Physics-based 3D chromatin loop extrusion simulator for variant pathogenicity prediction
Paper · Quick Start · Results · Validation · Docker · Citation
LSSIM distribution across 12 variant categories (n = 1,103 HBB). LoF classes (nonsense, frameshift) cluster below 0.85; benign classes (intronic, synonymous) near 1.0. Red diamonds = 27 pearl variants.
| 63,153 variants analyzed, 13 loci |
AUC 0.977 HBB ROC performance |
27 pearls VEP-invisible HBB finds |
641 VUS pearl-like candidates |
ARCHCODE is a Discovery Engine, not a Prediction Tool.
ARCHCODE is an analytical mean-field loop extrusion simulator that discovers structural mechanisms of genomic variants. It builds wild-type and mutant 3D chromatin contact maps using Kramer-rate cohesin kinetics, then compares them via Structural Similarity Index (SSIM) to reveal how a variant disrupts local chromosome architecture.
Unlike sequence-based predictors (VEP, SpliceAI, CADD) that classify variants, ARCHCODE discovers variants that act through structural mechanisms invisible to sequence-level annotation:
- Enhancer-promoter loop disruption
- CTCF boundary erosion
- Cohesin loading site alteration
Key discoveries from applying ARCHCODE to 63,153 variants across 13 loci:
- 27 "pearl" variants on HBB — VEP-blind, SpliceAI-blind, CADD-ambiguous, but ARCHCODE-detected
- 641 pearl-like VUS candidates for reclassification across 8 tissue-matched loci
- Structural blind spot validated by 9 orthogonal methods (VEP, SpliceAI, CADD, MPRA, gnomAD, conservation, AlphaGenome RNA/ATAC, Hi-C, cross-species)
Not competing with ML predictors — ARCHCODE creates a new category: structural mechanism discovery.
Nine independent lines of evidence confirm the blind spot: VEP (<0.30 for all 27), SpliceAI (0.00 for all 20 SNVs), CADD v1.7 (median 15.7 — ambiguous), MPRA Kircher 2019 (no signal), MaveDB SGE/DMS (orthogonal, r ≈ 0), gnomAD v4 (85% absent), cross-species conservation (r = 0.82), genome-wide scaling (13 loci), yet ARCHCODE LSSIM < 0.92 for all 27.
+----------------------------------------------------------------------+
| ARCHCODE Pipeline v2.8 |
+----------------------------------------------------------------------+
| |
| ClinVar API --> 30,318 variants across 9 loci |
| | (HBB 1,103 + BRCA1 10,682 + CFTR 3,349 + |
| | TP53 2,794 + MLH1 4,060 + LDLR 3,284 + |
| | SCN5A 2,488 + TERT 2,089 + GJB2 469) |
| | |
| v |
| +--------------------+ +----------------------+ |
| | ARCHCODE Engine | | Ensembl VEP v113 | |
| | (Kramer kinetics) | | + SpliceAI plugin | |
| | | | + CADD v1.7 | |
| | WT contact map | | sequence predictors | |
| | MUT contact map | | | |
| | LSSIM comparison | | | |
| +--------+-----------+ +----------+-----------+ |
| | | |
| v v |
| +---------------------------------------------------+ |
| | Quadrant Analysis (HBB) | |
| | Q1: Both detect (199) Q2: ARCHCODE only (27) | |
| | Q3: VEP only (127) Q4: Neither (750) | |
| +---------------------------------------------------+ |
| | |
| v |
| ROC + Youden --> AUC = 0.977 (HBB) |
| Threshold: LSSIM < 0.994 |
| Sens = 0.966 | Spec = 0.988 |
| |
| Per-locus threshold calibration (9 loci) |
| Tissue-specificity gradient (matched --> mismatch) |
| |
+----------------------------------------------------------------------+
git clone https://github.com/sergeeey/ARCHCODE.git
cd ARCHCODE
npm install
npm run build
npx tsx scripts/generate-unified-atlas.ts # Process all variants
npm test # Run test suitedocker build -t archcode .
docker run -v $(pwd)/results:/app/results archcodeSee docker-compose.yml for persistent data volume configuration.
