Supplemental Material

HCM participants
In total 710 patients with a clinical diagnosis of HCM, either seen in the inherited cardiomyopathy service or referred for
CMR imaging, were consecutively enrolled into a prospective registry at the National Institute for Health Research (NIHR)
Royal Brompton Hospital Cardiovascular Biobank project between 2009-2015, of whom 436 were included in this study.
All participants provided written informed consent and the study was approved by the National Research Ethics Service
(19/SC/0257). HCM diagnosis was independently adjudicated by a cardiomyopathy specialist based on established clinical
and CMR criteria where all patients met the American Heart Association criteria for diagnosis.10 This was defined as a wall
thickness of 15mm or greater, or 13–14mm if there was a first degree relative with HCM, not explained by another cardiac
or systemic disease-causing abnormal loading conditions, or had disproportionate apical wall thickness and tapering in
keeping with an apical HCM phenotype.35
Patients were excluded from analysis based on age (< 16 years at time of CMR), missing demographic or clinical data,
contraindication to CMR, previous history of septal ablation, cardiac transplantation or myectomy at baseline. A history of
hypertension or diabetes was documented, as well as current medication at time of enrolment to the study. The cohort un-
derwent detailed clinical, imaging and genetic assessment. All patients underwent CMR for assessment of cardiac chamber
volumes and function (1.5T, Siemens Sonata or Avanto, Siemens Medical Systems, Erlangen, Germany). Variables reported
were collected at enrolment to the study. The CMRs selected for analysis were those closest to the date of enrolment
or the first diagnostic study available. Where present, left ventricular outflow tract obstruction (LVOTO) was confirmed
through stress echocardiography. Unrelated Singaporean patients with a diagnosis of HCM (n = 60) were prospectively
recruited from the National Heart Centre Singapore. Patients gave written informed consent to participate which was ap-
proved by the Singhealth Centralised Institutional Review Board (2020/2353) and Singhealth Biobank Research Scientific
Advisory Executive Committee (SBRSA 2019/001v1). Singaporean subjects underwent an equivalent CMR protocol at 1.5T
(Aera, Siemens, Erlangen, Germany) or 3T (Ingenia, Philips, Best, Netherlands).
Conventional CMR analysis was undertaken by accredited operators using semi-automated software (CMRtools, Car-
diovascular Imaging Solutions, London, UK).

UK Biobank participants
The UKB study recruited 500,000 participants aged 40 to 69 years old from across the United Kingdom between 2006 and
2010 (National Research Ethics Service, 11/NW/0382).36 This study was conducted under terms of access approval number
40616. In each case, written informed consent was provided.
A sub-study of UKB invited participants for CMR for assessment of cardiac chamber volumes and function using a
standard protocol (1.5T, Siemens Aera, Siemens Medical Systems, Erlangen, Germany).37 As a reference population, we
selected 16,691 participants that did not meet criteria for left ventricular hypertrophy and were classified as genotype neg-
ative (SARC-NEG) by having no variants in genes that may cause or mimic HCM (see Sequencing and variant categorisation).

Cardiac phenotyping using machine learning

Segmentation of the cine images in both UKB and HCM groups was performed using a deep learning neural network
algorithm developed and optimised in-house. The performance of image annotation using this algorithm is equivalent to
a consensus of expert human readers and achieves sub-pixel accuracy for cardiac segmentation.38 The label maps were
super-resolved and registered to a cardiac atlas enabling consistent quantitative three-dimensional phenotypic analysis
within and between patient groups.39
Myocardial wall thickness was measured along radial line segments connecting the endocardial and epicardial surfaces
perpendicular to the myocardial centreline and excluding trabeculae. Chamber volumes and mass were calculated from
the segmentations according to standard post-processing guidelines.40 Myocardial strain analysis was performed using
non-rigid free-form deformation image registration.41 Trabecular traits were quantified using fractal dimension (FD) anal-
ysis where a higher value indicates more complex trabeculation.42

Sequencing and variant categorisation

Panel sequencing was completed in the HCM patient cohort, as previously described.43 The patients were sequenced
using either a custom SureSelect capture panel targeting genes associated with inherited cardiac conditions or the Illumina
TruSight Cardio panel. Sequencing was performed on either the SOLiD 5500xl platform, or the Illumina HiSeq, MiSeq or
NextSeq platforms. The Singaporean cohort underwent targeted genetic sequencing using the TruSight Cardio panel and
an equivalent pipeline as previously reported.44
Patients were divided into three genetic strata. Patients carrying at least one potentially-causative rare variant (allele
frequency <0.00004)45 in any of 8 sarcomere-encoding genes robustly associated with HCM were considered genotype
positive. These were further stratified into (i) those carrying variants previously confidently classified as pathogenic / likely
pathogenic (SARC-P/LP) in ClinVar, and confirmed on our review, or else curated as P/LP according to ACMG criteria using

