Leveraging in-silico deep learning and computational analyses to predict the pathogenicity of ROBO4 variants of uncertain significance in aortic aneurysm and dissection patients

Abstract

Variants of uncertain significance (VUS) in genes implicated in thoracic aortic aneurysm (TAA) present clinical challenges due to ambiguous pathogenicity and low patient representation. This study investigates the pathogenic potential of missense VUS in the ROBO4 gene, previously associated with vascular integrity and ascending aortic aneurysm. Clinical and genetic data from five patients with heterozygous ROBO4 VUS and thoracic aortic aneurysms or dissections were analyzed. Computational tools including AlphaFold2, AlphaMissense, REVEL, PolyPhen-2, SIFT, FATHMM, MutationTaster2, GranthamMatrix, and PhastCons were utilized to predict pathogenicity and structural impacts. Patients exhibited varying severities of aortic pathology, from elective aneurysm repairs to extensive familial aneurysmal histories. Structural modeling revealed significant differences in residue positions and biochemical properties, particularly for extracellular domain variants affecting critical beta-sheet structures involved in vascular stability. Notably, patient-specific predictions aligned computational evidence with clinical severity, suggesting potential genotype-phenotype correlations. For example, a variant (Q44P) showed strong pathogenic predictions coinciding with severe familial presentations. These computational predictions, validated by clinical data, highlight a novel and efficient workflow for evaluating VUS pathogenicity, informing precision medicine, and guiding counseling for aortic degenerative diseases. Ultimately, we demonstrate the value of integrating computational modeling with clinical data to decipher the pathogenic significance of genetic variants in cardiovascular diseases.

Introduction

Aortic aneurysm disease is a life-threatening vascular condition associated with high morbidity and mortality. Although thoracic aortic aneurysms (TAA) occur only in 6–10 per every 100,000 people[1], they may develop unbeknownst to the patient until a dissection occurs, which results in a mortality of over 90% despite advances in surgical intervention [2]. Further complicating the complex etiology of TAAs is the large impact of genetic predisposition alongside environmental factors (e.g. hypertension, smoking), with approximately 20% of TAA cases showing familial clustering [1].

Genetic mutations play a pivotal role in the development of thoracic aortic aneurysms and dissections, with several well-known associated syndromes, such as Marfan syndrome (FBN1) and Loeys-Dietz syndrome (TGFBR1, TGFBR2, SMAD2, etc.)[3, 4]. However, many cases of familial aortic disease remain unexplained, highlighting the importance of investigating secondary genes involved in aortic wall integrity. A recent study at our institution investigating 1,034 patients with thoracic aortic disease revealed that up to 27% have variants of uncertain significance (VUS). Most primary disease-causing genes have lower VUS rates as compared to secondary genes [5]. Amongst these genes, ROBO4 is a novel contributor to vascular pathologies, including bicuspid aortic valve (BAV), aberrant aortic remodeling, and ascending aortic aneurysm [6]. In fact, BAV, a well-observed congenital defect present in 1–2% of the population, is frequently associated with ascending aortic aneurysm and is often inherited in an autosomal dominant manner with incomplete penetrance [7].

ROBO4, which encodes the roundabout guidance receptor 4, is a vascular-specific transmembrane protein that maintains endothelial integrity and vascular function. It inhibits angiogenesis and endothelial cell migration and is expressed predominantly in angiogenic endothelium, particularly the aorta [8–10]. Structurally, ROBO4 comprises three major domains: an extracellular domain responsible for ligand binding, adhesion, and endothelial barrier signaling[10–12], a transmembrane domain, and a cytoplasmic domain for intracellular signals and cytoskeletal regulation [11]. Unlike canonical TAA-associated genes such as TGFBR1/2, where pathogenic variants cluster within kinase or regulatory regions, ROBO4 variants have been reported across all domains [6]. This broad distribution suggests that disruption of any structural component may compromise ROBO4’s vascular-stabilizing function and lead to large vessel pathologies.

