Heart failure drugs mostly lower blood pressure, block stress hormones, or remove fluid. They ease the load on the heart, but they do not repair the failing heart muscle cells. A recent single-nucleus atlas of 61 failing and 18 control human hearts found that different disease-causing mutations damage the heart through different cell-level programs, not one shared "final common pathway" (Pathogenic variants damage cell composition and single cell transcription in cardiomyopathies). If that holds, treatment has to target the program a given genotype actually drives. That atlas showed genotype is predictable from cell-type expression using a graph neural network, but it did not produce a short, ranked list of genes or gene combinations that a biologist could test.
We propose to extract those lists with random forests, ensembles of decision trees whose split choices can be read off as feature importances. We also fix the flaw that makes naive forests unreliable on atlas data. Nuclei from the same heart are not independent samples. A forest trained on nuclei will happily learn "which donor is this" and report donor-specific genes as genotype markers. Our method resamples and evaluates at the level of whole hearts. It uses iterative random forests to find gene combinations, not just single genes, and it scores each finding by how often it survives when different hearts are left out. The claim fits in two sentences. Within each cardiac cell type, genotype is encoded partly by non-additive gene combinations that are stable across donors. A donor-grouped interaction search recovers these combinations and predicts the genotype of unseen hearts better than additive per-donor signatures.
Where this sits
The method comes from iterative Random Forests to discover predictive and stable high-order interactions. That paper reweights features by importance over repeated forest fits, mines frequently co-occurring features along decision paths with random intersection trees, and wraps everything in an outer bootstrap that assigns stability scores. It was validated on fly enhancers and human splicing, where samples are roughly exchangeable. Its outer bootstrap draws individual samples, which is exactly wrong when thousands of samples share a donor.
The closest cardiac use is Epistasis regulates genetic control of cardiac hypertrophy. It developed lo-siRF on 29,661 UK Biobank MRI-derived left ventricular mass phenotypes and about 15 million variants, prioritized interacting loci including CCDC141, IGF1R, TTN and TNKS, and confirmed an 8.5% non-additive cell-size reduction in stem-cell-derived cardiomyocytes. That work finds germline interactions behind a single static trait in one-row-per-person data. We look for expression interactions inside cell types, conditional on a known pathogenic genotype, in data with a nested nucleus-in-donor structure. The epistasis paper motivates the bet that cardiac biology is non-additive. Its loci also give us an external check: genes it nominates should be enriched among our cardiomyocyte findings if the two lines of evidence agree.
CellBRF supplies a second ingredient, class balancing before forest training so rare cell states are not drowned out. CellBRF uses this for unsupervised clustering on non-cardiac atlases with self-generated labels. We use balancing for a supervised genotype question and apply it per donor, so that rare states like the profibrotic vFB2 fibroblasts are represented without letting one donor's nuclei dominate.
Other cardiac random forest work constrains the design more than it supplies it. Performance discrepancy mitigation in heart disease prediction for multisensory inter-datasets and A Wide and Deep Transformer Neural Network for 12-Lead ECG Classification use forests as one-shot filters on low-dimensional clinical or ECG features. Neither examines interactions or stability. Feature selection and transformation by machine learning reduce variable numbers and improve prediction for heart failure readmission or death is a cautionary result. On administrative data, principal components beat interpretable selected subsets (AUC 0.66 vs 0.62). Interpretability can therefore cost accuracy, and we must measure that cost rather than assume it away.
Research Program
The unit of analysis is one cell type at a time: cardiomyocytes, fibroblasts including vFB2, endothelial cells, endocardium, and immune cells. The label is the donor's genotype class (for example TTN, LMNA, RBM20, PKP2, genotype-negative, or control, using the groupings available in the atlas). Features are normalized gene expression of that cell type's nuclei.
The method changes iterative random forests in three places. First, every bootstrap, including the outer stability loop and the inner tree bootstraps, samples donors, then subsamples a fixed number of nuclei per donor after per-donor class balancing. Second, feature importance used for reweighting is permutation importance computed on held-out donors. We do not use impurity reduction, the within-tree measure of how much a split purifies its node, because it is biased toward high-variance continuous features and is computed on in-sample nuclei. Third, an interaction counts as stable only if it is recovered in a large fraction of donor-level resamples. Its predictive value is then tested by adding it as an explicit product feature to a donor-level model and scoring on held-out hearts.
To calibrate false discoveries, the whole pipeline is rerun with genotype labels permuted across donors. This gives a null distribution of stability scores for each cell type. A finding is reported only if its stability exceeds the 95th percentile of that null.