Curran et al. 2023 |

 the semi-automated CardioClassifier decision support tool46 (n = 107), as previously published,5 ; and (ii) those carrying
sarcomeric variants of uncertain significance (SARC-VUS), comprising variants in the same 8 genes, that are consistent
with known disease mechanisms and sufficiently rare, but with insufficient evidence to classify robustly as P/LP.47,48
Individuals were classified as genotype negative (SARC-NEG) if they had no rare protein-altering variant (minor allele
frequency <0.001 in the UKB and the Genome Aggregation Database)49 in any of 25 genes that potentially cause HCM
(definitive or moderate evidence according to international curation)50 or cause syndromes that can present with isolated
left ventricular hypertrophy (genocopies).50 In order to generate the most robust set of true genotype negatives, individu-
als carrying a protein-altering variant in any these 25 genes, but that was not sufficiently rare to be considered potentially
causative of monogenic HCM was excluded from the analysis. Further details are given in Supplementary Materials. Com-
mon genetic variation contributes substantially to HCM risk and we also assessed the relationship between phenotype and
polygenic score (PGS) derived from a case-control HCM genome wide association study (GWAS) in the 100,000 Genomes
Project.32

Variant curation pipeline

This pipeline can be found on GitHub (https://github.com/ImperialCollegeLondon/HCM-taxonomy) and has been previously
published.5 All genetic data was annotated using Ensembl Variant Effect Predictor (VEP; version 105)51 with plugins for
NMD, SpliceAI (version 1.3.1),52 ClinVar (version 2022 01 15),53 gnomAD (version r2.1),49 and LOFTEE.49 The VEP output
was analysed using R (version 3.6.0).
Protein-altering variants, defined using MANE transcripts, that had a MAF of <0.1% in gnomAD for variants identified
in cases, and <0.1% in gnomAD and UK Biobank for variants identified in the UK Biobank, were included in the analy-
ses. Protein altering variants were specified as high or moderate impact by Sequence Ontology and Ensembl, with the
addition of splice region variants for further curation. The variants were filtered for genes and protein consequences
of interest,50 to include 8 definitive-evidence sarcomeric HCM genes (MYH7, MYBPC3, MYL2, MYL3, ACTC1, TNNI3, TNNT2,
TPM1), 3 medium-evidence HCM genes (CSRP3, TNNC1, JPH2), 2 intrinsic cardiomyopathy genes (ACTN2 (moderate classifi-
cation), PLN (definitive classification)), and 12 syndromic genes that can cause isolated left ventricular hypertrophy (FHL1,
TTR, FLNC, GLA, LAMP2, PRKAG2, PTPN11, RAF1, RIT1, ALPK3, CACNA1C, DES). FLNC, ALPK3, ABCC9, CRYAB, MYO6, and RIT1, were
not sequenced in cases, but were analysed in UK biobank. No protein-altering variants were identified in TNNC1 in the case
cohort.
Splice region variants (outside the canonical splice donor and acceptor sites) were assessed in two ways; i) via ClinVar
report: splice region variants found pathogenic with at least 2 star evidence for HCM in ClinVar and reported functional
evidence for splicing were termed “splice confirmed”; if the functional evidence was unclear the protein consequence
remained unchanged; if there was functional evidence of an alternative mechanism to splicing, the protein consequence
was renamed (e.g. missense variant); ii) via prediction threshold: of the splice region variants, they were excluded if they
did not meet the spliceAI threshold of >0.8, and these thresholds were used to identify potentially splice-causing variants
of those splice region variants identified with a non-synonymous consequence flag (e.g. intron variant).
The pipeline then consisted of three main filtering steps which resulted in an output of four columns of binary code
flagging genotype status (heterozygous, compound heterozygotes, and homozygotes, combined) as “1”: SARC-NEG – Indi-
viduals who do not harbor any rare non-synonymous variants in any of the 25 genes of interest. This was a stringent filter
to identify an unambiguous genotype-negative control group.
SARC-VUS – Individuals harboring rare variants in one or more of the 8 definitive HCM-associated sarcomere-encoding
genes. Rare variants were restricted to known disease-associated variant classes. This step separated the variants into
two subsets: i) Loss of function (LoF) alleles (group A), which contained only the gene MYBPC3, and filters for the protein
consequences of stop gained, splice acceptor variant, splice donor variant, frameshift variant, and splice region variant
(with additional in silico evidence of an effect on splicing). LOFTEE was incorporated in this step to exclude loss of function
(LoF) variants that were flagged as “low confidence” (LC) and other LOFTEE flags, such as “NAGNAG sit” requiring reannota-
tion to non-LoF variant status; ii) Protein altering (PAV) alleles (group B), which included all 8 sarcomeric genes, including
MYBPC3, and filters for the protein consequences of missense variant, inframe insertion, and inframe deletion.
Both groups included additional positional annotation (LoF variants found in the last exon or 55bp into the penulti-
mate exon or stop gained variants with a NMD flag using the NMD plugin), this included variants that introduce a protein-
truncating variant (PTC) and predicted to lead to nonsense-mediated decay (NMD). The variants flagged ‘coding sequence
variant’ and ‘protein altering variant’ were manually curated, as were ‘stop lost’ and ‘start lost’ which were examined via EN-
SEMBL sequence and UCSC Genome Browser to identify in-frame rescues nearby. To be included in the SARC VUS group,
the variants were required to meet a maximum gnomAD filter allele frequency (FAF) threshold for HCM (<0.00004) and
excluded variants deemed P/LP for DCM on ClinVar.
SARC-P/LP is as SARC-VUS, plus annotated as P/LP according to the ACMG guidelines.48 Variants were reviewed if re-
ported as P/LP for HCM by at least one submitter in ClinVar, or if flagged as P/LP by the CardioClassifier decision support
software.46 Variants that did not meet either of these criteria were not individually reviewed.

Curran et al. 2023 |

 Outcome measures
Data were collected to measure all-cause mortality in the HCM cohort. Outcomes were verified through search of the NHS
Shared Care Records. Patients were followed up for a median of 10.2 years from date of study enrollment.