We investigate five TAA patients harboring ROBO4 missense variants from a single base change, identified through genetic testing at a quaternary academic aortic center. Many ROBO4 missense variants remain difficult to classify due to limited patient data, scarce molecular validation, and rare segregation analyses. To address this, we developed a workflow integrating clinical data with in-silico prediction tools to assess the structural and functional impact of these variants. By analyzing evolutionary conservation, molecular changes, and predicted protein stability, we elucidate potential genotype-phenotype correlations and provides a scalable model for interpreting VUS across genes implicated in thoracic aortic disease.

Methods

Patient selection and clinical presentation

This study was approved by the Institutional Review Board at Yale New Haven Health (IRB #: 200002755). Written informed consent was obtained from all patients. All patients with positive ROBO4 VUS genetic testing and relevant aortic histories were included. Clinical data were collected from electronic health records. Academic cardiologists obtained aortic measurements pre-repair and post-repair through echocardiography and long-term cardiac assessments (2 years ± 3 months) through computed tomography angiography (CTA).

Genetic analysis

Genetic testing results were provided through Yale DNA Diagnostics or Invitae. ​​DNA is extracted from patient biosamples and quality-assessed. Whole-genome sequencing is performed on the Illumina platform, and genetic variants are identified using the Broad Institute’s Genomic Analysis ToolKit best practices. Candidate genes analyzed for TAAs were ACTA2, ADAMTS10, BGN, CBS, COL1A1, COL1A2, COL3A1, COL4A5, COL5A1, COL5A2, EFEMP2, ELN, FBN1, FBN2, FLNA, FOXE3, GATA4, GATA5, GATA6, HCN4, HEY2, LOX, MAT2A, MED12, MFAP5, MIB1, MYH11, MYLK, NOTCH1, PKD1, PKD2, PRKG1, ROBO4, SKI, SLC2A10, SMAD2, SMAD3, SMAD4, SMAD6, SOX18, TCF7L2, TGFB2, TGFB3, TGFBR1, TGFBR2, TGFBR3, and THSD4. Variants are annotated with allele and population frequencies. An ABMGG-certified clinical molecular geneticist interpreted the results.

Computational pathogenicity predictions

To assess the potential impact of ROBO4 missense variants on protein function and structure, the following computational scores were employed. All in-silico modeling and computations utilized Homo sapiens (human) genome assembly GRCh37 (hg19) to match patient genetic testing.

1. Alpha Missense [13] is a deep learning model that predicts the likelihood of missense variants being pathogenic based on structural information. Scores < 0.34 indicate a ‘likely benign’ variant, >0.56 are ‘likely pathogenic,’ and 0.34–0.56 were classified as ‘ambiguous.‘[14].

2. REVEL [15] is an ensemble-based pathogenicity prediction tool that integrates multiple individual variant effect predictors to estimate missense variant pathogenicity. Scores < 0.4 indicate a likely benign variant, >0.6 indicate a likely pathogenic variant, and intermediate values (0.4–0.6) are considered ambiguous.

3. SIFT [16] predicts the functional impact of an amino acid substitution based on evolutionary conservation at the specific position in a protein family. A substitution at a highly conserved amino acid is more likely to be deleterious. SIFT classifies variants as “tolerated” or “deleterious,” where < 0.05 is considered deleterious.

4. PolyPhen-2[17] predicts the potential impact of amino acid changes on protein function by analyzing both structural and sequence-based features, using both Mendelian diseases and general variants as a basis. PolyPhen-2 helps classify variants as benign (0.00–0.15.00.15), possibly damaging (0.15–0.85), or probably damaging to protein function (0.85–1.00.85.00).

5. FATHMM [18] predicts the pathogenicity of amino acid substitutions through comparing the wild-type and mutant sequences in conserved homologous domains. The weighted method emphasizes the likelihood of variant being pathogenic based on known human variants. A score <−1.5 is considered pathogenic, and >−1.5 indicate a neutral variant.

6. MutationTaster2[19] predicts SNV pathogenicity by integrating publicly available databases, analyzing variants around splicing and regulatory sites, and physicochemical properties. It outputs a binary classification as “polymorphism” or “disease-causing.”

7. GranthamMatrix [20, 21] is a scoring system used to evaluate the impact of amino acid substitutions based on side chain atomic composition, polarity, and volume. The amino acid substitution is categorized as “conservative” (5–60) indicating minimal protein functional changes, “moderately non-conservative” (60–100) indicating moderate impact, and “radical” (>100) indicating severe impact on the protein’s function.

