Multi-Cohort
Cardiac Target
Discovery Engine
Unifying 12 public datasets: 6 single-cell cardiac atlases, 3 bulk RNA-seq cohorts, 3 mass-spectrometry cardiac proteomes (Malecki 2025 human LV, GTEx, Takasugi 2024 mouse), and Open Targets v4 human genetics with strict HAGR GenAge, EuropePMC, and mouse knockout causal direction auditing. All candidates are at computational identification; none has been tested in the lab.
Executive Summary
Traditional transcriptomic drug discovery in heart failure suffers from four systematic blind spots: (1) neglecting non-myocyte stromal, endothelial, pericyte, and myeloid niches; (2) assuming mRNA levels track protein abundance in aged post-mitotic myocardium; (3) conflating compensatory protective stress responses with pathogenic disease drivers; and (4) failing to audit candidate lists against mouse knockout phenotypes and human genetic colocalisation.
In our target discovery audit, Nilus Lab executed an unblinded multi-dataset discovery and independent causal audit across 12 public datasets (11 human, 1 mouse). Without prior gene weighting, the pipeline recovered the clinical heart failure benchmark ATP2A2 (SERCA2a) blind at rank 76 of 5,677 genes (3rd of the 11 genes left after all filters; $Z_{\text{DCM}} = -7.994$, Malecki human LV protein log2FC = −0.479, $p = 0.00039$, $q = 0.066$), alongside candidate TMEM65 (rank 74 of 5,677; $Z_{\text{DCM}} = -8.008$). All targets are at the computational identification stage; none has been tested in the lab.
Our subsequent audit (evaluating HAGR databases, historical literature aliases under MG29 and AAVS1, pre-registered SAP v3 GTEx 11-tissue cross-tissue testing, and IMPC knockout phenotypes) eliminated three compensatory protective stress brakes (ARHGAP1, NOL3, CIRBP), removed systemic non-cardiac ageing markers (ZMAT3 elevated across 9/10 non-cardiac tissues; LTBP2 in 6/10), removed COL18A1 and CBR4, and removed ZSWIM8 (not measurable in human heart proteomes; knockout mice have heart defects). Three computational candidates remain: SYPL2 (lower in DCM cardiomyocytes, Reichart $q = 2.0\times 10^{-5}$, MAGNet $q = 0.19$, protein measurable in 4 of 25 hearts; note its $H_4 = 0.973$ is a liver eQTL at a 44-gene locus, not heart evidence), PPP1R12C (Malecki human LV protein $+0.534$, $p = 1.1\times 10^{-4}$, $q = 0.037$ in the main test, $q = 0.12$ in two other tests; MAGNet DCM $q = 2.4\times 10^{-5}$; Reichart $q = 0.13$; first intron holds the AAVS1 locus; Bhagwan et al., F1000Research 2020), and TMEM65 (rank 74 of 5,677; $Z_{\text{DCM}} = -8.008$; Malecki protein $p = 0.86$, no change; mouse heart protein $-0.42$, $q = 0.00014$; cardiac knockdown in mice causes dilated cardiomyopathy; Teng et al., Nat Commun 2022). None of the four changes with age in GTEx left ventricle ($n = 431$).
Four-Angle Discovery Architecture
Executed Pipeline
pct_in_celltype ≥ 0.10 & (mean_celltype / mean_vCM) ≥ 0.50
# Angle 2: Dual-Species Cardiac Proteomics
Malecki_Human_LV (12 young vs 13 older, 25 hearts; p<0.05) ∩ Takasugi_Mouse (30 vs 6 months, p<0.05)
# Angle 3: DCM vs. Ageing Decomposition (5,677 Genes)
q_Reichart < 0.05 & q_MAGNet < 0.05 → 451 Concordant Genes (out of 5,677 genes)
# Angle 4: Open Targets v4 Genetics-First
GWAS_Coloc_H4 ≥ 0.80 & Tractability_Verified == True
Empirical Cohort Scale
Governing Statistical Formulations
$$Z_{\text{meta}} = \frac{\sum_{k=1}^{K} w_k \hat{\beta}_k}{\sqrt{\sum_{k=1}^{K} w_k}}, \quad w_k = \frac{1}{\text{SE}(\hat{\beta}_k)^2}$$
Eliminates ambient cardiomyocyte mRNA contamination across non-myocyte single-nucleus droplets inside raw
.h5ad count matrices.
Identifies proteins that accumulate or deplete in aged human LV ($p < 0.05$) and mouse heart ($p < 0.05$) despite flat mRNA ($|Z_{\text{RNA}}| < 1.0$).
Compensatory Brake Blacklist & Knockout Audit
Up-regulated genes in aged or failing myocardium frequently represent endogenous protective brakes rather than pathogenic drivers. Inhibiting a compensatory brake accelerates heart failure. Our audit verifies every candidate against HAGR GenAge, EuropePMC, and IMPC mouse knockout phenotypes before advancing any target. Every number traces to a named public dataset and an analysis script. The full audit trail is available to partners on request.
Candidate Targets & Clinical Benchmark
PPP1R12C (Malecki human LV protein +0.534, p = 1.1×10−4, q = 0.037 in the main test, q = 0.12 in two alternative tests; MAGNet DCM q = 2.4×10−5; Reichart q = 0.13). Its first intron holds the AAVS1 safe-harbour locus (Bhagwan et al., F1000Research 2020). Candidate at computational identification.
TMEM65 (rank 74 of 5,677 genes; $Z_{\text{DCM}} = -8.008$; Malecki protein p = 0.86, no change; mouse heart protein −0.42, q = 0.00014). Published: cardiac knockdown in mice causes dilated cardiomyopathy (Teng et al., Nat Commun 2022). Candidate at computational identification.
SYPL2 (lower in DCM cardiomyocytes, Reichart q = 2.0×10−5; MAGNet q = 0.19; protein measurable in only 4 of 25 hearts). Note: its H4 = 0.973 is a liver eQTL at a locus shared with 44 genes, not heart genetic evidence. Candidate at computational identification.
Positive control ATP2A2 (SERCA2a) was recovered blind (rank 76 of 5,677 genes; 3rd of 11 after all filters; $Z_{\text{DCM}} = -7.994$; Malecki protein −0.479, p = 0.00039, q = 0.066). None of the four targets changes with age in GTEx left ventricle (n = 431). Removed: ZSWIM8 (not measurable in human heart proteomes; knockout mice have heart defects), COL18A1, CBR4, ZMAT3, and LTBP2. None has been tested in the lab.