henhaobio
Sign in

bioRxivGeneticsPreprint

Machine learning-based prediction of human structural variation and characterization of associated sequence determinants

Daven LimiD, Runyang Nicolas LouiD, Nilah M. IoannidisiD, Peter H. SudmantiD

PDFDOICodeData
In-depth full-text digest available · Jump to digest ↓

Abstract

Structural variants (SVs) represent a major source of genetic diversity and play key roles in human disease and evolution. Yet, the extent to which local sequence context shapes the likelihood of structural variant formation remains poorly quantified. Here, we develop machine learning models to predict the occurrence of SVs across the human genome and characterize genomic determinants associated with their formation. We developed both a sequence only-based convolutional neural network (CNN) model as well as a random forest approach integrating diverse genomic annotations. Both models achieve high predictive performance individually (>90% AUROC) which can be further improved in an ensemble. The predictive ability of these models demonstrates that SV-prone regions can be accurately inferred from sequence context. Model interpretability techniques reveal key genomic contributors to SVs, including effects of sequence motifs such as microhomology and non-canonical DNA structures, as well as the presence of SV hotspots. We find that different classes of SVs exhibit distinct sequence determinants, with transposable elements and inversions displaying particularly unique signatures. Moreover, predicted SV probability correlates with allele frequency and gene functional constraint, indicating the potential utility of the model for variant effect prediction. These findings demonstrate that machine learning models trained on local sequence features can identify unstable genomic regions and provide a framework for quantifying SV susceptibility and SV variant effects in personalized genomics.

In-depth summary

Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 4 (15 Jul 2026) · CC BY 4.0

A sequence-only convolutional neural network and an annotation-based random forest predict human structural variant breakpoints from 400 bp windows (AUROC 0.903, 0.904); their ensemble reaches 0.916.

The degree to which local sequence context sets where human structural variants (SVs) form is unquantified, and prior machine learning predictors of rearrangements targeted smaller genomes (a eukaryotic pathogen, Arabidopsis) using annotations rather than raw sequence. This work fits raw-sequence and annotation-based models to population-scale long-read human SV calls, yielding per-position susceptibility scores and interpretable class-specific determinants.

Key results

  1. ShownBoth a sequence-only residual CNN and an annotation-based random forest separate SV breakpoint windows from random background with high accuracy; an ensemble that feeds CNN probability into the forest performs best, and all exceed logistic regression.

    Held-out test set (n=69,530): ensemble AUROC 0.916, AUPRC 0.936; random forest 0.904, 0.930; CNN 0.903, 0.926; logistic regression 0.890, 0.916. At a fixed 5% false positive rate the ensemble gave a 6% higher true positive rate than logistic regression; gains from longer context plateaued beyond 400bp (tested 100bp to 2000bp). · Development of machine learning models to…; Dataset construction

  2. ShownPerformance carried over to held-out chromosomes in Human Pangenome Reference Consortium (HPRC) genomes and to the T2T-complete CHM13 reference, where the CNN scored marginally higher than on GRCh38.

    Trained on HGSVC3 minus chromosomes 9, 10, 13, 17, 21 and tested on those chromosomes in 3 HPRC individuals: CNN AUROC above 0.86, AUPRC above 0.90; ensemble AUROC above 0.92, AUPRC above 0.94. CHM13 (188,500 SV calls >50bp): CNN AUROC 0.910 vs 0.903 on GRCh38, AUPRC 0.930 vs 0.926. · Development of machine learning models to…; Data sources

  3. ShownPredictability differs sharply by SV class: simple-repeat, SINE/Alu and LINE/L1 SVs are captured best by the sequence-only CNN, whereas no model predicts inversions particularly well and the CNN does extremely poorly on them.

    AUROC, CNN vs random forest / logistic regression / ensemble: simple repeat 0.994 vs 0.992/0.986/0.993; SINE/Alu 0.925 vs 0.614/0.626/0.807; LINE/L1 0.823 vs 0.676/0.677/0.745. Inversions (n = 300): CNN 0.480. The CNN also trailed the random forest on non-repeat SVs. · Model performance across SV classes; Model construction for SV subclasses

  4. ShownPerturbation analyses show the CNN learned known SV-linked sequence signals: breakpoint-spanning microhomology, G-quadruplex and Z-DNA motifs, GC- and CpG-rich k-mers, and the TAAA L1-endonuclease target motif at transposable element (TE) insertion sites.

    Saturation mutagenesis: sensitivity concentrated within ∼15bp of TE breakpoints, favoring TAAAA from position −1. 4-mer insertion effect vs Alu breakpoint enrichment ρ = 0.361, p=3.18e-09; top 5 motifs shifted probability ∼0.3-0.5. Greater k-mer pair homology raised probability, more for longer motifs; G4 and Z-DNA inserts exceeded length-matched random controls. · Effect of microhomologies, non-canonical DNA…; In silico saturation mutagenesis for TE…; K-mer, motif, microhomology, and non-canonical…; In silico saturation mutagenesis for TE…

  5. SuggestedPredicted SV probability tracks population and functional signals: common SVs score higher than rare ones, loss-of-function (LoF) tolerant genes score higher than constrained genes, and predictions anticorrelate with CADD-SV deleteriousness, hinting at variant-effect utility.

