Nilus Lab Institutional Reference · Audited

Multi-Cohort Target Discovery
& Causal Audit Catalog

The definitive technical specification for the Nilus Lab 12-Dataset Cardiac Ageing & Heart Failure Discovery Pipeline . Documenting the 12 public datasets (11 human, 1 mouse), the 4 orthogonal discovery angles, the 4-gate causal directionality audit, the blind recovery of positive control ATP2A2 (SERCA2a), and the candidate targets (all at computational identification stage).

Public Datasets

12 Datasets

11 Human + 1 Mouse Datasets

Multi-Cohort Concordance

451 Genes

FDR q < 0.05 in Reichart SC & MAGNet Bulk

Positive Benchmark

SERCA2a Control

Blind Recovery (Rank 76 of 5,677; #3 of 11)

Audited Portfolio

3 Candidates + 1 Control

5 Genes Removed from Candidates

Section 01

12-Dataset Cardiac Data Registry

All target discovery and causal auditing is executed on 12 public cardiac datasets (11 human, 1 mouse). Zero synthetic cells or interpolated expression values are used.

Cohort / Study	Accession	Modality	Sample Size
SYPL2 Mitsugumin-29 (MG29) · Candidate	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.	DOWN: Reichart SC CM DCM (q=2.0e-5), MAGNet DCM q = 0.19. Low-abundance in bulk MS (measurable in 4 of 25 hearts in Malecki 2025; GTEx LV p = 0.25, n = 7). Note: H4 = 0.973 is a liver eQTL at a 44-gene locus, not heart evidence. No change with age in GTEx LV (n=431). Computational identification only.	RESTORE / OVEREXPRESS Cardiomyocyte-targeted gene replacement
PPP1R12C MBS85 Regulatory Subunit · Candidate	Targets PP1 catalytic subunit to dephosphorylate myosin light chain (MLC2v). Intron 1 holds the AAVS1 locus (Bhagwan et al., F1000Research 2020); requires unedited iPSC-CM lines for in vitro screens.	UP: MAGNet DCM q = 2.4e-5, Reichart DCM q = 0.13; Malecki 2025 human LV protein log2FC = +0.534 (p=1.1e-4; q=0.037 in main test, q=0.12 in two alternative tests). No change with age in GTEx LV (n=431). Computational identification only.	INHIBIT / KNOCKDOWN ASO/siRNA knockdown or PP1-docking disruptor
ZSWIM8 Removed from Candidates	E3 ubiquitin ligase adaptor. Not measurable in either human heart proteome (Malecki 2025 or GTEx LV). Zswim8 knockout mice have severe heart defects and die as newborns.	Human heart MS: not measurable (quantified in only 7 of 25 hearts in Malecki 2025; absent in GTEx LV). KO is lethal in mice with cardiac defects. Recommendation to inhibit withdrawn.	REMOVED · NOT MEASURABLE Permanently removed from candidate portfolio
ZMAT3 Removed · Systemic Ageing	p53-activated zinc-finger RNA-binding protein. Universal systemic marker.	GTEx 11-tissue cross-tissue model shows elevation in 9 of 10 non-cardiac tissues. Not cardiac-specific.	REMOVED · SYSTEMIC Universal p53 ageing marker, not a cardiac target
CBR4 Removed · Discordant	Mitochondrial fatty-acid synthase subunit.	Malecki human LV protein increases +0.269 (p = 0.0497), directly contradicting the restoration thesis.	REMOVED · DISCORDANT Human protein direction discordance invalidates restoration thesis

Section 02

Four Orthogonal Discovery Angles

Angle 1 Multi-Cohort DCM Concordance

451 Dual-Concordant Cardiomyocyte Genes

Differential expression was evaluated across Reichart 2022 single-cell LV and MAGNet bulk LV myocardium. Exactly 451 genes achieved Benjamini–Hochberg $\text{FDR} < 0.05$ in both cohorts with concordant disease direction, out of 5,677 genes evaluated.