8. PhastCons [22] calculates the probability that each nucleotide or amino acid is part of a evolutionarily conserved element across multiple orthologous alignments. High scores (close to 1) indicate strong evolutionary conservation and functional importance. Low scores (close to 0) suggest low conservation and possibly low importance.

The Ensemble Variant Effect Predictor [23] was used to calculate AlphaMissense, REVEL, SIFT, and PolyPhen-2. FATHMM, MutationTaster2, GranthamMatrix, and PhastCons was calculated from the web application (https://www.mutationtaster.org/) compared to the wild-type ROBO4 (UniProt Q8WZ75) [24]. The Ensembl ID for the protein transcript used for all computational pathogenicity scores was ENST00000306534.8[25].

In-silico folding models for ROBO4 variants using AlphaFold2

Protein folding models were generated for each variant using ColabFold, a tool that combines the fast homology search of MMseqs2 with the deep-learning protein structure prediction capabilities of AlphaFold2 [26–28]. Each generated variant had its associated confidence metric using the predicted local distance difference test (pLDDT). The DALI server was used to align generated protein structures and verify AlphaFold’s generated variants [29]. Higher Dali z-scores indicates structural similarity to wild-type [30]. We used UCSF ChimeraX [31] to calculate the following physical parameters of ColabFold variant structure deviations compared to wild-type (UniProt Q8WZ75).

1. Root Mean Square Deviation (RMSD): RMSD was calculated to measure the structural deviations between the wild-type and mutant protein models:

   1. All-atom RMSD: The RMSD was also computed for all 1007 atom pairs between the wild-type and mutant structures for comprehensive view of all deviations.
   2. Pruned RMSD: The RMSD was calculated by excluding regions with high flexibility. Amino acids preserved by RMSD was also reported, indicating they were most conserved for folding and potentially important to protein structure.

2. Distance Between Variant and Wild-Type: The distance between the missense residue and its corresponding wild-type was measured to assess the local structural displacement.

3. Solvent Accessible Surface Area (SASA): We calculated changes in the surface exposure of the variant compared to the wild type, which affects protein stability and interactions.

4. Contacts and Hydrogen Bonds: Changes in the number of contacts (interatomic distances) between the substitution and neighboring residues were analyzed. Hydrogen bonds were measured to determine if the variant altered the protein’s secondary or tertiary structure.

Results

Genetics

Five patients with a history of aortic aneurysms or dissections and ROBO4 VUS were identified from our institution’s genetic database. These patients presented with the following ROBO4variants: p.K646R (c.1937 A > G), p.S271G (c.811 A > G), p.P296L (c.887 C > T), p.L745M (c.2233 C > A), and p.Q44P (c.131 A > C). These missense variants were classified as rare, as determined by their population frequencies in the Genome Aggregation Database below 0.01[32]. All variants were heterozygous and autosomal dominant. Patient 2 with p.S271G (c.811 A > G) had a comorbid VUS in TGFBR1 p.M1T (2T > C) and PKD1 p.C2229Y (6686G > A).

Clinical patient presentation

The clinical presentations, genetic attributes, and echocardiographic findings associated with each patient’s surgical course are detailed in Table 1.

Table 1.

Genotypic and phenotypic presentation of AscAA patients harboring ROBO4 VUS pre- and post-repair

