Degui Zhi

This paper develops efficient algorithms for detecting long exact matches between genome assemblies in compressed space and sublinear time, revealing novel structural variations in centromeres.

On the All of Us workbench, a Recomb-Mix-based cloud pipeline runs whole-genome local ancestry inference for an estimated $0.01 per sample, versus $0.43 for FLARE and $4.58 for RFMix.

The paper introduces sum-h2, a fast framework for evaluating genetic discovery from deep learning-derived phenotypes using heritability measures, and establishes a Genetic Discovery Arena for benchmarking.

TorchGWAS2 improves genome-wide association testing in related samples by using a deterministic variance-correction algorithm, achieving faster and more powerful results across large-scale omics data.

PerturbLDM predicts single-cell transcriptional responses to perturbations using a latent diffusion model, outperforming existing methods in drug-dose-cell line combinations and capturing key biological programs.

The paper introduces algorithms for constructing RLBWT-based compressed full text indexes in compressed space, reducing memory usage for large-scale pangenome analysis.

This paper introduces SLAB, an algorithm for identifying haplotype block cores using PBWT, and applies it to UK Biobank data to detect selection signals.

This study evaluates the replicability of unsupervised deep learning image phenotypes across different models, data splits, and cohorts, finding significant stability and biological relevance.

HiFiMAP provides high-resolution IBD mapping at SNP level, outperforming window-based methods and detecting novel associations in UK Biobank and brain imaging data.

The paper uses deep learning dynamics to uncover epoch-specific genetic architectures of brain morphology from UK Biobank data.

This paper introduces a multimodal contrastive learning framework to derive heritable brain representations from paired T1- and T2-weighted MRIs, improving genetic discovery and revealing shared biological targets.

This study compares deep learning methods to predict cognitive decline using brain MRI and clinical data over two years, showing the value of integrating image and tabular data.

The paper introduces HP-PBWT, a parallel PBWT algorithm for efficient haplotype matching in large biobank datasets, achieving significant speed-ups in both IO-included and IO-excluded benchmarks.

This study uses deep learning to derive image phenotypes from cardiac MRI and performs GWAS, identifying 323 SNPs and 628 genes associated with cardiac traits and diseases.

This study conducts a multi-ancestry GWAS on periodontitis using EHR data from the All of Us Research Platform, identifying significant genetic loci associated with periodontitis grade and other phenotypes.

This paper introduces a deep learning framework to analyze fractional anisotropy images, identifying genetic variants associated with white matter microstructure and linking them to neurological disorders.

Recomb-Mix improves local ancestry inference accuracy in admixed populations by integrating existing methods with a graph collapsing trick, outperforming current approaches in diverse scenarios.

This study introduces a framework that predicts protein-ligand binding affinity by folding proteins, determining binding conformations, and using 3D structures, achieving performance comparable to existing docking-free methods.

This paper uses a vision transformer autoencoder to extract brain imaging phenotypes and identifies 10 novel genetic loci associated with brain structure, 3 of which were previously unreported.

This paper introduces Dynamic μ-PBWT, a memory-efficient data structure for dynamic updates of haplotype panels, supporting efficient queries and reducing memory usage compared to existing methods.

This paper defines haplotype matches in pangenome graphs and provides efficient algorithms for querying them using GBWT and related data structures.

The paper introduces JAGWAS, a multi-phenotype GWAS tool that identifies more genetic loci for brain imaging phenotypes than single-phenotype GWAS.

The paper introduces a privacy-preserving genotype imputation method using haplotype classifiers, improving accuracy and throughput compared to linear models.

The paper introduces ROH-DICE, a method to identify ROH diplotypes in UK Biobank, finding associations between clustered ROHs and diseases like autoimmune disorders and haemochromatosis.

ProxyTyper generates proxy panels to protect genetic data privacy during genotype imputation with minimal accuracy loss.

FastRecomb efficiently estimates population-specific genetic maps from large biobank-scale genotyping data, achieving state-of-the-art performance at 10k resolution.

The paper introduces the MPSC problem as an alternative to the LS model for haplotype threading, enabling efficient haplotype reconstruction on large panels.

RaPID-Query enables fast and accurate genealogical search by identifying IBD segments between a query and a large database, achieving low error rates and quick processing times.

This paper introduces an unsupervised deep learning method to derive brain imaging phenotypes from MRI data, identifying new genetic loci associated with these endophenotypes.

The paper introduces Syllable-PBWT, a space-efficient data structure for haplotype long-match queries, reducing memory use by over 100 times compared to existing methods.