Angle 2 Dual-Species Cardiac Proteomics

Malecki 2025 Human LV & Takasugi 2024 Mouse Heart

Intersects mass-spectrometry proteomics across 25 human left-ventricle hearts (Malecki et al. 2025: 12 donors aged 14–25 vs 13 aged 51–65; 3,431 proteins tested) and mouse heart aging proteomics (Takasugi et al. 2024, 30 vs 6 months). Recovers PPP1R12C (sole candidate passing FDR at $q = 0.037$ in main test) and TMEM65 (mouse heart protein $-0.42, q = 0.00014$).

Angle 3 DCM vs. Ageing Decomposition

Cardiomyocyte Failure & Positive Control Recovery

Cross-tabulates 451 FDR-concordant DCM genes against chronological ageing to separate candidates and unsteered blind recovery of positive control ATP2A2 (SERCA2a) at rank 76 of 5,677 genes ($Z_{\text{DCM}} = -7.994$, Malecki protein log2FC = -0.479, $p = 0.00039, q = 0.066$; 3rd of 11 genes left after all filters). None of the four targets changes with age in GTEx LV ($n = 431$).

Angle 4 Causal Genetics & Lineage Filtration

Human Genetic Colocalisation & Causal Safety

Integrates Open Targets v4 human genetics, literature alias archives, and knockout directionality audits. Systematic screening eliminated compensatory brakes (ARHGAP1, NOL3, CIRBP), removed systemic non-cardiac markers (ZMAT3, LTBP2, COL18A1), and removed ZSWIM8 (not measurable in human heart proteomes; knockout mice have heart defects).

Section 03

Four-Gate Causal Directionality & Safety Audit

Standard pipelines assume that every gene upregulated in heart failure is a disease driver that should be inhibited. In reality, failing cardiomyocytes upregulate critical protective compensatory brakes to survive stress. Inhibiting a protective brake causes catastrophic cardiac decompensation. Every candidate target must pass four mandatory gates:

GATE 1

Intrinsic Lineage Purity

Raw-count CM vs. non-CM verification across 592k nuclei to eliminate ambient RNA bleed-through.

GATE 2

Knockout Directionality

Cross-reference against published mammalian KO/overexpression phenotypes to separate drivers from protective brakes.

GATE 3

Proteomic & Tissue Concordance

Verification against human LV proteomes (Malecki 2025, GTEx) and mouse heart (Takasugi 2024).

GATE 4

Human Genetic & Pleiotropy

Screening against human LoF intolerance, arrhythmia loci, and multi-organ toxicity (holding pleiotropic receptors like LGR4).

Section 04

Candidate Targets & Evaluated Portfolio

Target Gene	Biological Mechanism	Multi-Cohort Human Evidence	Therapeutic Direction & Modality
SYPL2 Mitsugumin-29 (MG29) · Candidate	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.	DOWN: Reichart SC CM DCM (q=2.0e-5), MAGNet DCM q = 0.19. Low abundance in bulk MS (measurable in 4 of 25 hearts in Malecki 2025; GTEx LV p = 0.25, n = 7). Note: H4 = 0.973 is a liver eQTL at a 44-gene locus, not heart evidence. No change with age in GTEx LV (n=431). Computational identification only.	RESTORE Cardiomyocyte-targeted gene replacement
PPP1R12C MBS85 Regulatory Subunit · Candidate	Targets PP1 catalytic subunit to dephosphorylate myosin light chain (MLC2v). Intron 1 holds the AAVS1 locus (Bhagwan et al., F1000Research 2020); requires unedited iPSC-CM lines for in vitro screens.	UP: MAGNet DCM q = 2.4e-5, Reichart DCM q = 0.13; Malecki 2025 human LV protein log2FC = +0.534 (p=1.1e-4; q=0.037 in main test, q=0.12 in two alternative tests). No change with age in GTEx LV (n=431). Computational identification only.	INHIBIT / KNOCKDOWN ASO/siRNA knockdown or PP1-docking disruptor
ZSWIM8 Removed · Not Measurable	E3 ubiquitin ligase adaptor. Not measurable in either human heart proteome (Malecki 2025 or GTEx LV). Zswim8 knockout mice have severe heart defects and die as newborns.	Human heart MS: not measurable (quantified in only 7 of 25 hearts in Malecki 2025; absent in GTEx LV). Knockout is lethal in mice with cardiac defects. Recommendation to inhibit withdrawn.	REMOVED · NOT MEASURABLE Permanently removed from candidate portfolio
CBR4 Removed · Discordant	Mitochondrial fatty-acid synthase subunit.	Malecki human LV protein increases +0.269 (p = 0.0497), directly contradicting the restoration thesis.	REMOVED · DISCORDANT Human protein direction discordance invalidates restoration thesis
ZMAT3 Removed · Systemic Ageing	p53-activated zinc-finger RNA-binding protein. Universal systemic marker.	GTEx 11-tissue cross-tissue model shows elevation in 9 of 10 non-cardiac tissues. Not cardiac-specific.	REMOVED · SYSTEMIC Universal p53 ageing marker, not a cardiac target

