David Bonet

Genome-wide association studies have successfully identified thousands of genetic associations, yet their predominant reliance on European-descent populations limits insights into the full spectrum of human genetic diversity and its impact on disease. Admixture mapping offers a powerful, complementary approach by leveraging differences in haplotype frequencies across ancestral backgrounds to identify risk loci for complex traits. Here, we perform a large-scale, multi-ancestry admixture mapping study across 415,792 unrelated individuals in the UK Biobank, examining associations between local haplotype ancestry and 108 phenotypes. Our approach identifies 13 genome-wide significant ancestry-phenotype associations, recovering previously reported signals while uncovering four novel ancestry-associated findings, including new risk loci for atrial fibrillation, dermatitis, and angina pectoris. To overcome the limited resolution of traditional admixture mapping, we implemented a conditional fine-mapping framework, which enabled us to localize four putatively causal variants. In silico variant effect prediction and eQTL integration revealed regulatory and missense effects predominantly localized to lung, and immune tissues, aligning with captured phenotypes such as asthma, dermatitis, and hypothyroidism. Notably, our findings demonstrate striking genetic heterogeneity, revealing how the same clinical phenotype can arise through distinct genetic pathways depending on the ancestral background. Overall, this work highlights the critical importance of modeling local ancestry structure to refine genetic associations, uncover novel disease mechanisms, and improve the equitable translation of genomic medicine.

Tabular data underpins decisions across science, industry, and public services. Despite rapid progress, advances in deep learning have not fully carried over to the tabular domain, where gradient-boosted decision trees (GBDTs) remain a default choice in practice. We present iLTM, an integrated Large Tabular Model that unifies tree-derived embeddings, dimensionality-agnostic representations, a meta-trained hypernetwork, multilayer perceptrons (MLPs), and retrieval within a single architecture. Pre-trained on more than 1,800 heterogeneous classification datasets, iLTM achieves consistently superior performance across tabular classification and regression tasks, from small datasets to large and high-dimensional tasks. After light fine-tuning, the meta-trained hypernetwork transfers to regression targets, matching or surpassing strong baselines. Extensive experiments show that iLTM outperforms well-tuned GBDTs and leading deep tabular models while requiring less task-specific tuning. By bridging the gap between tree-based and neural methods, iLTM offers a new framework for tabular foundation models for robust, adaptable, and scalable tabular learning.

Motivation: Recent advances in inference of the ancestral recombination graph (ARG), which describes how segments of chromosomes trace back through recombination and shared lineages, have made it possible to reconstruct genome-wide genealogies for large cohorts, but it remains difficult to summarize and use this information for population genetic analyses. Results: We present ARGformer, an encoder-only transformer that learns context-dependent embeddings with a self-supervised masked objective finetuned with contrastive learning for downstream retrieval tasks. We train ARGformer on genealogies from coalescent simulations and on genealogies inferred from ancient and present-day Homo sapiens genomes. Using only these learned embeddings, without access to genotype matrices, ARGformer captures patterns of global population structure and supports ancestry inference through clustering and nearest-neighbor retrieval. On genealogies that include archaic hominins, ARGformer can highlight Denisovan-derived segments in Oceanian genomes and reveals Oceanian-like ancestry in South American Indigenous populations. Availability and Implementation: ARGformer is available at https://github.com/AI-sandbox/ARGformer.