Statistical analysis and data modelling

Statistical analysis was performed with R (version 4.0.3) and RStudio Server (version 1.2; Boston, MA), unless otherwise
stated. Variables were expressed as percentages if categorical, mean ± standard deviation (SD) if continuous and normal,
and median ± inter-quartile range (IQR) if continuous and non-normal. Baseline anthropometric data were compared by
Kruskal-Wallis tests and, if differences were identified, a Wilcoxon test was used for pairwise comparisons with Benjamini-
Hochberg adjustment for multiple testing. Clustering of clinical data from HCM patients was performed using UMAP (uni-
form manifold approximation and projection)54 for dimensionality reduction followed by a K-means algorithm evaluated
with a silhouette score to find the optimal number of clusters.
The association between genotype and three-dimensional phenotype was assessed by fitting univariable regression
models at each vertex of the cardiac mesh, controlling for false discovery, with corrected beta coefficients plotted on the
epicardial surface.5,55 Clustering of subjects by their 3D left ventricular wall thickness, adjusted for age, sex and ancestry,
was performed through partitioning the shared nearest neighbour (SNN) graph56 with the multilevel refinement Leiden
algorithm.57 The SNN graph and its partitions were determined using the functions available in Seurat.58 The clusters
were visually inspected by UMAP projection and their stability was assessed through bootstrapping using fpc. We used
DDRTree (discriminative dimensionality reduction via learning a tree) to project the 3D left ventricular wall thickness into
a 2D tree structure to visualise the distribution of HCM phenotypes.12,59
The predictive power of the DDRTree mapping for P/LP genotype was tested with a generalized additive model (GAM)
fitted on the tree coordinates, using a 10-fold cross validation repeated 3 times. Survival probability to median observed
age was estimated with a Cox proportional hazards model fitted on the cubic spline of tree coordinates of each individual
and their relative position in the tree. Association between the DDRTree mapping and polygenic risk score was assessed
with a logistic regression GAM model using the subjects’ coordinates as independent variables and the binarized polygenic
risk score (thresholded at median) as outcome. The predictions were obtained using a 10-fold cross validation.
The primary survival analysis was performed in individuals from the HCM cohort with chronological age as time-scale
and adjusting for genetic sex and ancestry (dichotomised by white European ancestry). Participants with SARC-P/LP vari-
ants were compared with pooled participants without variants and with SARC-VUS carriers. Proportional hazards assump-
tion as assessed using Schoenfeld residuals was not violated.

Estimation of wall thickness from meshes

Three dimensional meshes with the local wall thicknesses (WT) were generated from the myocardial segmentations of the
end systolic (ES) and end diastolic (ED) phases, using previously published methods.55,60,61 In order to make the modelling
of the WT computationally tractable, the meshes were decimated by 99%. Specifically, because of the one-to-one corre-
spondence between subjects’ meshes vertices and atlas vertices, the decimation was applied to the atlas only, and the
closest resulting vertices to the original atlas were selected from all meshes. This allowed us to preserve the correspon-
dence between the individual vertices across all meshes.

Unsupervised analysis of wall thickness values

Wall thicknesses were firstly adjusted by age at cardiac magnetic resonance (CMR) imaging, sex and ancestry using a linear
regression, and normalized afterwards using the Seurat package for R.58 To remove the intrinsic correlations between the
values, due to the spatial nature of the data, and to reduce the effect of statistical noise, the data was compressed applying
a principal components analysis (PCA), and retrieving the first 50 principal components.
A shared nearest neighbor graph (SNN) was built from the compressed data, where the nodes corresponded to the
subjects, and the edges corresponded to the Jaccard index between the nearest neighbours of each pair of subjects. The
nearest neighbors corresponded to the 20 subjects with the smallest cosine distance from each subject. Thus, the SNN
graph was partitioned using the multilevel Louvain algorithm, with resolutions varying from 0.1 to 1 in steps of 0.1. Both
SNN and its partitioning were performed using Seurat package for R.58
The optimal resolution was chosen by inspecting the cluster stability with clustree, corresponding to the partitioning
before any mixing of cluster assignment was visible62 (Supplementary Figs. II, III). Additionally, stability of the clusters was
assessed by rerunning the clustering on 1000 random subsets, each with a size equal to 80% of the whole cohort, using
the function clusterboot from the fpc package for R.
If more than one partitioning was found corresponding to the same branch structure of the clustree plot, that with
the greater stability was chosen (Supplementary Tables I, II).