Section 05

Positive Control Benchmark & Candidate Targets

Target Gene	Biological Mechanism & Literature Context	Multi-Cohort Human Evidence	Therapeutic Direction & Modality
ATP2A2 (SERCA2a) Positive Control Benchmark	Sarcoplasmic reticulum Ca2+ ATPase 2a; master regulator of diastolic Ca2+ reuptake and systolic SR Ca2+ load (CUPID benchmark). Completely unsteered blind recovery (zero script hardcoding).	Rank 76 of 5,677 genes in DCM (Z_DCM = -7.994, Reichart SC q = 0.023, MAGNet q = 2.5e-12; 3rd of 11 genes left after all filters); Malecki LV protein log2FC = -0.479 (p = 0.00039, q = 0.066). No change with age in GTEx LV (n=431).	POSITIVE CONTROL SERCA2a positive control confirming pipeline biological sensitivity
TMEM65 Candidate Target (Rank 74)	Stabilizes Connexin-43 (GJA1) at the intercalated disc and regulates mitochondrial Ca2+ homeostasis. Published: cardiac knockdown in mice causes dilated cardiomyopathy (Teng et al., Nat Commun 2022).	Rank 74 of 5,677 genes in DCM (Z_DCM = -8.008); mouse heart protein -0.42 (q = 0.00014); Malecki LV protein p = 0.86 (no change). No change with age in GTEx LV (n=431). Computational identification only.	RESTORE Cardiomyocyte-targeted gene replacement
COL18A1 Removed · No LV Ageing	Basement-membrane collagen cleaved into Endostatin.	Removed from portfolio: does not change with age in GTEx left ventricle (p = 0.991, n = 431). Malecki protein log2FC = +0.559 (p = 0.048); GTEx LV protein p = 0.027 (q_BH = 1.0).	REMOVED · NO LV AGEING Permanently removed from candidate portfolio
LTBP2 Removed · Systemic Fibrosis	Fibroblast extracellular matrix glycoprotein.	Removed from portfolio: general systemic fibrosis marker elevated in 6 of 10 non-cardiac tissues in GTEx 11-tissue model. Not cardiac-specific.	REMOVED · SYSTEMIC Permanently removed from candidate portfolio

Section 06

Rejected Protective Compensatory Brakes & Pleiotropy Holds

The following genes are strongly dysregulated across multiple datasets and would be nominated as top hits by naive differential expression pipelines, but were blocked by the Causal Safety Gate:

ARHGAP1 (p50RhoGAP) REJECTED DO NOT INHIBIT

Upregulated in DCM (Reichart single-cell and MAGNet bulk) and higher with age in human LV protein (Malecki +0.50, p = 0.011). However, ARHGAP1 inactivates pathological RhoA–ROCK signaling. In mice, most Arhgap1 knockouts died as newborns and survivors aged early (HAGR GenAge). Upregulation in human HF is a protective brake.