Two follow-ups connect findings back to therapy. We test whether stable cardiomyocyte interactions are enriched for regulators the file identifies as disease brakes or checkpoints, such as LAPTM5, USP20 and its target STAT3, CIP, and the epistasis loci. We also ask whether the selected genes change across the compensated, HFpEF-like and HFrEF-like stages of the transverse aortic constriction mouse series (GSE182985 for mRNA, GSE182912 for circRNA, from Differential mRNA Expression and Circular RNA-Based Competitive Endogenous RNA Networks in the Three Stages of Heart Failure in Transverse Aortic Constriction Mice). The mouse data have too few animals for discovery, so they serve only as a stage-relevance check.
Baselines
The naive reference is a majority-genotype predictor scored on held-out donors, together with a standard nucleus-level random forest with impurity importance. We expect the standard forest to look excellent under random nucleus splits and to collapse under donor splits. Showing that gap is itself a result.
The strongest additive competitor is per-donor pseudobulk. Expression is averaged per cell type per heart, then fed to an L1-penalized multinomial logistic regression (the lasso, which keeps a small set of genes), with genes preselected by edgeR differential expression as in the atlas paper. It respects donor structure and is what a careful bioinformatician would do. It gets the same donors, the same leave-donors-out folds, and the same tuning budget, with penalty strength chosen by nested cross-validation.
The strongest existing nonlinear model is the atlas's graph attention network. For it, no reported result found on Lacuna for held-out-donor genotype accuracy; the file states that genotype prediction succeeded but gives no number. We will retrain it on our folds.
For rare-state handling, CellBRF is the reference. It reports average Adjusted Rand Index of 0.66 on small and 0.57 on large benchmark datasets, beating FEAST and DUBStepR by 0.12 on average (CellBRF). We will compare vFB2 recovery using genes selected by CellBRF against our selected genes.
Ablations isolate each ingredient. We swap donor bootstrap for nucleus bootstrap, held-out-donor permutation importance for impurity importance, and per-donor balancing for none. We also drop interaction features, keeping only marginal selected genes.
The method wins if it beats pseudobulk lasso on held-out-donor balanced accuracy and macro-AUC in at least the cardiomyocyte and fibroblast compartments, with paired bootstrap confidence intervals over donors excluding zero. It should also use a gene budget no larger than the lasso's and come within noise of the retrained graph network.
Why not just…?
Why not just run pseudobulk differential expression and a lasso? This is the real threat. If genotype signals are additive within cell types, it will match us at a fraction of the cost. The no-interaction ablation settles the question directly. If interaction features add nothing on held-out donors beyond marginal genes, we will report that genotype programs are additive at this resolution. That is a useful negative, because it would mean simpler signatures suffice for patient stratification. Our edge can only come from non-additivity, which the epistasis paper found at the germline level but nobody has tested within cell types.
Why not just use standard random forest or permutation importance on nuclei? Thousands of correlated nuclei per heart let trees memorize donor identity, so importance will reward donor-specific genes. The nucleus-bootstrap ablation, scored under donor-held-out evaluation, will show how much apparent signal is leakage. We expect this to be large, because every rare state is concentrated in a few hearts.
Why not just use the atlas's graph attention network and read its attention weights? Attention weights are not a stable ranked gene list, and they give no false-discovery control. The comparison here is whether a parsimonious, null-calibrated set of genes and gene pairs keeps most of the network's held-out accuracy. If the network is much more accurate, we will say the interpretable view loses real signal, mirroring the PCA result in the readmission study.
Decisive Checks
Three outcomes decide the project. The first is whether stable interactions in each cell type exceed the label-permutation null at all. The second is whether adding them as product features lowers held-out-donor error relative to the additive pseudobulk lasso. The third is whether the selected genes are enriched for the independently nominated regulators and epistasis loci, and whether they change in the staged mouse data. A pass on all three yields concrete, genotype-specific gene pairs for knockdown in stem-cell-derived cardiomyocytes, the same validation route the epistasis paper used. A failure on the second check says the program is additive and that pseudobulk signatures are the right tool for stratifying patients.
References
- Pathogenic variants damage cell composition and single cell transcription in cardiomyopathies
- Iterative random forests to discover predictive and stable high-order interactions
- Epistasis regulates genetic control of cardiac hypertrophy
- CellBRF: a feature selection method for single-cell clustering using cell balance and random forest
- Performance discrepancy mitigation in heart disease prediction for multisensory inter-datasets
- A Wide and Deep Transformer Neural Network for 12-Lead ECG Classification
- Feature selection and transformation by machine learning reduce variable numbers and improve prediction for heart failure readmission or death
- Differential mRNA Expression and Circular RNA-Based Competitive Endogenous RNA Networks in the Three Stages of Heart Failure in Transverse Aortic Constriction Mice