Patient ID	1		2		3		4		5*
ROBO4 VUS	p.K646R (c.1937 A > G)		p.S271G (c.811 A > G)		p.P296L (c.887 C > T)		p.L745M (c.2233 C > A)		p.Q44P (c.131 A > C)
Frequency	21/206,748		0		12/247,912		1/271,080		6/160,564
Sex	Male		Male		Female		Male		Female
Age	68		70		59		61		51
	Cardiovascular status prior to TAA repair
BAV	Bicuspid		Trileaflet		Trileaflet		Trileaflet		Trileaflet
AoV Calcification	Yes		Yes		No		No		No
Ao stenosis	Yes		No		No		No		No
Ao regurgitation	Yes		Yes		No		Yes (AoV prolapse)		No
LV hypertrophy	Concentric		Concentric		No		Concentric		Concentric
Hypercholesterolemia	Yes		Yes		Yes		Yes		No
Hypertension	Yes		Yes		Yes		Yes		Yes
Hx of AKI/CKD	AKI		No		No		No		No
	Aortic and cardiac function parameters pre- and post-repair on CTA and echocardiography
	Pre	Post	Pre	Post	Pre	Post	Pre	Post	Pre	Post
AoV PK vel (m/s)	2.4	1.8	1.1	1.2	1.0	1.4	1.4	2.1	1.5	-
Ao Mn Grad (mmHg)	13	4	2	3	2	3	2	10	-	-
AVA (cm2)	1.7	1.8	3.7	2.4	1.5	1.8	3.5	2.3	-	-
DI	0.49	0.7	0.69	0.69	0.84	0.78	0.88	0.47	-	-
Ao Asc diameter (cm)	4.8	3.1	4.4	3.8	3.8	2.5	5.8	3.6	4.1	-
SOV diameter (cm)	4.4	3.9	5.1	3.3	3.6	3.2	4.7	3.9	4.1	-
LVEF (%)	62	55	70	60	67	57	69	65	67	-
RVSP (mmHg)	28	23	17	-	22	23	-	-	40	-
Bioprosthetic valve		Yes		Yes		No		Yes		-

Empty spaces with “-” indicate that the data is not available

BAV Bicuspid aortic valve, Ao Aortic, AoV Aortic valve, LV Left ventricular, AKI Acute kidney injury, CKD Chronic kidney disease, AoV PK vel Aortic valve peak velocity, Ao Mn Grad Aortic mean gradient, AVA Aortic valve area, DI Dimensionless index, Ao Asc Ascending aorta, SOV Sinus of Valsalva, LVEF: Left ventricular ejection fraction, RVSP Right ventricular systolic pressure

*Patient 5 has not undergone TAA repair. She is a candidate for prophylactic TAA repair and currently has a stable ascending aortic aneurysm

Patient 1 was a 68-year-old male with a history of hypertension, hypercholesterolemia, right bundle branch block, and acute kidney injury. They also presented with a bicuspid aortic valve associated with ROBO4variants[33, 34], severe aortic regurgitation, aortic stenosis, and concentric left ventricular hypertrophy. Pre-repair CTA revealed a significantly dilated ascending aorta (4.8 cm), and dilated sinus of Valsalva (SOV) on echocardiography (4.4 cm). It also reported aortic valve area (AVA) of 1.7 cm² and mean gradient (Mn Grad) of 13 mmHg, consistent with mild aortic stenosis, along with a dimensionless index (DI) of 0.49 consistent with aortic valve dysfunction [35]. Despite regurgitation, left ventricular ejection fraction remained normal. The patient underwent ascending aortic aneurysm repair, aortic root replacement, and thoracic hemiarch replacement. Follow-up CTA two years later confirmed post-surgical stability, with preserved ejection fraction.

Patient 2 was a 70-year-old male with a history of longstanding hypertension for nearly 40 years, hypercholesterolemia, and superior mesenteric artery and renal artery aneurysm, suggesting systemic aneurysmal disease. He presented with moderate aortic regurgitation, aortic valve calcification, and concentric left ventricular hypertrophy. On pre-repair CTA, the ascending aorta diameter was 4.4 cm. Pre-repair echocardiography revealed a significantly dilated SOV (5.1 cm) and no significant impairment of cardiac function. The patient underwent aortic root and transverse hemiarch replacement. Post-repair echocardiography showed no significant differences in cardiac function [36]. A follow-up CTA two years later confirmed stable post-operative changes, an unchanged, intact aortic graft, and stable aortic arch measurements (3.8 cm).

Patient 3 was a 59-year-old female with a history of hypercholesterolemia and hypertension, but did not present with any of aortic regurgitation, stenosis, or valvular insufficiency. Pre-replacement CTA indicated dilation of the ascending aorta (3.8 cm). The echocardiography also revealed a dilated SOV (3.6 cm), and otherwise normal valvular and left ventricular function. Due to the patient’s family history of cardiac arrest and insistence for repair, the patient underwent prophylactic ascending aortic replacement for TAA without rupture. They also received an aortic wrap due to extreme tissue fragility reported during operation. Echocardiography demonstrated intact valvular function. Follow-up CTA two years later showed a stable aortic diameter at 2.5 cm.

