Methodology & Verification Q&A

In our internal audit of leave-one-donor-out (LODO) transcriptomic aging clocks across the 54-donor PERIHEART right-atrial cohort, a Ridge regression clock achieved MAE 6.97 years, $r = 0.4606$ (recorded in the model file; could not be re-derived from raw data). Moreover, a 10-tissue GTEx specificity audit showed that generic senescence markers (PTCHD4, EDA2R, LINC02388) track chronological age systemically across non-cardiac organs rather than specifically in myocardium.

We therefore engineered four complementary discovery angles:

• Angle 1 (Non-Myocyte Single-Cell Ageing): Donor-level sex-adjusted OLS age slopes meta-analyzed across Cardiac Fibroblasts, Endothelial cells, Macrophages, and Pericytes across 4 single-cell cohorts.
• Angle 2 (Dual-Species Cardiac Proteomics): Intersects Malecki 2025 human LV proteomics (12 young vs 13 older donors, 25 hearts) and Takasugi 2024 mouse heart proteomics (30 vs 6 months) with multi-cohort RNA slopes to capture both multi-modal concordant targets (Angle 2A) and post-transcriptionally decoupled proteins (Angle 2B).
• Angle 3 (DCM vs. Ageing Decomposition): Cross-tabulates 451 FDR-concordant DCM genes (Reichart 2022 snRNA LV + MAGNet bulk LV, out of 5,677 genes) against chronological ageing to separate 125 Synergy Drivers from 115 Exhausted/Antagonistic responses.
• Angle 4 (Human Genetics-First Colocalisation): Requires Open Targets v4 GWAS colocalisation posterior probability $H_4 \ge 0.80$ or locus-to-gene human genetic support plus druggable tractability.

In single-nucleus RNA-seq of human myocardium, lysed ventricular cardiomyocytes release abundant sarcomeric transcripts (e.g., TTN, RYR2, MYH6, TNNT2) into the suspension droplet buffer ("ambient soup"), which artificially contaminate fibroblast, endothelial, and macrophage pseudobulk regressions.

To eliminate ambient soup false positives in Angle 1, we queried the raw .h5ad count matrices directly and required every non-myocyte candidate $g$ in target lineage $L$ to satisfy two hard mathematical thresholds: $$\text{PctExpr}(g, L) \ge 10\% \quad \text{and} \quad \frac{\text{MeanExpr}(g, L)}{\text{MeanExpr}(g, \text{vCM})} \ge 0.5$$ This filtered out cardiomyocyte-leaked transcripts while retaining lineage-expressed candidates such as PPP1R12C (Mural/Pericyte). Crucially, subsequent cross-tissue auditing under our pre-registered SAP v3 GTEx 11-tissue model removed ZMAT3 (elevated in 9/10 non-cardiac tissues) and LTBP2 (elevated in 6/10 non-cardiac tissues) as universal systemic markers, and removed COL18A1 because it does not change with age in GTEx left ventricle (p = 0.991, n = 431).

A classic failure mode in transcriptomic and proteomic drug discovery is mistaking a compensatory protective stress response (which rises as the heart attempts to protect itself) for a pathogenic disease driver. If a gene is upregulated in aged or failing myocardium, inhibiting it is only valid if its upregulation is maladaptive.

Our audit verified every candidate against HAGR GenAge, EuropePMC, and IMPC mouse knockout phenotypes, catching three compensatory brakes:

• ARHGAP1 (p50RhoGAP — Rejected Control): Flagged in HAGR GenAge Human (#249) and GenAge Models (#13). Arhgap1^−/− mice exhibit 5-fold elevated Cdc42-GTP, severe premature aging, and reduced lifespan. Inhibiting ARHGAP1 would accelerate aging.
• NOL3 (ARC — Apoptosis Repressor with CARD Domain — Rejected): Although upregulated in aged human LV RNA ($Z = +3.94$) and protein ($p = 0.0044$), Nol3^−/− knockout mice develop accelerated dilated cardiomyopathy, cardiomyocyte apoptosis, and fibrosis under biomechanical load. NOL3 is a compensatory survival shield, not an inhibitor target.
• CIRBP (Cold-Inducible RNA-Binding Protein — Rejected): Although upregulated in aged human LV, Cirbp loss exacerbates myocardial ferroptosis and ubiquinone (CoQ10) deficiency.