Analysis of 30,318 real ClinVar variants across 9 genomic loci using the unified ARCHCODE + Ensembl VEP v113 pipeline:
- ROC AUC = 0.977 on HBB — Youden optimum at LSSIM < 0.994 (Sensitivity 0.966, Specificity 0.988)
- 27 "pearl" variants — pathogenic by 3D chromatin structure but invisible to VEP, SpliceAI, and MPRA
- Tissue-specificity gradient — matched loci (HBB Δ = 0.111) → expressed (TERT Δ = 0.019) → mismatch (SCN5A/GJB2 Δ ≤ 0.006); defines ARCHCODE's domain of applicability
- Enhancer proximity drives discrimination — variants within ≤1 kb of H3K27ac peaks show 7× greater LSSIM separation (Δ = 0.039 vs genome-wide average Δ = 0.006); pearl variants cluster at median 831 bp from enhancers (Mann-Whitney p = 1.08 × 10−8 vs non-pearl pathogenic)
- Per-locus threshold calibration — optimal threshold varies from LSSIM < 0.977 (HBB, sens 92.9%, spec 99.1%) to LSSIM < 0.968 (TERT, sens 22.7%); SCN5A and GJB2 serve as deliberate negative controls (cell-type mismatch → no threshold works)
- Hi-C validation — Pearson r = 0.28–0.59 across 8 locus×cell-type combinations (all p < 10−82); best at MLH1 r = 0.59, HBB 95 kb r = 0.59, BRCA1 r = 0.53 (K562)
- CADD complementarity — pearl CADD median = 15.7 (ambiguous zone); 54 ARCHCODE-only variants vs 5,773 CADD-only confirms orthogonal detection modes
- Evolutionary conservation — pearl positions 3.3× more conserved than flanking background (phyloP 2.39 vs 0.73; permutation p = 0.0, n = 10,000); all fall within GERP constrained elements (scores 8.4–81.3)
353 real ClinVar HBB variants. Red = 27 pearl variants (VEP-blind, ARCHCODE-detected). Pearl zone: VEP < 0.30 AND LSSIM < 0.95.
HBB ROC. AUC = 0.977. Youden threshold LSSIM < 0.994.
Table: Top 5 Pearl Variants (of 27 total, all HBB)
| ClinVar ID | HGVS_c | Category | Significance | LSSIM | VEP | SpliceAI | Mechanism |
|---|---|---|---|---|---|---|---|
| VCV000869358 | c.50dup | frameshift | Pathogenic | 0.8915 | 0.15 | 0.00 | LoF, VEP misannotated |
| VCV002024192 | c.93-33_96delins... | splice_acceptor | Likely pathogenic | 0.9004 | 0.20 | 0.00 | Complex indel, VEP underscored |
| VCV000015471 | c.-78A>G | promoter | Pathogenic/LP | 0.9276 | 0.20 | 0.00 | Promoter-enhancer loop disruption |
| VCV000015470 | c.-78A>C | promoter | Pathogenic | 0.9276 | 0.20 | 0.00 | Promoter-enhancer loop disruption |
| VCV000036284 | c.-136C>T | promoter | Pathogenic/LP | 0.9277 | 0.20 | 0.00 | Promoter-enhancer loop disruption |
Sorted by LSSIM ascending (strongest structural disruption first). Full list: Supplementary Table S1.
ARCHCODE was applied to 9 clinically significant loci across 30,318 ClinVar variants to test generalizability beyond HBB:
| Locus | Disease | Chr | Variants | Pathogenic | Benign | Tissue match | ΔLSSIM | Pearls |
|---|---|---|---|---|---|---|---|---|
| HBB | β-thalassemia | 11 | 1,103 | 353 | 750 | Matched | 0.111 | 27 |
| BRCA1 | Breast/ovarian cancer | 17 | 10,682 | 7,062 | 3,620 | Partial | 0.006 | 0 |
| CFTR | Cystic fibrosis | 7 | 3,349 | 1,756 | 1,593 | Partial | 0.007 | 0 |
| TP53 | Li-Fraumeni syndrome | 17 | 2,794 | 1,645 | 1,149 | Partial | 0.009 | 0 |
| MLH1 | Lynch syndrome | 3 | 4,060 | 2,425 | 1,635 | Partial | 0.009 | 0 |
| LDLR | Familial hyperchol. | 19 | 3,284 | 2,274 | 1,010 | Partial | 0.002 | 0 |
| SCN5A | Brugada / Long QT | 3 | 2,488 | 928 | 1,560 | Mismatch | 0.003 | 0 |
| TERT | Telomerase / Cancer | 5 | 2,089 | 431 | 1,658 | Expressed | 0.019 | 0 |
| GJB2 | Hearing loss | 13 | 469 | 314 | 155 | Mismatch | 0.006 | 0 |
All loci use identical Kramer kinetics parameters (α = 0.92, γ = 0.80, kbase = 0.002; manually calibrated from literature ranges, Gerlich 2006, Davidson 2019). HBB shows the highest structural sensitivity, consistent with its well-characterized LCR enhancer-promoter architecture and compact 30 kb regulatory window. SCN5A (cardiac ion channel) and GJB2 (cochlear gap junction) serve as deliberate negative controls: K562 cell-type mismatch produces null discrimination, confirming that ARCHCODE's signal is biologically specific rather than a computational artifact.