Patient 4 was a 61-year-old male with history of sleep apnea, hypercholesterolemia, and hypertension. Pre-repair CTA showed ascending aorta measuring 5.9 cm. Preoperative echocardiography showed aortic valve prolapse and severe aortic regurgitation, along with left ventricular hypertrophy. The patient underwent aortic root and transverse hemiarch replacement with concurrent CABG. Post-operative echocardiography showed intact bioprosthetic aortic valve function [37]. Follow-up CTA imaging two years later showed patent grafts and stable aortic arch.

Patient 5 was a 51-year-old female with a past medical history of hypertension, Parkinson’s, and TAA, with a strong family history of aortic aneurysms and dissections. Her CTA on initial presentation revealed a dilated ascending aorta (4.0 cm). Her echocardiography also indicated a dilated sinus of Valsalva (4.1 cm) and left ventricular hypertrophy. These findings have remained stable since, with the ascending aorta measuring 4.1 cm on the most recent CTA 3.5 years later.

Patient 5’s father, with no variants and unknown ROBO4, underwent an emergent aortic dissection repair at age 64 and had a splenic artery aneurysm. Her mother, with a ZNF469 VUS and unknown ROBO4, had an elective ascending aortic replacement for a TAA at age 79. Brother A, with “negative” genetic testing, underwent elective aortic aneurysm replacement surgery for a 4.9 cm ascending aortic root aneurysm and aortic insufficiency. Brother B, with no genetic testing, had no aortic history. Her children (19, 22, and 25) currently have normal aortic dimensions on imaging (Table 2).

Table 2.

Amino-acid level analyses of ROBO4 VUS pathogenicity using computational algorithms and in-silico protein folding models

Patient ID	1	2	3	4	5
Variant	p.K646R (c.1937 A > G)	p.S271G (c.811 A > G)	p.P296L (c.887 C > T)	p.L745M (c.2233 C > A)	p.Q44P (c.131 A > C)
Missense Location	Cytoplasmic	Extracellular	Extracellular	Cytoplasmic	Extracellular
	Computational pathogenicity predictions
AlphaMissense	0.105	0.180	0.098	0.071	0.127
REVEL	0.271	0.122	0.103	0.057	0.110
SIFT	0.04	0.04	0	0.18	0.18
PolyPhen-2	0.545	0.917	0.265	0	0.675
FATHMM	−0.17	0.62	−0.14	−0.08	−0.33
MutationTaster2	Polymorphism	Polymorphism	Polymorphism	Polymorphism	Disease-causing
GranthamMatrix	26	56	98	15	76
PhastCons**	0.979, 0.976, 0.996	0.585, 0.619, 0.639	0.002, 0.112, 0.121	0.002, 0.002, 0	0.945, 0.922, 0.890
	In-silico variant folding models
RMSDall (Å)	37.281	31.642	46.123	36.460	40.529
RMSDpruned (Å)	1.032	0.596	0.677	0.577	0.712
Preserved AAs	209	209	217	247	201
AAwt-AAmut gap (Å)	8.316	3.309	2.709	74.899	1.488
SASAAA (Å2)	350.85	187.07	291.34	306.29	246.23
Hydrogen bonds	615	640	653	652	642
Model DALI score	22.3	25.3	24.7	25.8	19.7

The in-silico folding model predictions are compared to the wild-type ROBO4 model. Wild-type ROBO4 model counted 668 hydrogen bonds

RMSD Root mean square deviation from wild-type, AA: Amino acid, wt Wild-type, mut Mutant, SASA Solvent-available surface area

*Low sequence representation at these amino acid locations, and thus SIFT predictions are made at low confidence

**The three PhastCons scores represent the conservation of the nucleobase preceding the variant, the variant nucleobase itself, and the nucleobase immediately following the variant

