Deciphering epistatic genetic regulation of cardiac hypertrophy

Figures

• 

  Extended Data Fig. 1:. Distribution of LVM and LVMi measurements for 29,661 UK Biobank participants. Left ventricular mass (LVM, a) and LVM indexed to body surface area (LVMi, b) measurements were extracted from cardiac magnetic resonance imaging for 29,661 unrelated White British individuals via deep learning20. c, A high Pearson correlation of 0.92 was observed between these LVM and LVMi measurements.

• 

  Extended Data Fig. 2:. LVMi GWAS using BOLT-LMM and PLINK. Two-sided GWAS p-values from BOLT-LMM (a) and PLINK (b) identified LVMi-associated SNVs, of which the top lead SNV rs3045696 showed the highest significance in both analyses. Other labeled lead SNVs were significant in only one of the two methods. The red dashed line indicates the genome-wide significance threshold (p < 5E-8, two-sided). Notably, rs3045696 and rs67172995 were also stably prioritized by lo-siRF as epistasis interactor variants. Further details are provided in Supplementary Data 1.

• 

  Extended Data Fig. 3:. Differences in local stability scores between high and low LV mass highlight the importance of the lo-siRF-prioritized interactions between genetic loci. a, Schematic of local stability importance score computation. Given a locus (light blue transcript), the local stability importance score for an individual is defined as the proportion of trees for which at least one SNV (shaded black region) in the locus is used in the individual’s decision path. This computation (top) was performed for each individual (denoted by the stacked boxes). Then, a two-sided permutation test was conducted to assess the difference in these local stability importance scores between the low and high LVMi individuals (bottom). b, Differences in the distribution of local stability importance scores suggest that the identified interactions between genetic loci are important for differentiating individuals with high (dark gray) and low (orange) LVMi in the siRF fit. This result, evaluated on the validation data, is stable across the three binarization thresholds and is quantified by a nominal two-sided p-value from a permutation test given in the top right corner of each subplot. Given the use of 10,000…

• 

  Extended Data Fig. 4:. Top SNVs from lo-siRF-prioritized loci and interactions between loci. The most important SNVs and SNV-SNV pairs, as measured by their proportion of occurrence in the siRF fit, are annotated for the top lo-siRF-prioritized interactions between loci in a and top genetic loci in b. The y-axis shows the proportion of decision paths in siRF, for which the SNV or SNV-SNV pair occurs, averaged across all three binarization thresholds. In each of the interactions between genetic loci, rs7591091 in the CCDC141 locus appears most frequently, suggesting a key role in cardiac hypertrophy.

• 

  Extended Data Fig. 5:. Genes mapped from epistatic loci share transcription factors and splicing regulators. Gene pairs exhibiting strong co-association with transcription factors (TFs) and RNA-binding regulators (p < 0.05, two-sided Fisher’s exact test, Fig. 4e) are displayed. Horizontal bars indicate the number of shared TFs or RNA-binding regulators (bottom axis), with the top-ranked TF or regulator, determined by the lowest two-sided enrichment p-value (dots, top axis), labeled on each bar. Bars are color-coded by the corresponding gene set library. Detailed information can be found in Supplementary Data 7.

• 

  Extended Data Fig. 6:. Dendrograms from WGCNA network analyses Dendrograms from WGCNA control (a) and heart failure (b) networks show distinctive gene module structures and modular assignments for CCDC141, IGF1R, and TTN.

• 

  Extended Data Fig. 7:. Spiral-shaped inertial microfluidic channel for cell focusing and imaging. a, Schematic of an inertial microfluidic cell focusing device. Cell suspensions and fresh medium were introduced into the microfluidic device through the cell and sheath flow inlets, respectively, using a syringe pump and flowed down the 5-loop spiral microchannel with the same flow rate (1.2 mL/min). Inserted microscopy images show that randomly dispersed cells were separated by size and bifurcated into the top (large cells) or bottom (small cells) outlets. Scale bar, 10 μm. Outlet channels are connected to straight observation channels where flowing cells were further focused in the channel height direction and imaged using a high-speed camera for morphological feature extraction (Fig. 6c). b, Schematic of the cell focusing principle. The spiral microchannel has a cross-section with a slanted ceiling, resulting in different depths at the inner and outer side of the microchannel. This geometry induces strong Dean vortices (counter rotating vortices in the plane perpendicular to the main flow direction) in the outer half of the microchannel cross-section. The interplay between drag forces (FD) induced by…

• 

  Extended Data Fig. 8:. Epistatic genes non-uniformly reshape cardiomyocyte size distributions. a, Heatmap showing relative differences in cell sizes at various quantile levels between gene-silenced and scramble control conditions for unaffected and MYH7-R403Q variant cardiomyocytes. Larger quantiles correspond to larger cells in the cell size distribution. Dark red indicates strong size reduction at the specified quantile level in gene-silenced cells relative to scramble controls. The corresponding statistical differences (b) were evaluated by the maximum (two-sided) p-values across n = 3 (n = 4 for unaffected cells silencing CCDC141) independent cell batches using a bootstrap quantile test with 10,000 bootstrapped samples (exact p values per batch are available in Source Data). c, QQ-plots of averaged cell size quantiles across cell batches compare gene-silenced cells against scramble controls in unaffected (top row) and MYH7-R403Q variant (bottom row) cardiomyocytes, revealing a clear size-dependent effect of silencing CCDC141-IGF1R on correcting cardiomyocyte hypertrophy.