Cross-locus validation summary. ΔLSSIM = separation between benign and pathogenic mean LSSIM. Tissue-specificity gradient from matched (HBB) through expressed (TERT) to mismatch (GJB2, SCN5A).
Per-locus threshold analysis across 9 genomic loci. LSSIM discrimination heatmap ordered by decreasing ΔLSSIM. Green = high discrimination (favorable); red = low. HBB achieves 92.9% sensitivity; mismatch loci show minimal signal.
ARCHCODE was benchmarked against ENCODE Hi-C data (K562, MCF7, HepG2, GM12878) and deep learning chromatin predictors:
Pearson r (ARCHCODE vs Hi-C) across 8 locus×cell-type combinations. All p < 10−82.
| Model | Type | Hi-C r (range) | Training data | Speed | Reference |
|---|---|---|---|---|---|
| ARCHCODE | Physics (Kramer LEF) | 0.28–0.59 | 0 (analytical) | < 1 s | This study |
| Akita | Deep learning CNN | 0.59 | ~4,000 Hi-C maps | ~145 s (GPU) | Fudenberg et al. 2020 |
| Orca | Graph neural network | 0.71 | Multi-scale Hi-C | N/A | Zhou et al. 2022 |
Best ARCHCODE loci: MLH1 r = 0.59, HBB 95 kb r = 0.59, BRCA1 r = 0.53 (K562). ARCHCODE achieves Akita-level correlation on top loci with zero training data, interpretable physics, and 100× speed advantage. Parameters (α, γ, kbase) map directly to measurable biophysical quantities (cohesin residence time, processivity, loading rate).
Pearl variants (n = 27, all HBB) were evaluated against five independent predictors:
| Predictor | Pearl score | Detection? | Mechanism tested |
|---|---|---|---|
| VEP/SIFT | < 0.30 (all 20) | No | Protein sequence + canonical splice |
| SpliceAI | 0.00 (all 20 SNVs) | No | Deep-learning splice disruption |
| CADD v1.7 | median 15.7 | Ambiguous | Sequence conservation + annotations |
| MPRA (Kircher 2019) | mean −0.015 | No (p = 0.91) | Promoter-intrinsic transcription |
| ARCHCODE LSSIM | < 0.92 (all 27) | Yes | 3D enhancer-promoter contact |
Only ARCHCODE detects these variants, confirming they operate through enhancer-promoter contact disruption — a structural mechanism invisible to both rule-based (VEP) and neural-network-based (SpliceAI) sequence predictors, as well as episomal functional assays (MPRA).
SpliceAI scores were obtained via the Ensembl VEP REST API with SpliceAI plugin for all 20 pearl single-nucleotide variants. Every variant scored 0.00 across all four splice metrics (donor gain, donor loss, acceptor gain, acceptor loss). This extends the structural blind spot beyond VEP consequence annotation: pearl variants are invisible not only to rule-based classifiers but also to the highest-resolution deep-learning splice predictor currently available.
Cross-validation against the Kircher et al. 2019 MPRA dataset (MaveDB: urn:mavedb:00000018-a-1; 623 variants across the HBB promoter region, assayed in HEL 92.1.7 erythroid cells) yields null correlation between ARCHCODE LSSIM and MPRA functional scores (r = −0.21, p = 0.36; n = 22 allele-matched variants). This null is mechanistically expected: MPRA measures promoter-intrinsic transcriptional activity in an episomal context, stripped of the 3D chromatin architecture through which pearl variants are predicted to act.