Contextualizing pathogenicity through in-silico folding of missense variants

The structure of ROBO4 provides essential context for understanding how its variants may alter vascular stability. ROBO4 is a 1007-amino acid, single-pass transmembrane receptor consisting of an extracellular domain (residues 1–478), a transmembrane helix (479–501), and a cytoplasmic tail (502–1007) [11]. The extracellular region contains four domains with two distinct motifs: two immunoglobulin-like domains Ig1 (32–131) and Ig2 (137–224), and two fibronectin type-3 domains FnIII-1 (248–345) and FnIII-2 (347–442) [24]. Each form β-sandwich folds [8, 38, 39] stabilized by disulfide bonds and N-linked glycosylation sites. The four domains mediate ligand interactions and facilitate cell adhesion, primarily through Slit proteins that mediate migration [12] and endothelial adhesion molecules, overall maintaining vascular barrier integrity and suppresses aberrant angiogenesis [9, 11].

The cytoplasmic domain contains two ROBO-family conserved CC0 and CC2 motifs that recruit cytoskeletal and kinase-associated adaptors, enabling ROBO4 to regulate downstream endothelial migration and vascular permeability [40]. The transmembrane domain anchors ROBO4 within the plasma membrane and harbors a metalloproteinase cleavage site for converting the ectodomain to a soluble, circulating ROBO4 that exerts inhibition on VEGF and FGF-induced angiogenesis [9].

All five In-silico structures demonstrated high confidence folding across conserved domains, particularly within the β-sheet-rich Ig and FnIII domains that comprise the extracellular portion of ROBO4 (Fig. 1A and B). On qualitative inspection, variants showed areas of high alignment, such as the extracellular β-sheets in Fig. 1C, while others deviate from the wild-type, such as the intracellular alpha helices in Fig. 1D. The conservation of extracellular domains and relative misalignments in other domains is consistent with the pLDDT graphs in Fig. 1A.

Fig. 1.

In-silico folding models reveal areas of high confidence and alignment versus significant spatial deviation. a Predicted local distance difference test (pLDDT) confidence scores visualized across highest ranked predicted structure for the S271G VUS in patient 2. b Predicted per-residue LDDT confidence scores for the top five ranked structural models for the S271G VUS in patient 2 plotted against sequence position, where higher values indicate greater local reliability in the predicted structure. Note the higher average pLDDT in extracellular residues versus cytoplasmic residues. c Example area of high alignment in predicted structure (beta sheets). Yellow is wild-type, blue is variant. d Example area of high deviation in predicted structure (alpha helices). Yellow is wild-type, blue is variant

Rationalizing pathogenicity through quantitative computation and qualitative structure analysis

Table 1 shows the results of computational pathogenicity prediction scores and structural alignment between wild-type and variants. While being the gold standards of modern pathogenicity prediction, AlphaMissense and REVEL fail to predict significant harm in each patient’s missense variants. Further exploration was warranted through other computational scores and protein modeling.

Patient 1’s cytoplasmic K646R variant was classified as possibly damaging by PolyPhen-2. PhastCons suggests the amino acid is the highest conserved of all five variant locations, also supported by SIFT. The lysine to arginine substitution replaces the side chains from an amine to a guanidine, which are both positively charged at physiologic pH. While a biochemically similar replacement as suggested by GranthamMatrix, it also has sterically and electrochemically different impacts (ex. cation-π interactions) [41]. Furthermore, protein folding models shows the distance between the variant and wild-type residue is 8.316 Å – a structurally significant difference (for comparison, a guideline for amino acid size is 3.5 Å)[42]. Prior literature in protein structure comparison shows similar judgment for deleterious structure changes [43–45].

Patient 2’s extracellular S271G variant was predicted more confidently to be “damaging” by PolyPhen-2 (0.917) and FATHMM (0.62). PhastCons and SIFT predicts the amino acid to be of moderate evolutionary conservation. The serine to glycine substitution replaces the flexible glycine with a bulkier hydroxymethyl group capable of hydrogen bonds, and GranthamMatrix suggests a moderate biochemical and steric change. While structural analysis was unremarkable, the variant lies in the FnIII-1 domain: essential for ligand binding, adhesion, and signaling.

