Taedong Yun

Heart failure with preserved ejection fraction (HFpEF) affects over 30 million people and lacks disease-modifying therapies. Although genomic-led drug discovery increases success by more than 2.6-fold, HFpEF genomic discovery remains constrained by imprecise phenotyping in biobanks, with only two loci identified to date. Biobanks lack HFpEF diagnostic codes and echocardiograms, yet HFpEF diagnosis exists along a continuum and is inherently probabilistic, presenting an opportunity for multimodal prediction. Here we introduce TRI-modal Assessment and Discovery of HFpEF (TRIAD-HFpEF), a machine learning framework integrating electrocardiograms, cardiac magnetic resonance imaging, and biomarkers to assign HFpEF probabilities. Deployed in UK Biobank, these probabilities validate with respect to mortality, hospitalizations, and structural and functional HFpEF features. Genome-wide and proteomic analyses reveal over 90 novel loci, a 45-fold expansion, and distinguish causal proteins from non-causal biomarkers of disease progression, prioritizing 11 therapeutic targets and 7 non-causal biomarkers. We identify FLT3 as one of the 11 therapeutic targets, consistent with the reported 7-fold increased heart failure risk from FLT3 inhibitors in leukemia. We validate this finding by demonstrating significant worsening of diastolic function following FLT3 inhibitor treatment in an independent clinical cohort. Conversely, MPO emerged as one of the 7 non-causal biomarkers, aligning with three recent negative MPO inhibitor trials. TRIAD-HFpEF demonstrates that machine learning-derived phenotypes can unlock genetic discovery in complex syndromes, identifying actionable targets while deprioritizing associations reflecting disease consequences rather than causes.

UNSTRUCTURED Generative artificial intelligence (Generative AI) is transforming healthcare. With this evolution comes optimism regarding the impact it will have on mental health, as well as concern regarding the risks that come with generative AI operating in the mental health domain. Much of the investment in, and academic and public discourse about, AI-powered solutions for mental health has focused on therapist chatbots. Despite the common assumption that chatbots will be the most impactful application of GenAI to mental health, we make the case here for a lower-risk, high impact use case: leveraging generative AI to enhance and scale training in mental health service provision. We highlight key benefits of using generative AI to help train people to provide mental health services and present a real-world case study in which generative AI improved the training of veterans to support one another’s mental health. With numerous potential applications of generative AI in mental health, we illustrate why we should invest in using generative AI to support training people in mental health service provision.

Importance: Obtaining spirometry requires repeated testing and using the maximal values based on quality control criteria. Whether the suboptimal efforts are useful for the prediction of respiratory outcomes is not clear. Objective: To determine whether a machine learning model could predict respiratory outcomes and mortality based on suboptimal spirometry. Design: Observational cohorts (UK Biobank and COPDGene). Setting: Multi-center; population, and disease-enriched. Participants: UK aged 40-69; US aged 45-80, >10 pack-years smoking, without respiratory diseases other than COPD or asthma. Exposures: Raw spirograms (volume-time). Main outcomes and measures: To create a combined representation of lung function we implemented a contrastive learning approach, Spirogram-based Contrastive Learning Framework (Spiro-CLF), which utilized all recorded volume-time curves per participant and applied different transformations (e.g. flow-volume, flow-time). We defined “maximal” efforts as those passing quality control (QC) with the maximum FVC; all other efforts, including submaximal and QC-failing efforts, were defined as “suboptimal”. We trained the Spiro-CLF model using both maximal and suboptimal efforts from the UK Biobank. We tested the model in a held-out 20% testing UK Biobank subset and COPDGene, on 1) binary predictions of FEV1/FVC <0.7, and FEV1 Percent Predicted (FEV1PP) <80%, 2) Cox regression for all-cause mortality, and 3) prediction of respiratory phenotypes. Results: We trained Spiro-CLF on 940,705 volume-time curves from 352,684 UKB participants with 2-3 spirometry efforts per individual (66.7% with 3 efforts) and at least one QC-passing spirometry effort. Of all spirometry efforts, 61.6% were suboptimal (37.5% submaximal and 24.1% QC-failing). In the UK Biobank, Spiro-CLF using QC-failing and submaximal efforts predicted FEV1/FVC < 0.7 with an Area under the Receiver Operating Characteristics (AUROC) of 0.956, mortality with a concordance index of 0.647, and asthma with a 9-42% improvement versus baseline models. In COPDGene (n=10,110 participants), adding QC-passing, submaximal efforts did not improve the prediction of lung function or mortality; however, Spiro-CLF representations predicted asthma and respiratory phenotypes (joint test P ≤ 2 × 10−3). Conclusions and Relevance: A machine-learning model can predict respiratory phenotypes using suboptimal spirometry; results from all spirometry efforts may contain valuable data. Additional studies are required to determine performance and utility in specific clinical scenarios.