AlphaGenome RNA-seq and ATAC-seq tracks were used to assess whether ARCHCODE pearl variants produce independent regulatory signals at 1 bp resolution:
| Modality | Metric | Pearl (n = 23) | Benign (n = 23) | p-value |
|---|---|---|---|---|
| RNA-seq | Signal concentration ratio | 16.97 | 6.09 | 4.8 × 10−5 |
| RNA-seq | Δ at variant bin | 0.381 | 0.109 | 0.0014 |
| ATAC-seq | Signal concentration ratio | 11.15 | 6.39 | 0.0026 |
| ATAC-seq | Δ at variant bin | 0.268 | 0.098 | 0.029 |
Pearl variants show 2.8× higher RNA-seq signal concentration and 2.7× higher ATAC-seq Δ at the variant position compared to benign controls (all p < 0.05). The three-locus tissue gradient (HBB: 10/10 significant tests; BRCA1: 1/10; SCN5A: 0/10) mirrors ARCHCODE's tissue-specificity gradient, indicating biological specificity rather than simulation artifact.
Caveat: AlphaGenome's training data includes 4DN Hi-C from K562; partial overlap with our validation cell line cannot be excluded. The convergence of two mechanistically orthogonal approaches — analytical loop extrusion (ARCHCODE) and sequence-to-epigenome deep learning (AlphaGenome) — nonetheless strengthens the case for genuine regulatory disruption.
Ablation of effect strength encoding. Removing the effect-strength term from LSSIM degrades HBB discrimination; the full model recovers the biologically expected category ordering.
Enhancer proximity drives ARCHCODE structural discrimination. Variants within ≤1 kb of H3K27ac peaks show 7× greater LSSIM separation (Δ = 0.039) than the genome-wide average (Δ = 0.006). Pearl variants cluster at median 831 bp from enhancers (Mann-Whitney p = 1.08 × 10−8 vs non-pearl pathogenic).
| Layer | Technology | Purpose |
|---|---|---|
| Simulation engine | TypeScript 5.2 | Kramer-rate loop extrusion, contact matrices, LSSIM |
| 3D visualization | React 18 + Three.js (r181) | Interactive chromatin fiber viewer |
| State management | Zustand | Reactive simulation parameters |
| Styling | Tailwind CSS 4 | Responsive UI components |
| Data pipeline | Python 3.11 + matplotlib | ROC analysis, VEP integration, figure generation |
| Testing | Vitest | Physics regression tests, gold-standard validation |
| Build | Vite 5 | Fast HMR, optimized production builds |
| Containerization | Docker | Reproducible scientific environment |
| Dataset | Source | Version / Access |
|---|---|---|
| ClinVar variants | NCBI E-utilities (esearch + efetch) | Retrieved 2025-2026 |
| Hi-C contact maps | 4DN Data Portal (K562, MCF7, HepG2, GM12878) | ENCODE accessions per locus |
| MPRA functional scores | MaveDB | urn:mavedb:00000018-a-1 (Kircher et al. 2019, Nat Commun 10:3583) |
| SpliceAI scores | Ensembl VEP REST API + SpliceAI plugin | VEP v113 |
| CADD scores | Ensembl VEP REST API | CADD v1.7 |
| Evolutionary conservation | UCSC phyloP100way (hg38) | GERP constrained elements (Ensembl) |
| AlphaGenome tracks | AlphaGenome SDK v0.6.0 | RNA-seq, ATAC-seq, CTCF ChIP-seq |
Project Structure
ARCHCODE/
+-- manuscript/ # Publication (arXiv q-bio.GN pending)
| +-- body_content.typ # Main manuscript body (Typst)
| +-- main_ru.typ # Russian-language version
| +-- TABLE_S1_PEARLS.md # All 27 pearl variants (supplementary)
+-- figures/ # Publication figures (PDF + PNG)
| +-- fig1_ssim_violin.* # LSSIM distribution by category
| +-- fig2_roc_curves.* # ROC curve (HBB AUC 0.977)
| +-- fig3_pearl_quadrant.* # Pearl quadrant (ARCHCODE vs VEP)
| +-- fig4_hic_validation.* # Hi-C correlation heatmap
| +-- fig5_multilocus_summary.* # 9-locus bar chart
| +-- fig6_contact_maps.* # WT vs MUT contact map panels
| +-- fig7_ablation_barplot.* # Ablation analysis
| +-- fig8_enhancer_proximity.* # Enhancer proximity discrimination
| +-- fig9_tissue_heatmap.* # Tissue-specificity heatmap
| +-- fig10_alphagenome_validation.* # AlphaGenome multimodal validation
+-- results/
| +-- HBB_Unified_Atlas.csv # 1,103 HBB variants (unified pipeline)
| +-- integrative_benchmark.csv # 30,318 variant CADD concordance
| +-- integrative_benchmark_summary.json
| +-- per_locus_thresholds_summary.json
| +-- spliceai_pearl_variants.csv # SpliceAI scores for 20 pearl SNVs