Patient 3’s extracellular P296L variant also lies in the FnIII-1 domain. PhastCons suggests low conservation at position 296. PolyPhen-2, FATHMM, MutationTaster2 suggests a benign variant, although SIFT disagrees. GranthamMatrix suggests a large biochemical change, which is consistent with replacing a secondary amine with restricted rotational freedom with a free isobutyl side chain. However, because the variant occurs in a low conservation region, neither the computational scores nor structural parameters indicate a highly deleterious substitution.

Patient 4’s intracellular L745M variant occurs in a region of low evolutionary conservation as suggested by PhastCons and prior literature. Computational scores suggest low pathogenicity. The protein structure modeling suggests a high deviation of amino acid location between wild-type and variant after alignment (AA_wt-AA_mut gap = 74.9 Å). It also has the highest RMSD_all of all five variants, suggesting that the overall protein’s structure is the most perturbed across the entire primary transcript. However, because the substitution occurs intracellularly in a low conservation region, exact variant pathogenicity is difficult to ascertain.

Patient 5’s extracellular Q44P variant lies in the first immunoglobulin-like domain, which also responsible for ligand binding for downstream signaling. PolyPhen-2 suggests that the substitution is possibly damaging (0.675), which is consistent with the PhastCons score labeling the surrounding region as highly conserved. The flexible and bioactive amide of the glutamine is replaced with a restricted secondary amine of proline, and GranthamMatrix labels the substitution as a large impact. Patient 5’s missense variant was also the only one to be labeled as “Disease-causing” by MutationTaster. No significant deviations were noted on structural analysis.

Discussion

Developing a better understanding of how VUS convert to thoracic aortic disease is key to dissecting a key population of patients who may lack representation and medical certainty. As our prior study has shown, the picture is further complicated by a significant difference in the proportions of known pathogenic variants versus VUS in primary disease-causing genes versus secondary genes [5].

Critically, we can synthesize each patient’s computational pathogenicity scores and their variant structural analyses to reason through the patient’s clinical presentations. Patient 1’s K646R variant was shown to be pathogenic by substituting two electrochemically and sterically different amino acids in a region of high evolutionary conservation, resulting in its significant structural aberrations. Thus, the substitution could lead to the variant’s poor ability to inhibit aberrant large-vessel angiogenesis, leading to patient 1’s severe aortic dilation (4.8 cm), BAV, and aortopathy.

Patient 2’s presentation is difficult to attribute independently between their ROBO4 S271G variant and the concurrent comorbidity associated with their TGFBR1 start codon variant. However, given that most pathogenic TGFBR1 variants associated with Loeys-Dietz are clustered around its serine-threonine kinase domain, the patient’s aortopathy and concurrent systemic aneurysmal disease may be highly attributable to ROBO4. This conclusion is further supported by the patient’s pathogenic computational scores and the location of the variant in the important FnIII-1 domain.

Patient 3’s clinical presentation was arguably the most benign of the five: the patient showed mild, asymptomatic, and stable dilatation of the aorta (3.8 cm) and no aortopathy such as calcification, stenosis, or regurgitation. In fact, upon expert evaluation, aortic replacement was classified as non-urgent, and the procedure was only conducted “prophylactically” at patient’s request However, the provider noted “tissue fragility” during the operation that warranted an aortic wrap. Computational scores suggest that the patient’s variant is not a harmful change, although the amino acid substitution is biochemically significant and occurs in the important FnIII-1 domain. Neither computational scores nor in-silico modeling indicated significant pathogenicity, which may have led to patient 3’s relatively benign disease course.

Patient 4’s had the largest dilatation of the aorta at 5.8 cm with severe regurgitation. Computational scores failed to explain the L745M variant’s pathogenicity. However, structural modeling showed the most significant deviations from the wild-type of all five variants, possibly suggesting that the cytoplasmic domain’s amino acid substitution in a low conservation area may not be locally important, but could have broader impacts in the protein’s structure. Furthermore, while the cytoplasmic domain is non-catalytic, rescue experiments have shown it to be involved in VEGF-based signaling to prevent aberrant angiogenesis [11, 38] and Slit-mediated endothelial migration that maintains aortic wall integrity [46].