Curran et al. 2023 |

 DDTree modelling of wall thickness values
DDRTree was applied on the adjusted wall thicknesses for age at CMR, sex and ancestry. Following the same approach
described in the previous section, the first 50 principal components were calculated and used as input for the DDRTree
algorithm.19 All parameters were kept as default. The underlying tree structure, the result of the procedure, was auto-
matically partitioned into branches by considering as branching those points with a degree equal to 3. Subsequently, we
merged small branches into their larger neighbour, such that the main structure of the tree was preserved. This step aimed
at reducing the number of phenotypic sub-types and increasing the statistical power of the subsequent modelling. Firstly,
branches consisting of less than 5 individuals were merged to the closest connected branch. Subsequently, short terminal
branches, with a geodesic length smaller than 5% of the graph length were merged with their parent with respect to the
center of the tree, as they represented low-diversity individuals. The branches connected to these were then merged, as
bifurcation no longer existed (Supplementary Fig. VIII- IX).
A set of clinical and anthropometric measures were tested for statistical association with the tree branches. We first
applied a Kruskal Wallis test to continuous variables and 𝜒 2 test to discrete variables. Those variables showing a signifi-
cant association (Benjamini-Hochberg adjusted P < 0.05) were post-hoc tested. A Dunn test was used to test each pair of
branches, whereas the exact Fisher test was used for discrete variables in one-vs-all fashion. All P-values were adjusted
for multiple testing using Benjamini-Hochberg method.
Finally, we tested the consistency of the tree structure after removing 8 related subjects from our original analysis. The
association between the tree’s main axes and 1) genotype status, and 2) PRS remained unchanged (Supplementary Table
V-VI). Additionally, the relative position of the subjects was consistent between the two trees, as shown by the high value
of the correlation between the pseudotime of the two models (Spearman’s rho = 0.89, Supplementary Fig. X).

External validation
Wall thicknesses were residualized by linear regression using sex and age at scan as covariates. In this way, we could
compare the similarities between the intra-cohort statistical properties of WT values. The first 5 principal components were
estimated from the development cohort and used to project the adjusted Singaporean WT values. The development cohort
principal component scores were used to fit two random forest models aimed at predicting the x and y tree coordinates.
Performances of the models were evaluated using a 10-fold cross-validation, repeated 3 times. The fitted models were
therefore used to predict the tree coordinates of the Singaporean individuals, from their principal component scores.
Faithfulness of the tree mapping for the Singaporean cohort was evaluated by the “trustworthiness” M1 measure, which
estimates how observations that are similar in the original high dimensional space are placed close to each other in the
low dimensional space. It ranges from 0 to 1, with larger values corresponding to a better representative low dimensional
mapping.13
In order to evaluate the consistency of the local statistical patterns between the development tree and the projected Sin-
gaporean individuals, we considered Spearman’s correlation between nearest neighbour points in the tree. We estimated
the correlations between the adjusted WT of the closest points in the development tree, and compared their distribution
with that of the correlations between the Singaporean WT and their closest points in the development tree. In order to
test differences in the distributions, we used a Wilcoxon test for difference of medians, and a Kolmogorov Smirnov test for
difference of distributions.

Supplementary Tables I- XI and Figures I-XI

Supplementary Table I. Cluster stability at end diastole. Results from the analysis of the cluster stability for end diastolic wall thickness
(ED WT) in 1000 subsets. The possible values range between 0 and 1, where 0 means that the cluster is not stable, and 1 that the clusters
are identical in all repetitions. Between a resolution of 0.4 and 0.5, determined from the clustree plot (Supplementary Fig. II). We chose
the resolution of 0.5 because it is also characterized by more stable clusters. Cluster 2 is found as an intermediate state between cluster 0
and 1, and it is less stable than the other two.

resolution cluster 0 cluster 1 cluster 2

0.4 0.95 0.85 0.50
0.5 0.95 0.92 0.78

Curran et al. 2023 |

 Supplementary Table II. Cluster stability at end systole. Results from the analysis of the cluster stability for end systolic wall thickness
(ES WT) WT in 1000 subsets. The possible values range between 0 and 1, where 0 means that the cluster is not stable, and 1 that the
clusters are identical in all repetitions. Resolutions from 0.1 to 0.6 are determined from the clustree plot (Supplementary Fig. III). The
resolution of 0.1 was chosen because it is the lowest value among those with more stable clusters.

resolution cluster 0 cluster 1

0.1 0.96 0.96
0.2 0.96 0.96
0.3 0.96 0.96
0.4 0.96 0.96
0.5 0.96 0.96
0.6 0.94 0.95

Supplementary Table III. Pathogenic/likely pathogenic variant prediction from tree coordinates. Fitted parameters for the GAM
model used to predict individuals with P/LP variants using the 2 tree coordinates. 1 OR, Odds ratio; CI, Confidence interval

ED ES
Characteristic log(OR)1 95% CI1 p-value log(OR)1 95% CI1 p-value
(Intercept) -1.1 -1.4, -0.93 <0.001 -1.2 -1.4, -0.96 <0.001
s(Z1) 0.008 0.013
s(Z2) 0.2 0.062

Supplementary Table IV. Singaporean branch assignment. Singaporean HCM patients were assigned to the tree branches of their
nearest neighbours in the development tree. In both end diastole (ED) and end systole (ES), most of the individuals were assigned to
branches 1 and 4.

Branch ED ES
1 23 21
2 13 1
3 4 2
4 20 35
5 - 1

Supplementary Table V. Association between ES tree coordinates and genotype status. Coefficients of the logistic regression
between the ES tree coordinates and the genotype status (genotype == "P/LP"). The results for the whole cohort tree (left) and the
unrelated subjects tree are consistent. 1OR = Odds Ratio, CI = Confidence Interval

Full cohort Reduced cohort

Characteristic OR1 95% CI1 p-value OR1 95% CI1 p-value
(Intercept) 0.31 0.25, 0.39 <0.001 0.31 0.24, 0.38 <0.001
Z1 1.04 1.01, 1.06 <0.001 1.04 1.02, 1.06 <0.001
Z2 1.02 0.94, 1.11 0.6 0.99 0.91, 1.08 0.8