| +-- mpra_crossvalidation_summary.json
| +-- conservation_pearl_analysis.json
| +-- conservation_robustness.json
| +-- roc_unified.json # ROC analysis (AUC 0.977)
| +-- figures/ # Legacy figures (v1-v2)
+-- scripts/
| +-- generate_publication_figures.py # All 10 publication figures
| +-- generate-unified-atlas.ts # Main variant processing pipeline
| +-- calculate_roc_and_quadrants.py # ROC + quadrant analysis
| +-- run_vep_predictions.py # Ensembl VEP batch predictions
+-- src/
| +-- engines/ # Physics engines
| | +-- LoopExtrusionEngine.ts # Core Kramer-rate simulator
| | +-- MultiCohesinEngine.ts # Multi-cohesin extension
| | +-- contactMatrix.ts # Contact map generation
| +-- domain/ # Biophysical constants and models
| +-- components/ # React + Three.js UI
| +-- __tests__/ # Test suite (regression)
+-- config/ # Simulation parameters (per-locus JSON)
+-- Dockerfile # Reproducible container
+-- docker-compose.yml # Volume-mounted configuration
+-- CLAUDE.md # Scientific Integrity Protocol
- Mean-field approximation — analytical Kramer-rate model, not full stochastic Monte Carlo simulation
- Cell-type specificity — all simulations use K562 CTCF/H3K27ac annotations; loci mismatched to K562 expression (SCN5A cardiac, GJB2 cochlear) produce null signal by design
- Hi-C validation — ENCODE Hi-C r = 0.28–0.59 across loci (significant, p < 10−82); pilot HUDEP2 Capture Hi-C r = 0.16 (not significant, small sample)
- Parameters manually calibrated — α = 0.92, γ = 0.80 from literature ranges (Gerlich 2006, Davidson 2019), not fitted to data
- No missense sensitivity — ARCHCODE models chromatin topology, not protein folding; missense variants are detected only indirectly via CADD-derived effect strength
- HBB-centric pearls — all 27 pearl variants are on HBB; generalization of the pearl-detection paradigm to other loci requires matched regulatory annotations
- MPRA episomal context — the Kircher 2019 null (r = −0.21) is mechanistically expected but does not rule out alternative non-structural explanations for pearl pathogenicity
- AlphaGenome training overlap — AlphaGenome was trained on 4DN Hi-C including K562; validation against AlphaGenome tracks is not fully independent
- Pearl candidates require experimental validation — the 27 pearl variants and 641 pearl-like VUS candidates are computational findings only; no clinical reclassification without RT-PCR, Capture Hi-C, or functional assay
- AlphaGenome as auxiliary only — AlphaGenome RNA/ATAC signal supports structural blind spot narrative but is not a replacement for experimental validation (same training domain; use for prioritization only)
This project follows a strict Scientific Integrity Protocol governing all AI-assisted development:
- No phantom references (every DOI verified before use)
- No invisible synthetic data (all mock data watermarked)
- No post-hoc claims as pre-registered
- Transparent parameter provenance (MEASURED / CALIBRATED / ASSUMED)
This protocol was developed after a self-audit identified risks of AI-generated hallucinations in scientific code. See CLAUDE.md for the full protocol.
For Codex-assisted tasks, use:
- AGENTS.md for plan-first execution and approval gate
- Codex Zero-Hallucination Gates
- Implemented vs Verified report template
Manuscript prepared for submission to arXiv (q-bio.GN). Preprint pending — citation will be updated upon assignment.
Boyko, S.V. (2026). ARCHCODE: 3D Chromatin Loop Extrusion Simulation Reveals Structural Pathogenicity Invisible to Sequence-Based Predictors — Evidence from 30,318 ClinVar Variants across Nine Genomic Loci. arXiv q-bio.GN (preprint pending).
@article{boyko2026archcode,
title = {ARCHCODE: 3D Chromatin Loop Extrusion Simulation Reveals Structural
Pathogenicity Invisible to Sequence-Based Predictors ---
Evidence from 30,318 ClinVar Variants across Nine Genomic Loci},
author = {Boyko, Sergey V.},
year = {2026},
note = {arXiv preprint q-bio.GN (pending)},
url = {https://github.com/sergeeey/ARCHCODE}
}MIT License — See LICENSE
ARCHCODE v2.8 · Updated 2026-03-04 · Sergey V. Boyko · sergeikuch80@gmail.com