    Allele frequency trend: CNN p=1.6e-7, ensemble p=1.2e-22. Gene-level predictions (30 sampled sites per gene) vs gnomAD v4.0 LOEUF ρ = 0.205 (p = 4.42e-168), pLI ρ = −0.166 (p = 8.63e-39), observed SV count ρ = 0.330. CADD-SV PHRED: CNN ρ = −0.294, ensemble ρ = −0.335. · Gene functional constraint and tolerance; Constraint and tolerance

Methods, data and reproducibility

Breakpoint-centered 400 bp windows from HGSVC3 SVs versus random background were classified by a one-hot residual 1D CNN, a 1000-tree random forest on genomic annotations, logistic regression, and an ensemble adding CNN output to the forest; interpretation used Gini/permutation importance, in silico insertions and saturation mutagenesis, DeepSHAP and TF-MoDISco.

  • HGSVC Phase 3 SV calls against GRCh38 from near-T2T long-read assemblies of 65 individuals (130 haplotypes), paired with an equal number of random background windows · human · n = 173,824 SVs + 173,824 background (347,648 windows)
  • HGSVC3 small-variant and SV callsets used to derive neighboring-variant density features · human · n = 23,462,179 SNPs; 3,128,354 small ins; 3,049,659 small del
  • HGSVC3 SV calls against the T2T-complete CHM13 reference (CNN-only benchmark) · human · n = 188,500 SV calls (>50bp)
  • HPRC PAV SV callsets for HG00658, HG01099 and HG02027, evaluated on held-out chromosomes 9, 10, 13, 17, 21 · human · n = 3 individuals
  • RepeatMasker-defined SV subclass datasets (Alu, L1, SVA, simple repeat, non-repetitive, inversion), each balanced against background windows · human · n = Alu 4,111; L1 652; SVA 2,081; simple 14,650; nonrep 37,601; inv 300
  • Full-length Alu insertions from the test set used for in silico saturation mutagenesis · human · n = 1367
  • UCSC Genome Browser GRCh38 annotations: phyloP, recombination rate, RepeatMasker, cCREs, DNase accessibility, TF ChIP-seq peak clusters · human (GRCh38 reference)
  • gnomAD v4.0 gene constraint metrics (LOEUF, pLI) and CADD-SV PHRED scores for test-set SVs · human
Methods & tools
Residual 1D CNN (PyTorch 2.0.1), Random forest and logistic regression (scikit-learn 1.6.1), RepeatMasker 4.1.5, DeepSHAP, TF-MoDISco, DNABERT k-mer embeddings, PAV SV caller (source of HGSVC3 and HPRC calls), CADD-SV, gnomAD v4.0 LOEUF and pLI, SciPy 1.14.1 (Spearman correlation), UCSC Genome Browser annotations, transformers 4.41.2
Software & environment
Python 3.10.14, PyTorch 2.0.1, scikit-learn 1.6.1, transformers 4.41.2, SciPy 1.14.1, RepeatMasker 4.1.5, plus NumPy, pandas, matplotlib; random seed 42. No container or lockfile described.
Assay conditions & compute
GPU-enabled hardware used when available; GPU model, runtime and memory footprint not reported.
Reproducibility gaps
Availability statement gives only a GitHub repository; release of trained weights, processed feature tables or genome-wide prediction tracks is not mentioned.
Limitations (5)
  • Only highly localized sequence context is modeled (400 bp windows); the authors call this a major limitation that underlies poor prediction of inversions, which depend on long-range homology (CNN AUROC 0.480). · stated by the authors
  • The authors note that the association between predicted SV probability and gene loss-of-function tolerance, though highly significant, is weak in magnitude. · stated by the authors
  • The models omit replication timing, chromatin organization and 3D genome architecture; the authors propose that adding these could improve predictions, especially for complex events. · stated by the authors
  • The top-ranked feature, neighboring SV count, comes from the same HGSVC3 callset that defines positives, and external tests used HPRC calls from the same PAV caller in 3 genomes, so hotspot clustering and caller-specific breakpoint placement may inflate scores. · noted in this summary
  • Evaluation uses a balanced design (equal SV and background windows), so AUPRC and predicted probabilities are not calibrated to the genome-wide rarity of breakpoints, limiting direct use as per-base mutability rates. · noted in this summary

Read this if you study mechanisms of human structural variation or need sequence-based SV susceptibility scores, with interpretable class-specific signals for TE insertions, repeat-associated SVs and inversions.

Citations
1

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISKalra '26Catona '26Sharo '22Kumar '20Gjoni '26Ganel '16
64% semantic matchBioinformatics
Archit Kalra, Luis F. Paulin, Fritz Joachim Sedlazeck·bioRxiv·Jan 30, 2026