Supplementary Table VI. Association between ES tree coordinates and PRS. Coefficients of the logistic regression between the ES tree
coordinates and PRS (PRS == "high"). The results for the whole cohort tree (left) and the unrelated subjects tree are consistent. 1OR = Odds
Ratio, CI = Confidence Interval

Full cohort Reduced cohort

Characteristic OR1 95% CI1 p-value OR1 95% CI1 p-value
(Intercept) 0.99 0.78, 1.25 >0.9 0.96 0.76, 1.22 0.8
Z1 1.00 0.98, 1.02 >0.9 1.00 0.98, 1.03 0.8
Z2 1.14 1.05, 1.25 0.002 1.12 1.03, 1.22 0.012

Curran et al. 2023 |

 A

Branch 1 2 3 4

1.32

Beta Coefficients
0

-1.32

B
Weight Body Surface Area ED Volume ACE/ARB ASA/Clopi Beta Blocker

1 1 1
1
* 1
** 1

2 Value 2 Value 2 Value

Branch

Branch

Branch
No No No
3 Yes 3 Yes 3 Yes
2 2 2

** * **
Branch

Branch

Branch

4 4 4
***

***

***
***

***

***

3 3 3 0 50 100150 0 50 100150 0 50 100150

Num. subjects Num. subjects Num. subjects
**

4 4 4 Hypertension Infarction LV Gadolinium

70 80 90 100 1.7 1.8 1.9 2.0 2.1 100 120 140 160
1 1
** 1
*** Value
Weight BSA EDV 2 Value 2 Value 2
Branch

Branch

Branch
1
No No 2
3 Yes 3 Yes 3 3
Left Ventricle Mass LV Max Wall Thickness SV 4
4
*** 4 4

1 1 1 0 50 100150 0 50 100150 0 50 100150

Num. subjects Num. subjects Num. subjects

LVOTO Most Aff. Level

***

***

2 2 2
Branch

Branch

Branch
***
***

***
***

***

1
** 1
*
***

***

***

3 3 3 Value
2
*** Value 2
**
Branch

Branch

Apex
***

**

No Base
3 Yes 3
Mid
4 4 4
4
** 4

150 200 250 15.0 17.5 20.0 22.5 25.0 80 90 100 110 120 0 50 100150 0 50 100150
LVM LV Max WT SV Num. subjects Num. subjects

Supplementary Figure I. Phenotypic tree from 3D end-diastolic wall thickness. a. The projection of patients’ 3D end-diastolic (ED) wall thickness (WT) by the
DDRTree dimensionality reduction reveal the presence of four main branches that are associated to specific morphological changes of the myocardium. Each branch is
represented by the decimated ED atlas mesh, coloured accordingly to the beta coefficients resulting from testing the average difference between each branch individual
and the other subjects. The yellow contour denotes the areas with a beta significantly different from zero. b. The continuous and discrete phenotypic variables found to
be significantly associated to at least one branch. For left ventricular (LV) Gadolinium, labels are as follows: 1: None, 2: Minimal, 3: Moderate and 4: Severe. The
significance for the enrichment of discrete variables is reported within the bars. ACE, Angiotensin-converting enzyme inhibitors; Aff, affected; ARB, Angiotensin receptor
blockers; ASA, aspirin; Clopi, clopidogrel; LVOTO, Left ventricular outflow tract obstruction; SV, stroke volume. Only the significant pairs are reported with the symbols:
∗ 𝑃 ≤ 0.05; ∗∗ 𝑃 ≤ 0.01; ∗∗∗ 𝑃 ≤ 0.001; ∗∗∗∗ 𝑃 ≤ 0.0001, 𝑛 = 436.

Curran et al. 2023 |

 A C
Cluster 0 1 2
0.1 0.3 0.5 0.7 0.9
res.
0.2 0.4 0.6 0.8 1

4 0 1

0 1
3
0 1
UMAP2

2 0 2 1

0 2 1
1
0 2 1 3

0 1 2 0 3

1 2 0 3
−1 2 1 0 3
0 2 4 6 1 2 3 0 4
UMAP1

B D
Genotype NEG PLP VUS
Number of subjects

150
4
Genotype
3
100 NEG
UMAP2

2 P/LP
VUS
1 50

0
0
−1
0 1 2
0 2 4 6 Cluster
UMAP1

Supplementary Figure II. Selection of optimal resolution for clustering of end diastolic wall thickness. The optimal resolution for the Louvain partitioning is found
by inspecting the clustree plot (c.). In this case, the value of 0.5 was chosen, corresponding to the resolution with stable branching before any assignment mixing
(diagonal arrows), and the largest bootstrapping stability (Supplementary Table I). The UMAP projections in a. and b. show the parallelism between the clusters and the
genotypes. d. Genotype proportions by cluster.

Curran et al. 2023 |

 A C
Cluster 0 1
0.1 0.3 0.5 0.7 0.9
res.
0.2 0.4 0.6 0.8 1

0 1
8
0 1

0 1
UMAP2

0 1
6
0 1

0 1

4 0 2 1 3

0 2 1 3

0 2 4 1 3

0.0 2.5 5.0 7.5 10.0 0 4 2 5 1 3

UMAP1

B D
Genotype NEG PLP VUS

200
Number of subjects

8 150
Genotype
NEG
UMAP2

100 P/LP
6
VUS
50
4
0
0 1
0.0 2.5 5.0 7.5 10.0 Cluster
UMAP1

Supplementary Figure III. Selection of optimal resolution for clustering of end systolic wall thickness. The optimal resolution for the Louvain partitioning is found
by inspecting the clustree plot (c.). In this case, the value of 0.1 was chosen, corresponding to the resolution with stable branching before any assignment mixing
(diagonal arrows), and the largest bootstrapping stability (Supplementary Table II). The UMAP projections in a. and b. show the parallelism between the clusters and the
genotypes. d. Genotype proportions by cluster.