• 

  Extended Data Fig. 9:. Effects of lo-siRF-prioritized genes and gene-gene interactions on hypertrophic and non-hypertrophic cell morphology. Relative differences in median cell size (a), normalized peak number (b), and roundness error (c) were analyzed separately for cells size-sorted into the top (hypertrophic cells, green) and bottom (non-hypertrophic cells, orange) microchannel outlets (Extended Data Fig. 7), as well as for all cells combined (blue). For both unaffected (left) and MYH7-R403Q variant (right) cardiomyocytes, the gene-silencing induced relative differences in each morphological feature (a-c) were quantified as (ms – mc)/mc × 100%, where ms and mc denote median values in gene-silencing and scrambled control conditions, respectively. Squares represent mean relative differences across n = 3 independent cell batches (n = 4 for unaffected cells silencing CCDC141), with overlaid points indicating medians of individual replicates. Violins display mean relative differences from 1,000 bootstrap samples, with error bars indicating standard deviations. Asterisks indicate significant differences compared to scramble controls, based on the maximum p-values of two-sided Mann-Whitney U test across…

• 

  Fig. 1:. Schematic of the study workflow. The study workflow includes four major stages: (a) derivation of left ventricular mass from cardiac magnetic resonance imaging (green boxes); (b) computational prioritization of epistatic drivers (orange); (c) functional interpretation of the hypothesized epistatic genetic loci (purple); and (d) experimental confirmation of epistasis in cardiac tissues and cells (blue). Abbreviations: MRI, magnetic resonance imaging; LV, left ventricle; LVM, left ventricular mass; LVMi, left ventricular mass indexed by body surface area; SNV, single-nucleotide variant; GWAS, genome-wide association study; PLINK26 and BOLT-LMM27, two different GWAS software packages; lo-siRF, low-signal signed iterative random forest; ANNOVAR74, a software for functional annotation of genetic variants; CADD36, combined annotation dependent depletion, which scores the deleteriousness of variants; RegulomeDB35, a database that scores functional regulatory variants; ChromHMM34, a multivariate Hidden Markov Model for chromatin state annotation; eQTL, expression quantitative trait locus; sQTL, splicing quantitative trait locus; Hi-C, high-throughput chromosome conformation capture;…

• 

  Fig. 2:. Low-signal signed iterative random forest (lo-siRF) workflow. a, Lo-siRF took as input single-nucleotide variant (SNV) data and cardiac MRI-derived left ventricular mass indexed by body surface area (LVMi) from 29,661 UK Biobank participants. b, Two genome-wide association studies (GWAS) were applied using different software to reduce the dimensionality of SNVs. c, LVMi was binarized into high and low categories using three thresholds (shown as stacked boxes). d, For each threshold, a signed iterative random forest (siRF) was fitted using the GWAS-filtered SNVs to predict the binarized LVMi, achieving prediction accuracy better than or comparable to other methods, before interpreting the model fit. e, siRF aggregates SNVs into genetic loci using ANNOVAR74 and finally ranks loci and their interactions based on a stability-driven importance score averaged across the three binarization thresholds.

• 

  Fig. 3:. Lo-siRF finds epistasis between genetic risk loci for left ventricular hypertrophy. a, Circos plot illustrating genetic risk loci identified by lo-siRF (green, circle 2) and regions after clumping FUMA-extracted SNVs in LD (r2 > 0.6) with lo-siRF-prioritized SNVs (black, circle 3). Circle 1 shows the top 300 epistatic SNV-SNV pairs ranked by occurrence frequency in lo-siRF (green), SNV-gene linkages (FDR < 0.5) from GTEx37 V8 cis-eQTL data of heart and skeletal muscle tissues (purple), and 3D chromatin interactions33 from Hi-C data of left ventricular tissue (GSE87112). Circle 5 and 6 show occurrence frequencies and the number of partner SNVs in epistasis (normalized by the maximum value per locus) identified by lo-siRF, respectively. Circle 7 displays the ChromHMM34 core-15 chromatin state for left ventricle (LV), right ventricle (RV), right atrium (RA), and fetal heart (Fetal). Circle 8 presents a GWAS Manhattan plot (points, two-sided p < 0.05 from PLINK26), where SNVs are color-coded by their maximum r2 value relative to the 283 lo-siRF-prioritized SNVs. The outer heatmap layer represents LD-linked SNVs (r2 > 0.6) extracted from the FUMA reference panel, which lack GWAS…