched)
  ESR1   p.D538G    delta_ll = -0.402  (ligand-independent ER activation)
  BACE1  p.D289N    delta_ll = +0.002  (near neutral — ACMG PVS1 context needed)

What Was Fixed vs the POC

POC Issue	Fix
Target-Lock saturation (0.352–0.355 band)	Step-z-score normalization in scorer.py
Chromatin 15% → −0.013 AUROC	Downweighted to 10%; regulatory increased to 20%
Off-target: no bulge/CFD weighting	anti_hallucination_flags documents gap; CFD TODO
15 guides only in structural validation	Expanded framework: top-5/step reproducible script
IDH1/IDH2 used as glioma negatives	Hard negatives added: IDH1, IDH2, FLT3, ABL1, etc.
PM2 always fires without gnomAD	Fixed in ACMG classifier (requires real gnomAD data)
All Evo2 paper tests failed	New score_variant_conditional endpoint (conditional LL)

Scoring Methods

Target-Lock Formula

Target_Lock = 0.35×Functionality + 0.35×Essentiality + 0.20×Regulatory + 0.10×Chromatin

Brain-Met variant (WEIGHTS_BRAIN_MET):
  0.33×Functionality + 0.33×Essentiality + 0.24×Regulatory + 0.10×Chromatin
  (Higher regulatory: BBB transit involves splice isoforms)

Assassin Score Formula

Assassin = 0.37×Efficacy + 0.30×Safety + 0.30×Mission + 0.03×Structure

Scoring Method

• Functionality: 1 / (1 + exp(delta_ll / 0.5)) — conditional LL from Evo2
• Essentiality: 1.0 for frameshift/nonsense; min(1.0, abs(delta_ll)/1.5) for missense
• Regulatory: |min_delta| / (|min_delta| + 1) — splice/noncoding impact
• Chromatin: Enformer accessibility proxy [0,1]

Brain Met Target Universe

7 BrM-specific steps (vs 8 general metastasis steps in POC):

1. primary_tumor_escape — escape from breast/lung/melanoma primary
2. intravasation — blood/lymph entry
3. circulation_survival — CTC anoikis resistance
4. bbb_transit ← NEW — blood-brain barrier crossing
5. cns_colonization — brain micrometastasis establishment
6. brain_niche_adaptation — astrocyte/neuron microenvironment adaptation
7. brm_angiogenesis — brain-specific angiogenic switch

19 positive targets (evidence from AURORA, MSK-MET, GSE237446, GSE205033, literature) 9 hard negatives (IDH1, IDH2 = glioma/NOT BrM; ABL1, FLT3 = hematologic; etc.)

Reproducibility

• Seed: 42 (matches manuscript)
• Model: crispro-evo2-v9 (A100, evo2_1b_base, conditional LL scoring)
• All variants scored against real hg38 sequences from UCSC API
• Results timestamped and saved to data/brain_met/pipeline_results_{ts}.json
• Anti-hallucination flags in every result object

Research Use Only

This framework is for research purposes only and has not been validated for clinical use. All predictions require experimental validation before therapeutic application.