Curran et al. 2023 |

 ES Branch 1

ES Branch 2

ES Branch 3

ES Branch 4

ES Branch 5
0.4

0.36
ED Branch 1 0.37 0.09 0.05 0.14 0.09
0.32

0.28

ED Branch 2 0.21 0.22 0.1 0.07 0.08

0.24

Jaccard Index
0.2

0.16
ED Branch 3 0.02 0.08 0.22 0.18 0.03
0.12

0.08

ED Branch 4 0.01 0 0.04 0.4 0.01

0.04

Supplementary Figure IV. Co-occurrence between end diastolic and end systolic tree branches. The Jaccard index of the subjects membership for the branches in
the DDRTree from ED and ES WT shows that branches 1 to 4 have the largest co-occurrence and they can be considered capturing a similar phenotypic subpopulation.
Branch 5 in end systolic DDRTree is not found in the end diastolic DDRTree and consists of an average sub-type of the cohort.

Frame ES
0.950
Survival probability (69 yrs)

0.925

0.900

0.875

0.00 0.25 0.50 0.75 1.00

Position in branch 4

Supplementary Figure V. Survival probability in end systolic branch 4. The more distal regions of branch 4 correspond to lower probability of survival at a
chronological age of 69 years. The OR between the distal and base points of the branch is 0.9773.

Curran et al. 2023 |

 A B

ED − Projected Singaporean data ES − Projected Singaporean data

set RBH Sing. set RBH Sing.

0
Z2

Z2
−4
−5

−8

−20 −10 0 10 −20 −10 0 10 20

Z1 Z1

Supplementary Figure VI. Predicted tree coordinates for the Singaporean cohort. The coordinates predicted by the two random forest models for the Singaporean
cohort follow the original spatial distribution of the development cohort, with few points falling outside the main structure of the tree, in both end diastole (a.) and end
systole (b.). Sing, Singaporean patients; RBH, Royal Brompton Hospital patients.

A ED B ES

0.5 0.5
Spearman's rho

Spearman's rho

dataset dataset
RBH RBH
Sing. Sing.
0.0 0.0

−0.5
−0.5
RBH Sing. RBH Sing.
Set Set

Supplementary Figure VII. Similarity between nearest tree points. Spearman’s correlation between the adjusted wall thickness of the nearest RBH points in the tree
follow the same distribution of the nearest Singaporean and RBH points, in both end diastole (a.) and end systole (b.). Sing, Singaporean patients; RBH, Royal Brompton
Hospital patients.

Curran et al. 2023 |

 1 4 7 10 13 2 4 6 10 13
colour
colour 2 5 8 11 14 3 5 9 11 15

3 6 9 12 15

5
4
5 2
4 5

Component 2
2 63
Component 2

5 0
63
0

1
1 −5 7
−5 7

−20 −10 0 10 −20 −10 0 10

Component 1 Component 1

colour 10 16 18 21 State_merged 10 16 18 21

5
4 6
3
2
2
5
Component 2

63
0

1
1
−5 7 7

−20 −10 0 10
Component 1

Supplementary Figure VIII. Branch merging for ED phase. Intermediate results of the branch merging process for the ED phase data. The graph at the top-left shows
the branch labels determined by ‘monocle‘ that were progressively merged: a) short leaf branches, b) branches that did not have bifurcation into two different states. The
final branches qualitatively correlate with the hierarchical structure of the tree, where the root state correspond to the center of the tree.

1 4 7 10 1 3 7 10
colour
colour 2 5 8 11 2 5 8 11

3 6 9
4

4 5
5 4
2 1
Component 2

0
4
Component 2

0 2 1 3
3

−4
−4

−8 −8
−20 −10 0 10 20 −20 −10 0 10 20
Component 1 Component 1

State_merged 1 2 11 12 14
colour 1 2 11 12 14

4
1

5
4
2 1
Component 2

0 2
4
3 3

−4 5

−8
−20 −10 0 10 20
Component 1

Supplementary Figure IX. Branch merging for ES phase. Intermediate results of the branch merging process for the ES phase data. The graph at the top-left shows the
branch labels determined by ‘monocle‘ that were progressively merged: a) short leaf branches, b) branches that did not have bifurcation into two different states. The
final branches qualitatively correlate with the hierarchical structure of the tree, where the root state correspond to the center of the tree.

Curran et al. 2023 |

 A C
Pseudotime
0 10 20 30 40

5 40 Spearman's rho = 0.8898 (p < 0.0001)

42 1
0
Component 2

3 Spearman's rho = 0.8898 (p < 0.0001)

−4 Pseudotime new tree

30

−8
−20 −10 0 10 20
Component 1 20

B
Pseudotime
0 10 20 30 10

5.0

2.5 1 0
5 3 8
2
Component 2

0.0 0 10 20 30 40
4 6
Pseudotime original tree
−2.5 7
−5.0

−7.5
−20 −10 0 10 20
Component 1

Supplementary Figure X. Relative positions of subjects in the original and unrelated subjects tree. The pseudotime calculated from the whole cohort tree (A) and
that without unrelated subjects (B) shows that the subjects relative positions are highly correlated (Spearman’s rho = 0.89) (C). This suggests that the tree generated after
removing the related subjects was consistent with the one from the whole cohort.