Taedong Yun, Eric Yang, Mustafa Safdari, Jong Ha Lee, Vaishnavi Vinod Kumar, S. Sara Mahdavi, Jonathan Amar, Derek Peyton, Reut Aharony, Andreas Michaelides PhD, Logan Douglas Schneider, Isaac Galatzer-Levy, Yugang Jia, John Canny, Arthur Gretton, Maja Mataric. Findings of the Association for Computational Linguistics: ACL 2025. 2025.

Although high-dimensional clinical data (HDCD) are increasingly available in biobank-scale datasets, their use for genetic discovery remains challenging. Here we introduce an unsupervised deep learning model, Representation Learning for Genetic Discovery on Low-Dimensional Embeddings (REGLE), for discovering associations between genetic variants and HDCD. REGLE leverages variational autoencoders to compute nonlinear disentangled embeddings of HDCD, which become the inputs to genome-wide association studies (GWAS). REGLE can uncover features not captured by existing expert-defined features and enables the creation of accurate disease-specific polygenic risk scores (PRSs) in datasets with very few labeled data. We apply REGLE to perform GWAS on respiratory and circulatory HDCD—spirograms measuring lung function and photoplethysmograms measuring blood volume changes. REGLE replicates known loci while identifying others not previously detected. REGLE are predictive of overall survival, and PRSs constructed from REGLE loci improve disease prediction across multiple biobanks. Overall, REGLE contain clinically relevant information beyond that captured by existing expert-defined features, leading to improved genetic discovery and disease prediction.

High-dimensional clinical data have become invaluable resources for genetic studies, due to their accessibility in biobank-scale datasets and the development of high performance modeling techniques especially using deep learning. Recent work has shown that low dimensional embeddings of these clinical data learned by variational autoencoders (VAE) can be used for genome-wide association studies and polygenic risk prediction. In this work, we consider multiple unsupervised learning methods for learning disentangled representations, namely autoencoders, VAE, beta-VAE, and FactorVAE, in the context of genetic association studies. Using spirograms from UK Biobank as a running example, we observed improvements in the number of genome-wide significant loci, heritability, and performance of polygenic risk scores for asthma and chronic obstructive pulmonary disease by using FactorVAE or beta-VAE, compared to standard VAE or non-variational autoencoders. FactorVAEs performed effectively across multiple values of the regularization hyperparameter, while beta-VAEs were much more sensitive to the hyperparameter values.

High-dimensional clinical data are becoming more accessible in biobank-scale datasets. However, effectively utilizing high-dimensional clinical data for genetic discovery remains challenging. Here we introduce a general deep learning-based framework, REpresentation learning for Genetic discovery on Low-dimensional Embeddings (REGLE), for discovering associations between genetic variants and high-dimensional clinical data. REGLE uses convolutional variational autoencoders to compute a non-linear, low-dimensional, disentangled embedding of the data with highly heritable individual components. REGLE can incorporate expert-defined or clinical features and provides a framework to create accurate disease-specific polygenic risk scores (PRS) in datasets which have minimal expert phenotyping. We apply REGLE to both respiratory and circulatory systems: spirograms which measure lung function and photoplethysmograms (PPG) which measure blood volume changes. Genome-wide association studies on REGLE embeddings identify more genome-wide significant loci than existing methods and replicate known loci for both spirograms and PPG, demonstrating the generality of the framework. Furthermore, these embeddings are associated with overall survival. Finally, we construct a set of PRSs that improve predictive performance of asthma, chronic obstructive pulmonary disease, hypertension, and systolic blood pressure in multiple biobanks. Thus, REGLE embeddings can quantify clinically relevant features that are not currently captured in a standardized or automated way.

