📄 health informatics

Topological Deep Learning Identifies Polygenic Variant Clusters Across Familial Multimorbid Disorders

The paper introduces PolyCLIP-T, a topology-guided multimodal framework that leverages whole-genome sequencing and persistent homology to identify stable clusters of polygenic variants across familial multimorbid disorders, effectively overcoming the limitations of traditional rule-based pipelines in detecting non-coding, structural, and low-penetrance genetic drivers.

Original authors: Vomo-Donfack, K. L., Bousquet, G., Falgarone, G., Ginot, G., Morilla, I.

Published 2026-06-09

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Vomo-Donfack, K. L., Bousquet, G., Falgarone, G., Ginot, G., Morilla, I.

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

The Big Problem: Finding a Needle in a Haystack

Imagine your DNA is a massive library containing 3.2 billion books (base pairs). When a doctor sequences a patient's genome, they get a list of about 4 to 5 million "typos" or differences compared to a standard reference book.

The current way doctors find the "bad" typos that cause disease is like using a strict rulebook (called ACMG/AMP guidelines). This rulebook is great at finding the "smoking guns"—obvious, high-impact errors that guarantee a disease (like a single broken engine part in a car). However, it often misses the "sneaky" problems:

Typos in the footnotes or margins (non-coding DNA) that don't look important on their own.
A collection of tiny, harmless-looking scratches that, when added together, ruin the car's performance (polygenic risk).
Errors that only show up when combined with other specific errors in the family.

Because the rulebook is so strict, it leaves doctors with a massive list of "suspects" (Variants of Uncertain Significance) and no clear way to figure out which ones actually matter for a specific family.

The New Solution: PolyCLIP-T (The "Topological Detective")

The authors introduce a new tool called PolyCLIP-T. Instead of looking at each typo one by one and asking, "Is this bad?", they ask, "Do these typos hang out together in a specific neighborhood?"

Think of the genome not as a flat list, but as a giant, multi-dimensional city map.

The Old Way: You look at one house (variant) and check if it has a "Bad" sign on the door. If the sign isn't there, you ignore the house.
The PolyCLIP-T Way: You look at the whole city. You notice that certain houses, even if they don't have "Bad" signs, are all clustered together in a specific, strange-looking neighborhood. If you see a group of houses that look similar and are all occupied by people with the same illness, that neighborhood is suspicious, even if no single house looks dangerous on its own.

How It Works: Three Simple Steps

1. The Translator (Contrastive Learning)
First, the tool translates DNA into two different languages simultaneously:

Language A: The raw DNA sequence (the letters A, C, T, G).
Language B: The biological meaning (what the DNA actually does, like which genes it turns on or off).

It uses a "dual-encoder" system (like a translator who speaks both languages fluently) to map these two languages onto the same chart. It learns that a specific DNA sequence and its biological effect should sit next to each other on the map. If a DNA change causes a big shift in the biological meaning, it moves far away on the map.

2. The Neighborhood Watch (Topology)
Once all the DNA variants are mapped onto this chart, the tool uses a mathematical technique called Persistent Homology.

Imagine dropping a net over the city map. As you pull the net tighter, you see which houses stick together.
Some houses stick together only when the net is very tight (noise).
Other houses stick together even when the net is loose. These are stable clusters.
PolyCLIP-T looks for these stable clusters. If a group of variants stays together across different family members who are sick, the tool flags them as a "coherent group." It's like finding a gang of suspects who always hang out together, even if you can't prove the crime of any single one of them yet.

3. The Filter (The Funnel)
The tool takes millions of raw variants and filters them down. Instead of just looking for rare, high-impact errors, it looks for groups that are:

Topologically stable: They form a tight, persistent cluster.
Biologically coherent: They share similar functions (e.g., they all affect immune response or DNA repair).
Family-relevant: They appear in the sick relatives but not the healthy ones.

What They Found

The researchers tested this on six families where members suffered from a mix of cancer, autoimmune diseases, and heart problems.

The Result: The tool successfully narrowed down millions of variants to a tiny, manageable list of about 100 candidates per family.
The Discovery: It found "sneaky" variants that the old rulebook missed. These included:
- Non-coding variants: Typos in the "footnotes" of the DNA that regulate genes.
- Low-penetrance variants: Small errors that aren't dangerous alone but are dangerous when combined with others.
- Cross-disease links: It found that the same groups of genes were involved in cancer, heart disease, and autoimmune issues simultaneously. This suggests these diseases might share a common "root cause" in the family's DNA, rather than being separate, unrelated problems.

