Abstract
Although genetic variant effects often interact non-additively, strategies to uncover epistasis remain in their infancy. Here, we develop low-signal signed iterative random forests to elucidate the complex genetic architecture of cardiac hypertrophy, using deep learning-derived left ventricular mass estimates from 29,661 UK Biobank cardiac MRIs. We report epistatic variants near CCDC141, IGF1R, TTN, and TNKS, identifying loci deemed insignificant in genome-wide association studies. Functional genomic and integrative enrichment analyses reveal that genes mapped from these loci share biological process gene ontologies and myogenic regulatory factors. Transcriptomic network analyses using 313 human hearts demonstrate strong co-expression correlations among these genes in healthy hearts, with significantly reduced connectivity in failing hearts. To assess causality, RNA silencing in human induced pluripotent stem cell-derived cardiomyocytes, combined with novel microfluidic single-cell morphology analysis, confirms that cardiomyocyte hypertrophy is non-additively modifiable by interactions between CCDC141, TTN, and IGF1R. Our results expand the scope of cardiac genetic regulation to epistasis.
Introduction
Cardiac hypertrophy, the thickening of the left ventricular wall, is a primary risk factor for heart failure and sudden cardiac death. While genome-wide association studies (GWAS) have successfully identified numerous risk loci associated with heart structure, these models typically operate under the assumption of additivity—where each genetic variant contributes independently to the trait. However, biological systems are inherently interactive. This phenomenon, known as epistasis, occurs when the effect of one gene is dependent on the presence of another.
Uncovering these interactions has historically been hampered by the "combinatorial explosion" of testing millions of variant pairs and the inherently weak signals of individual genetic components. This study introduces a novel machine learning framework designed to move beyond additive models and map the epistatic landscape of cardiac hypertrophy.
The lo-siRF Framework
To navigate the complexity of gene-gene interactions, the researchers developed lo-siRF (low-signal signed iterative random forests). Built upon the Predictability, Computability, and Stability (PCS) framework, lo-siRF offers several key innovations:
- Iterative Refinement: The algorithm grows random forests and iteratively re-weights features to stabilize decision paths, making it more robust to the noise inherent in genetic data.
- Signed Interactions: Unlike standard models, lo-siRF tracks whether a mutation is present (+) or absent (-). This allows it to model "switch-like" biological behaviors where the direction of an effect matters.
- Locus-Level Aggregation: To overcome the sparsity of variant-level data, the model aggregates signals into stable gene-level "loci" using stability-driven importance scores.
Phenotyping and Discovery
The study utilized a massive dataset from the UK Biobank, involving 29,661 unrelated individuals. To ensure high-quality phenotyping, the team used a fully convolutional neural network to automatically segment cardiac MRI images, calculating the Left Ventricular Mass index (LVMi) for each participant.
By applying lo-siRF to this cohort, the researchers identified epistatic interactions involving the genes CCDC141, IGF1R, TTN, and TNKS. Notably, several of these loci—particularly those near IGF1R—were statistically insignificant in traditional univariate GWAS. This suggests that a portion of the "missing heritability" in complex diseases may reside in interactions that only become visible when genes are analyzed in combination.
Functional Validation and Network Rewiring
To bridge the gap between statistical association and biological mechanism, the authors integrated transcriptomic data from 313 human hearts (177 failing and 136 non-failing). They found that the genes identified by lo-siRF are highly co-expressed in healthy hearts, forming a functional regulatory network. However, in failing hearts, this network connectivity is significantly weakened or "rewired," suggesting that the breakdown of these epistatic relationships is a hallmark of cardiac pathology.
Further validation was conducted using human induced pluripotent stem cell-derived cardiomyocytes (hiPSC-CMs) carrying a known hypertrophic mutation (MYH7-R403Q). Using siRNA silencing and a novel spiral microfluidic system for high-throughput single-cell imaging, the team tested how these genes interact to influence cell morphology.
Key Results
The experimental results confirmed the presence of non-additive interactions:
- Synergistic Rescue: Digenic (two-gene) silencing of CCDC141 and IGF1R resulted in a 9.3% reduction in median cell size. This effect was significantly greater than the sum of silencing either gene individually, confirming a synergistic epistatic interaction.
- Morphological Fingerprinting: Beyond simple size, the interactions influenced complex morphological traits such as cell "roundness" and internal texture, providing a more nuanced view of how these genes regulate the hypertrophic phenotype.
- Model Performance: lo-siRF achieved a balanced classification accuracy of approximately 55%, outperforming standard methods like LASSO and traditional Random Forests in this low-signal genetic context.
Limitations and Future Directions
While this study represents a significant step forward, several limitations remain. The analysis was conducted primarily on individuals of White British ancestry to maximize signal stability; future work must validate these interactions in more diverse populations. Additionally, while lo-siRF can theoretically detect higher-order interactions (3-way or 4-way), the current signal-to-noise ratio limited the stable identification to pairwise interactions. Future iterations of the pipeline may integrate functional annotations—such as shared regulatory pathways—directly into the machine learning model to further enhance discovery power.
Summary updated