4-Angle Discovery
& 4-Gate Audit

Step-by-step architecture for Nilus Lab's computational cardiac target discovery and audit pipeline. All targets are at the computational identification stage; none has been tested in the lab.

RESEARCH USE ONLY (RUO) / NOT CLINICALLY VALIDATED: All multi-cohort statistical regressions, proteomic concordance scores, and knockout directionality audits presented here are strictly for preclinical target prioritization and research evaluation.

01 / Four-Angle Discovery Engine

Angle 01

Non-Myocyte Single-Cell Ageing & Ambient Soup Filter

Computes donor-level sex-adjusted pseudobulk OLS age slopes across Cardiac Fibroblast, Endothelial, Macrophage, and Pericyte lineages across 3 single-cell .h5ad cohorts (PERIHEART and CAREBANK, both right atrium, and Kanemaru 2023). Applies a raw-count lineage specificity gate to limit cardiomyocyte ambient RNA contamination:

PctExpr(gene, Lineage) ≥ 10%  AND  MeanExpr(gene, Lineage) / MeanExpr(gene, vCM) ≥ 0.5 (to be raised to 1.0 in the next run)
→ ZMAT3 and LTBP2 were removed by the GTEx 11-tissue test as general ageing genes; COL18A1 was removed because it does not change with age in GTEx left ventricle (p = 0.99).
Angle 02

Dual-Species Cardiac Proteomics (Multi-Modal & Decoupled)

Intersects Malecki et al. 2025 human left-ventricle proteomics (12 donors aged 14–25 vs 13 aged 51–65; 3,431 proteins tested), GTEx human heart TMT proteomics (left ventricle $n=7$, atrial appendage $n=11$), and Takasugi et al. 2024 mouse heart ageing proteomics (30 vs 6 months) with human RNA age slopes:

• PPP1R12C: higher in older human LV (Malecki p=1.1×10−4; q=0.037 in the main test, q=0.12 in two alternative tests)
• CBR4 removed: its protein rises with age (+0.269), against the restore hypothesis
• ZSWIM8 removed: not measurable in either human heart proteome (4 of 12 young and 3 of 13 older hearts; absent from GTEx LV), and Zswim8 knockout mice have heart defects
Angle 03

DCM Heart Failure vs. Chronological Ageing Decomposition

Tests 5,677 cardiomyocyte genes in single-nucleus Reichart 2022 ventricular cardiomyocytes (52 DCM vs 18 healthy donors) and bulk MAGNet LV (DCM vs non-failing, $n=326$) to isolate 451 FDR-concordant DCM genes ($q < 0.05$ in both, same direction), then compares them with ageing and human LV proteomics:

• TMEM65: Z_DCM=−8.01 (rank 74 of 5,677); human LV protein unchanged with age (p=0.86)
• ATP2A2 / SERCA2a (positive control, found blind): Z_DCM=−7.99 (rank 76 of 5,677; 3rd of the 11 genes left after all filters); human LV protein lower with age (p=3.9×10−4)
• SYPL2 / MG29: lower in DCM cardiomyocytes (Reichart q=2.0×10−5), not significant in MAGNet tissue (q=0.19)
• None of the three changes with age in GTEx left ventricle (n=431)
Angle 04

Human Genetics-First Colocalisation (Open Targets v4)

Queries Open Targets v4 genetics and GWAS colocalisation tables ($H_4 \ge 0.80$) for human genetic support and tractability. The QTL tissue must be checked: SYPL2's colocalisation ($H_4 = 0.973$) is a liver eQTL at a locus shared with 44 genes, so it is not counted as heart evidence.

02 / Four-Gate Causal Direction Audit

Gate 01 / HAGR Databases GenAge & CellAge Check

Audits all candidates against HAGR GenAge Human (307 genes), GenAge Models (2,205 entries), and CellAge (948 genes). 0/15 discovery leads appear in any database (control ARHGAP1 flagged in GenAge #249/#13). Absence from these curated lists is weak evidence of novelty.

Gate 02 / Literature EuropePMC & Alias Audit

Queries live EuropePMC and historical literature aliases (e.g. MG29 for SYPL2, AAVS1 for PPP1R12C) to verify literature provenance and separate actionable leads from demoted systemic markers (ZMAT3, LTBP2, CBR4).

Gate 03 / Mouse Knockouts IMPC & KO Directionality

Removes genes whose rise in heart failure looks protective: dropped ARHGAP1, NOL3 (ARC), and CIRBP; placed LGR4 on hold due to macrophage vs. myocyte pleiotropy.

Gate 04 / Portfolio Audited Therapeutic Alignment

Current computational candidates: PPP1R12C, TMEM65 and SYPL2, with ATP2A2 (SERCA2a) as the positive control. Removed: ZMAT3, LTBP2, COL18A1, CBR4 and ZSWIM8. None has been tested in the lab.

03 / Reproducible Script Execution

Verification

Run Discovery & Audit Pipeline Locally

Every number in our target tables traces to a named public dataset and an analysis script. The scripts and the full audit trail are available to partners on request.