Curran et al. 2023 |

 Supplementary Table VII. Participant characteristics and CMR-derived cardiac measurements in UK Biobank. BSA, body surface
area; concentricity, (left ventricular mass / left ventricular end-diastolic volume); CMR, cardiac magnetic resonance imaging; DBP, diastolic
blood pressure; EDV, end-diastolic volume; EF, ejection fraction; ESV, end-systolic volume; FD, fractal dimension; LV, left ventricular; LVM,
left ventricular mass; LVMI, left ventricular mass index (LVM/BMI); peak diastolic strain rate, PDSR; RA, right atrial; RV, right ventricular; SBP,
systolic blood pressure; WT, wall thickness. *Medication for cholesterol, blood pressure, diabetes.

Characteristic UKBB n=16,691

Female 8,775 (52.5)
Age at scan, y 55 ± 7.5
White 14,683 (87.9)
BSA, m2 1.9 ± 0.2
LVEDV, ml 148 ± 33.5
LVESV, ml 60.4 ± 19
LVEF, ml 59.6 ± 6
LVM, g 86 ± 22.1
LVMI, g/m2 45.8 ± 8.5
LV maximum WT, mm 9.4 ± 1.6
Mean apical FD 1.21 ± 0.05
Mean basal FD 1.19 ± 0.03
Mean global FD 1.17 ± 0.03
LV global radial strain, % 45 ± 8.3
LV global circumferential strain, % -22.3 ± 3.4
LV global longitudinal strain, % -18.5 ± 2.8
LV radial PDSR -5.7 ± 2
LV longitudinal PDSR 1.7 ± 0.6
LV concentricity, g/mL 0.58 ± 0.08
Heart rate, min 69.5 ± 11.6
Hypertension 4,857 (29)
On medication* 2,241 (13.4)
SBP, mmHg 137.5 ± 18.1
DBP, mmHg 78.6 ± 9.9

Supplementary Table VIII. Characteristics of Singaporean HCM cohort. BSA, body surface area; SBP, systolic blood pressure.

Characteristic Singaporean HCM n=60

Female 11 (11.7)
Age at scan, y 58.9 ± 20
Chinese 52 (86.7)
BSA, m2 1.8 ± 0.2
SBP, mm Hg 137 ± 24.8
Genotype SARC-NEG 28 (46.7)
Genotype SARC-VUS 16 (26.7)
Genotype SARC-P/LP 16 (26.7)

Supplementary Table IX. Cumulative hazard model. All cause mortality in individuals with hypertrophic cardiomyopathy carrying
pathogenic or likely pathogenic sarcomeric variants (SARC-P/LP) compared to those without variants in genes that may cause or mimic
HCM (SARC-NEG) and those with variants of uncertain significance (SARC-VUS), adjusted for Age, Sex and Race. n = 436; P = 0.002. 1 HR =
Hazard Ratio, CI = Confidence Interval

Full model Genotype only

Characteristic HR1 95% CI1 p-value HR1 95% CI1 p-value
P/LP
N — — — —
Y 2.63 1.43, 4.86 0.002 2.62 1.42, 4.84 0.002
Race
White — —
Other 0.68 0.34, 1.37 0.3
Sex
F — —
M 1.15 0.72, 1.85 0.6

Curran et al. 2023 |

 Supplementary Table X. Cumulative hazard model excluding SARC-VUS. All cause mortality in individuals with hypertrophic
cardiomyopathy carrying pathogenic or likely pathogenic sarcomeric variants (SARC-P/LP) compared to those without variants in genes
that may cause or mimic HCM (SARC-NEG), adjusted for Age, Sex and Race. N = 395; P = 0.003. 1 HR = Hazard Ratio, CI = Confidence Interval

Full model Genotype only

Characteristic HR1 95% CI1 p-value HR1 95% CI1 p-value
P/LP
N — — — —
Y 2.66 1.43, 4.95 0.002 2.64 1.40, 4.98 0.003
Race
White — —
Other 0.68 0.30, 1.54 0.4
Sex
F — —
M 1.26 0.77, 2.06 0.4