A Real-World Example: Family F6

The researchers tested the tool on a "blind" family (Family F6) they hadn't seen before.

The Old Way: Found a known cancer gene (BRCA1). This explained the cancer, but not why the family also had Crohn's disease and strokes.
The PolyCLIP-T Way: Found a new cluster of variants involving genes like ERBB3, HLA-A, and LPA.
- These genes were linked to cancer, immune function, and heart health.
- Individually, these variants were considered "uncertain" by standard rules.
- Together, they formed a stable topological cluster that explained the family's complex mix of diseases.

The Bottom Line

PolyCLIP-T doesn't replace the doctor's rulebook; it acts as a smart filter that sits before the rulebook. It helps doctors stop staring at a million suspects and instead focus on the specific "neighborhoods" of DNA that are likely causing the trouble in families with complex, multi-disease histories.

It turns the question from "Which single letter is broken?" to "Which group of letters is acting strangely together?" This is a crucial step for understanding diseases that are caused by many small factors working in combination, rather than one big mistake.

Technical Summary: Topological Deep Learning Identifies Polygenic Variant Clusters Across Familial Multimorbid Disorders

Problem Statement

Whole-genome sequencing (WGS) offers comprehensive coverage of coding, non-coding, and structural variations, yet its clinical utility is constrained by an "interpretation bottleneck." Current bioinformatic pipelines, largely anchored in American College of Medical Genetics and Genomics/Association for Molecular Pathology (ACMG/AMP) criteria, excel at identifying highly penetrant Mendelian alleles but frequently fail to detect variants of low-to-moderate penetrance, non-coding alterations, and germline–somatic interactions. This rule-based, categorical logic is ill-suited for the "genome-wide data deluge" and for identifying variants that contribute to disease through quantitative, combinatorial mechanisms (oligogenic or polygenic models). In familial disorders characterized by multimorbidity (e.g., co-occurring cancer, autoimmune, and cardiovascular diseases), a substantial proportion of contributing variants may not meet classical pathogenicity thresholds (e.g., insufficient rarity or modest effect sizes), leading to their dismissal as Variants of Uncertain Significance (VUS). The authors argue that the field requires a shift from single-variant classification to the selection of coherent sets of variants that collectively point to a plausible disease mechanism.

Methodology: The PolyCLIP-T Framework

The authors introduce PolyCLIP-T, a topology-guided multimodal framework designed to transform variant selection from a classification problem into a geometric discovery task. The pipeline consists of three primary stages:

1. Unified Multimodal Representation and Contrastive Learning

PolyCLIP-T learns multimodal representations of genomic variants by aligning sequence, biological, and text-derived information within a shared latent space.

Dual-Encoder Architecture: Inspired by the CLIP framework, the model employs two encoders:
- Sequence Encoder: Uses a fine-tuned DNABERT-2 transformer to encode reference and alternate DNA sequences (±100 bp, extended to ±500 bp for splice-aware contexts) into sequence embeddings ( $h_{seq}$ ).
- Biological Encoder: A multi-layer perceptron (MLP) processes tabular annotation vectors, including functional consequences (VEP, CADD, SpliceAI), conservation metrics (phyloP, GERP++), population genetics (gnomAD), structural context (CNV, LOH), and pathway annotations.
Contrastive Learning: The model is trained using a symmetric InfoNCE (NT-Xent) loss to align sequence embeddings with biological feature embeddings. Crucially, it utilizes ontology-based semantic similarity (e.g., Gene Ontology, pathways) as a soft-label supervision strategy rather than relying on scarce and biased hard pathogenicity labels. This allows the model to learn functional similarity and disease-related relationships between variants.
Delta Embeddings: The displacement between reference and alternate embeddings ( $\Delta z_i = z_{alt} - z_{ref}$ ) in the shared latent space quantifies the molecular perturbation induced by each variant.

2. Topological Data Analysis (TDA) and Regularization

To preserve the intrinsic geometric structure of the variant manifold and prevent geometric collapse common in contrastive learning, the authors introduce Topological Diffusion Regularization (TDR).

TDR: A teacher-student framework where a teacher network (an exponential moving average of student parameters) maintains a stable reference topology. The loss function minimizes the discrepancy between student and teacher diffusion distances across multiple scales ( $t \in \{1, 3, 5\}$ ), ensuring that the global geometric structure of the variant space is preserved.
Persistent Homology: Applied to the integrated feature space (specifically the delta embeddings), persistent homology identifies stable topological components. A Vietoris–Rips filtration constructs a filtration on the distance matrix, recording the "birth" and "death" of topological features. Features with long lifetimes represent robust, well-separated clusters of variants that share sequence-level similarities and functional profiles.