Polygenic risk models trained in one ancestry often fail to perform well in others, in part due to linkage disequilibrium and allele frequency differences across ancestries. In response, separate models trained for specific ancestries have been introduced. However, this single ancestry approach is untenable in admixed groups, which have ancestry from multiple sources varying across the genome and between individuals. Here we present Combine, a biobank-scale sparse regression framework that augments per-variant genotype features with local ancestry dosages and fits all effects jointly using a variant-level group lasso penalty. In 99,298 admixed participants from the All of Us Research Program, Combine substantially outper-forms state-of-the-art multi-ancestry summary-statistic approaches (e.g., 144% relative improvement over PRS-CSx for white blood cell count). Furthermore, it matches or improves upon the predictive performance of highly optimized individual-level models (iPGS/snpnet) across seven of nine evaluated phenotypes, while uniquely providing locus-level interpretability to disentangle shared allelic effects from ancestry-linked tagging. An ancestry-specific extension, Combine-S, estimates haplotypic ancestry-associated SNP effects together with local-ancestry terms, enabling systematic identification of ancestry-dependent effect magnitudes and sign differences at established and plausible loci. Finally, we show that incorporating external GWAS evidence through group-specific penalty weights improves LDL cholesterol prediction without pre-filtering variants. Together, Combine provides a scalable framework for polygenic modeling that prioritizes efficient local ancestry-aware modeling and interpretation in admixed biobanks.

Recent advances in inference of the ancestral recombination graph (ARG), which describes how segments of chromosomes trace back through recombination and shared lineages, have made it possible to reconstruct genome-wide genealogies for large cohorts, but it remains difficult to summarize and use this information for population genetic analyses. We present ARGformer, an encoder-only transformer that learns context-dependent embeddings with a self-supervised masked objective finetuned with contrastive learning for downstream retrieval tasks. We train ARGformer on genealogies from coalescent simulations and on genealogies inferred from ancient and present-day Homo sapiens genomes. Using only these learned embeddings, without access to genotype matrices, ARGformer captures patterns of global population structure and supports ancestry inference through clustering and nearest-neighbor retrieval. On genealogies that include archaic hominins, ARGformer can highlight Denisovan-derived segments in Oceanian genomes and reveals Oceanian-like ancestry in South American Indigenous populations. ARGformer is available at https://github.com/AI-sandbox/ARGformer .

Precision medicine models often perform better for populations of European ancestry due to the over-representation of this group in the genomic datasets and large-scale biobanks from which the models are constructed. As a result, prediction models may misrepresent or provide less accurate treatment recommendations for underrepresented populations, contributing to health disparities. This study introduces an adaptable machine learning toolkit that integrates multiple existing methodologies and novel techniques to enhance the prediction accuracy for underrepresented populations in genomic datasets. By leveraging machine learning techniques, including gradient boosting and automated methods, coupled with novel population-conditional re-sampling techniques, our method significantly improves the phenotypic prediction from single nucleotide polymorphism (SNP) data for diverse populations. We evaluate our approach using the UK Biobank, which is composed primarily of British individuals with European ancestry, and a minority representation of groups with Asian and African ancestry. Performance metrics demonstrate substantial improvements in phenotype prediction for underrepresented groups, achieving prediction accuracy comparable to that of the majority group. This approach represents a significant step towards improving prediction accuracy amidst current dataset diversity challenges. By integrating a tailored pipeline, our approach fosters more equitable validity and utility of statistical genetics methods, paving the way for more inclusive models and outcomes.

Wake-up word spotting in noisy environments is a critical task for an excellent user experience with voice assistants. Unwanted activation of the device is often due to the presence of noises coming from background conversations, TVs, or other domestic appliances. In this work, we propose the use of a speech enhancement convolutional autoencoder, coupled with on-device keyword spotting, aimed at improving the trigger word detection in noisy environments. The end-to-end system learns by optimizing a linear combination of losses: a reconstruction-based loss, both at the log-mel spectrogram and at the waveform level, as well as a specific task loss that accounts for the cross-entropy error reported along the keyword spotting detection. We experiment with several neural network classifiers and report that deeply coupling the speech enhancement together with a wake-up word detector, e.g., by jointly training them, significantly improves the performance in the noisiest conditions. Additionally, we introduce a new publicly available speech database recorded for the Telefónica’s voice assistant, Aura. The OK Aura Wake-up Word Dataset incorporates rich metadata, such as speaker demographics or room conditions, and comprises hard negative examples that were studiously selected to present different levels of phonetic similarity with respect to the trigger words “OK Aura”.