Curran et al. 2023 |

 Dimensionality reduction and unsupervised clustering of clinical features
Participant features comprised demographic data, clinical characteristics, CMR and echocardiographic measurements, and
reported interventions and medicines (Supplementary Table XI). Missing values were inferred with the mice package for
R.63 Numerical features were converted to categorical variables by clustering groups of values into bins with a K-means
algorithm.64 All categorical variables were then transformed into binary variables with one-hot encoding.
Dimensionality reduction was performed on this collection of binary variables with UMAP (uniform manifold approxi-
mation and projection)65 using the following parameters: Dice metric, 25 components, 8 neighbouring sample points and
a minimum distance between points of 10−6 . Finally, unsupervised clustering was applied to the 25 resulting UMAP compo-
nents with a K-means algorithm assessed and optimised with a silhouette score, revealing three clusters. Genotype status
was found to be significantly associated with the clusters, using a 𝜒 2 test. A post-hoc exact Fisher test was then performed
to find cluster-specific enrichment. Adjustment for multiple testing was done with the Benjamini-Hochberg procedure,
P < 0.05. Cluster 1 was significantly enriched with genotype-negative (NEG) subjects while cluster 3 was associated with
genotype-positive (P/LP) and genotype-indeterminate (VUS) individuals. Feature importance from the initial set of partic-
ipant data was assessed by applying a Kruskal-Wallis test for numerical features and a 𝜒 2 test for categorical features.
Significant associations (adjusted with the Benjamini-Hochberg method, P < 0.05) were further tested for cluster-specific
enrichment: a Dunn test was used to test each pair of numerical features (Supplementary Fig. XIVA), while an exact Fisher
test looked for one-vs-rest differences in categorical features (Supplementary Fig. XIVB).
The clustering revealed features characterising each group: 1) older female participants with lower body surface area
(BSA), lower left ventricular (LV) volume, higher ejection fraction, hypertension, low activity score and on beta blockers
and diuretic medications; 2) male participants with higher BSA, higher LV mass, higher LV maximum wall thickness, hy-
pertension, moderate activity score and on medications for blood pressure (ACE/ARBs) and protective vascular (ASA/Clopi)
medications; 3) younger participants with a family history of hypertrophic cardiomyopathy (HCM), no clinical cardiovascular
risk factor and a high activity score.

Curran et al. 2023 |

 A
DBP (mmHg) LV max WT Ejection fraction Age Alcohol units/week

Stroke volume (ml) LVESV (ml) LV mass (g) SBP (mmHg) BSA (m2)

LVEDV (ml)

Supplementary Figure XI. Unsupervised clustering of clinical features and feature importance. a. Significant pairs of associations between identified clusters and
numerical features from the initial set of data. The line represents the interquartile range and median value. b. Significant one-vs-rest associations between clusters and
categorical features from the initial set of data grouped by significance level. The height of the curved bars illustrates the significance level (−𝑙𝑜𝑔10 𝑃 ). Only the significant
pairs are reported with the symbols: ∗ 𝑃 ≤ 0.05; ∗∗ 𝑃 ≤ 0.01; ∗∗∗ 𝑃 ≤ 0.001; ∗∗∗∗ 𝑃 ≤ 0.0001, 𝑛 = 436.

Curran et al. 2023 |

Supplementary Table XI. Participant clinical features. MRI, magnetic resonance imaging; HCM, hypertrophic cardiomyopathy; SCD, sudden cardiac death CCS,
Canadian Cardiovascular Society66 ; NYHA, New York Heart Association67 ; LV, left ventricular; RV, right ventricular; ACE, angiotensin-converting enzyme; ARB, angiotensin
receptor blockers; ASA, acetylsalicylic acid.

Measurements derived from MRI

Demographics Mean ± SD or n (%) and echocardiogram Mean ± SD or n (%)
Age at time of MRI scan (years) 57.3 (± 14.4) LV end-diastolic volume (mL) 137.6 (± 34.5)
Sex, male 310 (71.1%) LV end-systolic volume (mL) 37.3 (± 18.3)
Ethnicity LV stroke volume (mL) 100 (± 24.4)
non-Finnish European 353 (81%) LV ejection fraction 73.5 (± 8.5)
South Asian 52 (11.9%) LV mass (mL) 189 (± 67)
African 16 (3.7 %) LV maximum wall thickness (mm) 18.8 (± 4.5)
Others 14 (3.2 %) LV most affected segment
East Asian 1 (0.2 %) Anterior 43 (12.0%)
Inferior 8 (2.2%)
Clinical characteristics Lateral 11 (3.1%)
Septal 295 (82.6%)
Body surface area (m2 ) 2 (± 0.3)
LV most affected level
Diastolic blood pressure (mmHg) 76.2 (± 11.4)
Base 210 (59.3%)
Systolic blood pressure (mmHg) 133 (± 18.5)
Mid 78 (22.0%)
Pulse rate (bpm) 70.1 (± 13.7)
Apex 66 (18.6%)
Smoker 172 (39.4%)
Mitral regurgitation
Alcohol intake (units per week) 6.8 (± 12.3)
None 203 (54.9%)
Activity score
Minimal 124 (33.5%)
0 64 (14.7%)
Moderate 40 (10.8%)
1 68 (15.6%)
Severe 3 (0.8%)
2 252 (57.8%)
LV gadolinium
3 49 (11.2%)
None 53 (14.0%)
4 3 (0.7%)
Minimal 124 (32.7%)
Hypertension 175 (40.1%)
Moderate 133 (35.1%)
Diabetes mellitus 47 (10.8%)
Severe 69 (18.2%)
Coronary artery disease 46 (10.6%)
RV hypertrophy 45 (10.3%)
Myocardial infarction 21 (4.8%)
Coincident infarction 19 (4.4%)
Family history of HCM 85 (19.5%)
LV outflow tract obstruction 120 (27.5%)
Family history of SCD 70 (16.1%)
LV outflow peak velocity (m/s) 2.3 (± 0.7)
CCS Angina Grading Scale
0 57 (13.1%)
Interventions and medicines
I 277 (63.5%)
II 88 (20.2%) ACE inhibitors and ARBs 145 (33.3%)
III 13 (3.0%) ASA/clopidrogel 184 (42.2%)
IV 1 (0.2%) Beta blocker 210 (48.2%)
NYHA Classification of Heart Failure Diuretic 66 (15.1%)
No heart failure 48 (11.0%)
I 177 (40.6%)
II 176 (40.4%)
III 31 (7.1%)
IV 4 (0.9%)

Curran et al. 2023 |