Without any prior weighting toward calcium-handling genes, our unblinded screen recovered positive control 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, $p = 0.00039$, $q = 0.066$). None of the four targets changes with age in GTEx left ventricle ($n = 431$).

Even more importantly, the audit identified candidate targets that operate on the same excitation–contraction coupling apparatus:

• SYPL2 (MG29 — Candidate, Restore): In skeletal muscle, loss of MG29 disrupts T-tubule and triad structure (Nishi et al., J Cell Biol 1999); its role in heart is not established. Significantly lower in DCM cardiomyocytes (Reichart single-cell $q = 2.0\times 10^{-5}$; MAGNet bulk tissue $q = 0.19$). Its protein is measurable in only 4 of 25 human hearts in Malecki 2025. Crucially, our audit revealed that its $H_4 = 0.973$ colocalisation is a liver eQTL at a locus shared with 44 genes, not heart evidence. All candidates are at computational identification; none has been tested in the lab.
• PPP1R12C (MBS85 — Candidate, Inhibit): Just as clinical Inhibitor-1 (PPP1R1A) gene therapy inhibits Protein Phosphatase 1 (PP1) at the sarcoplasmic reticulum, PPP1R12C is the regulatory myosin-binding subunit scaffolding the active PP1 holoenzyme. Significant in MAGNet DCM ($q = 2.4\times 10^{-5}$; Reichart $q = 0.13$) and represents the sole candidate with FDR-significant validation in the Malecki 2025 Human LV Proteome (log₂FC = +0.534, $p = 1.1\times 10^{-4}$, $q = 0.037$ in the main test; $q = 0.12$ in two alternative tests). Importantly, the AAVS1 safe-harbour locus resides in its first intron (Bhagwan et al., F1000Research 2020), and in vitro validation requires unedited iPSC lines to avoid confounding. All candidates are at computational identification.

ZSWIM8 was initially evaluated as an E3 ubiquitin ligase adaptor in post-transcriptional screens. However, our audit found that ZSWIM8 protein is not measurable in either human heart proteome (quantifiable in only 7 of 25 hearts in Malecki 2025, and completely absent from GTEx LV).

Crucially, knockout directionality auditing revealed that Zswim8 knockout mice suffer from severe congenital heart defects and die as newborns. Because loss of ZSWIM8 is lethal and causes developmental heart failure, and its protein is not measurable in human adult heart tissue, all claims that its protein accumulates with age and all recommendations to inhibit ZSWIM8 have been permanently withdrawn.

TMEM65 is ranked 74 of 5,677 genes ($Z_{\text{DCM}} = -8.008$). In the Malecki 2025 human LV proteome, TMEM65 protein does not change with age ($p = 0.86$), while in mouse heart it falls (Takasugi 2024, $-0.42, q = 0.00014$). In published literature, cardiac-specific knockdown of Tmem65 in mice causes dilated cardiomyopathy (Teng et al., Nat Commun 2022). TMEM65 is an inner mitochondrial membrane and intercalated-disc protein that binds NCLX and stabilizes Connexin-43. None of the four targets changes with age in GTEx left ventricle (n = 431). TMEM65 is at the computational identification stage and has not been tested in the lab.

CBR4 (Carbonyl Reductase 4) was initially screened as an mtFAS II mitochondrial candidate ($Z_{\text{DCM}} = -7.230$). However, in the Malecki 2025 human left-ventricular proteome, CBR4 protein increases with age ($\log_2\text{FC} = +0.269, p = 0.0497$), directly contradicting the restoration hypothesis, while bulk cardiac ageing was null ($p = 0.127$). In accordance with our causal direction rules, CBR4 was disqualified and removed from the portfolio.

In Angle 4, LGR4 (GPR48) showed strong human genetic colocalisation ($H_4 = 0.91$, HFpEF GWAS), Open Targets cardiovascular genetic score 0.62, and downregulation in DCM ($Z_{\text{DCM}} = -7.57$). However, our literature and knockout audit revealed opposing cell-type-specific phenotypes: while cardiomyocyte LGR4 supports R-spondin/Wnt survival signaling, myeloid/macrophage-specific Lgr4 deletion reduces post-infarct AP-1 pro-inflammatory activation and improves healing. Because systemic agonism vs. antagonism carries opposing risks across cardiac lineages, we placed LGR4 on hold pending cell-type-restricted delivery validation.

Every number traces to a named public dataset and an analysis script. The full audit trail is available to partners on request.