Lastly, patient 5’s clinical presentation also has the support of the most suggestive family history of all five patients. The ROBO4 variant occurs in a highly conserved portion of the extracellular Ig-1 domain, and the computational scores suggest potentially significant pathogenicity. Although there is little structural deviation (Fig. 2A and B), the substitution occurs in the highly conservative Ig domain β-sheets. Given the lack of other variants in the patient or their family (beyond the ZNF469 VUS absent in patient 5), we can reasonably attribute the patient’s aortopathy to ROBO4.

Fig. 2.

Rationalizing patient aortic instability through differences in sequence-level biochemical structure. a Although patient 5’s Q44P variant places the amino acid not too far from wild-type (1.488 A), it exists in a highly conserved region of the protein. Yellow is wild-type, blue is variant. b The variant lies near one of the IgG beta sheets of the ROBO4 protein. Yellow is wild-type, blue is variant. c Categorizing patient 1–5’s variants based on their locations in the protein sequence

Furthermore, two mechanisms are consistent with the biochemistry of patient 5’s ROBO4 missense variant and their autosomal dominant inheritance pattern (as suggested by their heterozygosity and family pedigree). First, there is a significant reduction of functional ROBO4 in the maintaining the structural integrity of the aortic walls through VEGF-binding to the Ig domains [47]. Second, the extracellular domain of ROBO4, which can be cleaved and circulate in soluble form (sROBO4) to inhibit Slit-mediated migration and angiogenesis, is no longer functionally active [10]. These mechanisms are supported by both literature and patient 5’s case, further lending proof-of-concept to our workflow applied in real clinical practice.

Figure 2C summarizes the findings of our study based on the case studies of the five variants. We present patients with ROBO4 VUS and their clinical presentations and attempt to molecularly rationalize the evidence-driven mechanisms of their associated pathologies through in-silico modeling and computational pathogenicity predictions. We discerned justifiable clinical outcomes that highlight the pathophysiology of aortic aneurysms and further suggested the polygenic nature/multifactorial inheritance. Especially considering comorbid VUS, we highlight the importance of quantitative and qualitative evaluations of pathogenic SNPs for personalized care.

This study has limitations. With a small patient cohort (n = 5) and no control group, it is difficult to attribute findings exclusively to ROBO4 or to draw statistically significant conclusions. Despite our best efforts to collect familial data, there was limited family data for segregation analyses. Additionally, the analysis relied solely on modeling without experimental validation. These findings underscore that current methods should be applied cautiously and refined before clinical translation to ensure patient safety.

Nevertheless, this workflow proved to be a powerful, efficient, and cost-effective tool in cardiac surgery bioinformatics. Our in-silico methods and computational analyses yielded crucial biochemical insights into the pathogenic potential of ROBO4 VUS in ascending aortic aneurysm in a cost-effective manner, showing potential to inform clinical decision-making, advance precision medicine interventions, and even improve genetic counseling through widening the tools available for informing patients. Because these individual variants have very little clinical representation, yet together still represent a significant portion of patients, our method of studying these variants is but one necessary step towards delivering increasingly effective, precise, and tailored care.

Authors’ contributions

CL: Conceptualization; Formal Analysis; Investigation; Methodology; Software; Visualization; Writing – Original Draft.IH: Conceptualization; Data Curation; Formal Analysis; Investigation; Project Administration; Writing – Original Draft.MC, EE, HA: Conceptualization; Data Curation; Investigation; Methodology.JL: Formal Analysis; Investigation; Methodology; Software; Visualization.ASF, SV, WS: Formal Analysis; Investigation.RA, PV: Conceptualization; Data Curation; Investigation.All authors: Writing – Review & Editing.

Funding

This work had no funding to declare.

Data availability

There are restrictions to the availability of the patient datasets used in this study due to privacy considerations and institutional ownership by Yale New Haven Health, and thus data is not publicly available for these reasons. All relevant, de-identified data of the patients presented in this manuscript is provided in the tables, including patient demographics, comorbidities, and ROBO4 variant information.

Declarations

Ethics approval and consent to participate

The Yale University Institutional Review Board review board gave approval for this study under protocol number 200002755. Consent for study participation was obtained prior to this study, as indicated in the Methods section.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.