Background: Spirometry measures lung function by selecting the best of multiple efforts meeting pre-specified quality control (QC), and reporting two key metrics: forced expiratory volume in 1 second (FEV ₁ ) and forced vital capacity (FVC). We hypothesize that discarded submaximal and QC-failing data meaningfully contribute to the prediction of airflow obstruction and all-cause mortality. Methods: We evaluated volume-time spirometry data from the UK Biobank. We identified “best” spirometry efforts as those passing QC with the maximum FVC. “Discarded” efforts were either submaximal or failed QC. To create a combined representation of lung function we implemented a contrastive learning approach, Spiro gram-based C ontrastive L earning F ramework (Spiro-CLF), which utilized all recorded volume-time curves per participant and applied different transformations (e.g. flow-volume, flow-time). In a held-out 20% testing subset we applied the Spiro-CLF representation of a participant’s overall lung function to 1) binary predictions of FEV ₁ /FVC < 0.7 and FEV ₁ Percent Predicted (FEV ₁ PP) < 80%, indicative of airflow obstruction, and 2) Cox regression for all-cause mortality. Findings: We included 940,705 volume-time curves from 352,684 UK Biobank participants with 2-3 spirometry efforts per individual (66.7% with 3 efforts) and at least one QC-passing spirometry effort. Of all spirometry efforts, 24.1% failed QC and 37.5% were submaximal. Spiro-CLF prediction of FEV ₁ /FVC < 0.7 utilizing discarded spirometry efforts had an Area under the Receiver Operating Characteristics (AUROC) of 0.981 (0.863 for FEV ₁ PP prediction). Incorporating discarded spirometry efforts in all-cause mortality prediction was associated with a concordance index (c-index) of 0.654, which exceeded the c-indices from FEV ₁ (0.590), FVC (0.559), or FEV ₁ /FVC (0.599) from each participant’s single best effort. Interpretation: A contrastive learning model using raw spirometry curves can accurately predict lung function using submaximal and QC-failing efforts. This model also has superior prediction of all-cause mortality compared to standard lung function measurements. Funding: MHC is supported by NIH R01HL137927, R01HL135142, HL147148, and HL089856. BDH is supported by NIH K08HL136928, U01 HL089856, and an Alpha-1 Foundation Research Grant. DH is supported by NIH 2T32HL007427-41 EKS is supported by NIH R01 HL152728, R01 HL147148, U01 HL089856, R01 HL133135, P01 HL132825, and P01 HL114501. PJC is supported by NIH R01HL124233 and R01HL147326. SPB is supported by NIH R01HL151421 and UH3HL155806. TY, FH, and CYM are employees of Google LLC

Circular consensus sequencing with Pacific Biosciences (PacBio) technology generates long (10–25 kilobases), accurate ‘HiFi’ reads by combining serial observations of a DNA molecule into a consensus sequence. The standard approach to consensus generation, pbccs, uses a hidden Markov model. We introduce DeepConsensus, which uses an alignment-based loss to train a gap-aware transformer–encoder for sequence correction. Compared to pbccs, DeepConsensus reduces read errors by 42%. This increases the yield of PacBio HiFi reads at Q20 by 9%, at Q30 by 27% and at Q40 by 90%. With two SMRT Cells of HG003, reads from DeepConsensus improve hifiasm assembly contiguity (NG50 4.9 megabases (Mb) to 17.2 Mb), increase gene completeness (94% to 97%), reduce the false gene duplication rate (1.1% to 0.5%), improve assembly base accuracy (Q43 to Q45) and reduce variant-calling errors by 24%. DeepConsensus models could be trained to the general problem of analyzing the alignment of other types of sequences, such as unique molecular identifiers or genome assemblies.

Genome-wide association studies (GWASs) examine the association between genotype and phenotype while adjusting for a set of covariates. Although the covariates may have non-linear or interactive effects, due to the challenge of specifying the model, GWAS often neglect such terms. Here we introduce DeepNull, a method that identifies and adjusts for non-linear and interactive covariate effects using a deep neural network. In analyses of simulated and real data, we demonstrate that DeepNull maintains tight control of the type I error while increasing statistical power by up to 20% in the presence of non-linear and interactive effects. Moreover, in the absence of such effects, DeepNull incurs no loss of power. When applied to 10 phenotypes from the UK Biobank (n = 370K), DeepNull discovered more hits (+6%) and loci (+7%), on average, than conventional association analyses, many of which are biologically plausible or have previously been reported. Finally, DeepNull improves upon linear modeling for phenotypic prediction (+23% on average).