3. Variant Selection Criteria

Candidate variant sets are extracted from persistent components based on three intrinsic criteria:

Topological Significance: Component lifetime must exceed a threshold ( $\ell_{min} = 0.25$ ), ensuring robustness against noise.
Density Enrichment: The local density of variants within the component must be at least twofold higher than the background density in the same genomic region.
Functional Coherence: The group must show enrichment for at least one molecular signature of biological relevance (e.g., high CADD score, extreme rarity, LOH overlap, or structural variant co-occurrence).

The framework also employs a mixed-norm selection criterion grounded in optimal transport theory, filtering variants based on their distance from a compact core of known-benign profiles.

Key Contributions and Results

Cohort and Validation

The framework was developed and benchmarked on six families (14 individuals) selected for dense transgenerational multimorbidity (cancer, autoimmune, and/or cardiovascular disease). Five families (F1–F5) were used for development, while Family F6 served as a held-out independent test set.

Performance: PolyCLIP-T achieved a precision of 0.96 and recall of 0.67 in recovering known pathogenic variants from ClinVar, substantially outperforming baseline methods and individual encoders.
Label Efficiency: The model maintained high clustering quality (Adjusted Rand Index $\ge 0.70$ ) with as little as 1–5% of variants carrying disease labels, outperforming fully supervised classifiers.
Ablation Studies: Results confirmed that each component—sequence encoding, biological annotation, contrastive alignment, and topological regularisation—makes an independent and additive contribution to performance.

Discovery of Polygenic Modules

Applied to the test family (F6), PolyCLIP-T reduced 4.2 million raw variants to 112 interpretable candidates.

Non-Coding Recovery: Unlike conventional pipelines that focus on coding exons, PolyCLIP-T recovered a significant proportion of non-coding (38%) and structural candidates.
Cross-Disease Pleiotropy: The identified clusters revealed biologically coherent modules spanning disease categories. For example, genes such as ERBB3, HLA-A, and LPA were found to co-cluster, linking cancer, immune regulation, and cardiovascular risk.
Case Study (Family F6): While conventional analysis identified a pathogenic BRCA1 variant (explaining hereditary breast/ovarian cancer), PolyCLIP-T identified a distinct cluster of 15 constitutional variants (including ERBB3, HLA-A, LPA, PAX3, TNFRSF25). These variants, individually of modest effect and often classified as VUS, co-segregated with the broader multimorbid phenotype (cancer, Crohn's disease, stroke) and formed a stable topological component.

Significance and Claims

The paper posits that PolyCLIP-T provides an interpretable, scalable solution for genome-first investigations of disorders driven by polygenic architectures that evade single-variant analysis.

Paradigm Shift: The work advocates for a shift from "classification to selection," moving away from binary pathogenic–benign thresholds toward identifying coherent sets of variants that collectively indicate a disease mechanism.
Complementary Role: The authors explicitly state that PolyCLIP-T is not intended to replace the ACMG/AMP framework for definitive variant classification. Instead, it operates upstream as a filtering and prioritization layer to identify the small subset of variants most worthy of detailed manual curation.
Hypothesis Generation: The tool serves as a "discovery engine" that reduces the risk of prematurely discarding variants whose functional relevance may only become apparent when examined in combination. It generates hypotheses for specific genes and pathways (e.g., PI3K/AKT axis, immune-metabolic crosstalk) for further validation in larger cohorts.

Limitations and Future Directions

The authors maintain a modest stance regarding the current scope:

Cohort Size: The study relied on six families, prioritizing depth of characterization over sample size. Validation in larger, independent populations is essential to establish generalizability.
Circularity: The benchmarking against ClinVar introduces partial circularity because ClinVar classifications are partly derived from ACMG/AMP rules, which also feed into the model's scoring system. Independent, orthogonally curated validation sets are required for unbiased performance estimates.
Data Requirements: The current implementation requires multi-generational familial WGS data with tumor-normal pairs, limiting applicability to singleton or trio settings.
Ancestry Bias: Performance may be affected by biases in reference databases (e.g., gnomAD) toward European populations.

The authors conclude that while the framework shows promise for identifying constitutional susceptibility alleles in multimorbid families, prospective validation in independent multi-center cohorts is the necessary next step before clinical translation.