NOL3 (ARC) REJECTED DO NOT INHIBIT

Upregulated in all 3 snRNA atlases and reversed post-LVAD ($p=0.0029$). Encodes Apoptosis Repressor with CARD domain (ARC), which blocks Caspase-8 and BAX activation in post-mitotic cardiomyocytes. Nol3 knockout mice develop accelerated heart failure under biomechanical stress.

CIRBP REJECTED DO NOT INHIBIT

Upregulated in all 3 snRNA atlases and reversed post-LVAD ($p=0.00059$). Cold-inducible RNA-binding protein stabilizes KCND2/KCND3 ($I_{to}$ potassium channel) and ubiquinone biosynthesis transcripts. Cirbp knockout causes QT prolongation, arrhythmias, and bioenergetic defect.

LGR4 (GPR48) ON HOLD PLEIOTROPY

Downregulated across all 3 snRNA atlases, MAGNet, and GSE116250. However, LGR4 potentiates canonical Wnt/$\beta$-catenin signaling and regulates systemic bone, gut, and metabolic homeostasis. Held pending cardiomyocyte-restricted delivery validation.

Section 07

Governing Statistical Equations

1. Donor-Level Pseudobulk Aggregation

Prevents single-cell pseudoreplication by aggregating raw UMI counts $C_{c,g}$ across all nuclei $c \in \mathcal{C}_{d,k}$ for donor $d$ and cell type $k$:

$$E_{d,k,g} = \log_2 \left( \frac{\sum_{c \in \mathcal{C}_{d,k}} C_{c,g}}{\sum_{g'} \sum_{c \in \mathcal{C}_{d,k}} C_{c,g'}} \times 10^6 + 1 \right)$$

2. Ambient RNA Soup Purity Ratio

Eliminates non-myocyte extracellular matrix and endothelial transcript bleed-through before cardiomyocyte target ranking:

$$R_{\text{ambient}}(g) = \frac{\max_{k \neq \text{CM}} \bar{E}_{k,g}}{\bar{E}_{\text{CM},g}} \le 1.5$$

3. Multi-Cohort Meta-Analytic Score

Combines Benjamini–Hochberg adjusted significance across independent cohorts $m \in \{1,\dots,M\}$ conditioned on strict directional concordance:

$$S_{\text{meta}}(g) = \mathbb{I}(\text{sign concordant}) \cdot \sum_{m=1}^{M} -\log_{10}(\text{FDR}_{m,g})$$

4. Inverse-Variance Meta-Z Score

Combines standardized effect sizes across independent cardiac cohorts weighted by inverse standard error:

$$Z_{\text{meta}}(g) = \frac{\sum_{k=1}^K w_k Z_k(g)}{\sqrt{\sum_{k=1}^K w_k^2}}, \quad w_k = \frac{1}{\text{SE}_k^2}$$

Section 08

Scientific Integrity & Negative Result Disclosure

Why We Pivoted from Transcriptomic Age Clocks & Generative Cocktails to Multi-Cohort Causal Target Discovery

In earlier phases of the Zenith research program, we evaluated an ElasticNet transcriptomic biological age clock across healthy Human Cell Atlas donors ($N=18$) and tested generative transcription-factor cocktail heuristics. Rigorous Leave-One-Donor-Out (LODO) cross-validation and external evaluation on Koenig et al. 2022 ($N=25$ healthy donors) demonstrated that single-cohort transcriptomic age clocks achieve only modest donor-level correlation ($r = 0.46$, $\text{MAE} = 6.97\text{ yr}$ internally; $r = 0.256$ externally) due to technical batch variance and cell-composition confounding.

Rather than cherry-picking or concealing these limitations, Nilus Lab permanently retired synthetic perturbation trajectories, unverified cocktail claims (e.g., OSK/GMT-3/lncRNA heuristics), and decommissioned structural/LNP/robotics mock modules. All discovery and target validation on the platform is now anchored exclusively in the 12-dataset empirical discovery and 4-gate causal audit architecture.

Section 09

Verification Pipeline & REST API

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

Read Full Scientific Paper View Evidence Registry Scientific Q&A Developer API Reference