• Melanie Schirmer

The human microbiome encodes millions of microbial genes that play a critical role for our health. However, identifying disease-associated microbial genes and elucidating their role in disease etiologies remains challenging. We developed a gene-centric approach for metagenomic analysis that combines systematic functional profiling with additional gene neighborhood information to provide novel biological insights. After constructing a microbial gene catalog, we use information on their genomic neighborhood from the metagenomic assemblies to build a metagenomic assembly graph (MAGraph). Nodes represent genes and edges indicate their co-localization, while information such as gene abundance and annotation is encoded as node properties. This gene-centric approach can reveal important insights into the microbiome, including bacterial operon structures, disease-associated strain heterogeneity, and evolutionary diversification of genes. We applied our gene-centric approach to 9,053 metagenomic stool samples from 24 human cohorts as part of a meta-analysis and investigated microbiome functionality in inflammatory bowel disease (IBD) and colorectal cancer (CRC). This revealed several new functional disease insights: (1) We identified IBD and CRC signature genes that were consistently increased in patients across different cohorts. (2) We detected a duplication event of a nitrate reduction (nar) operon in Enterobacteriaceae spp., which facilitates bacterial blooming in the gut during inflammation. Interestingly, only the second copy of the nar operon was up-regulated during stress conditions and virulence induction. (3) Lastly, we identified a Tn916-like mobile genetic element shared across diverse bacterial lineages, which was a carrier of tetracycline resistance, thereby establishing a connection between this mobile element and widespread antibiotic resistance across species. In summary, gene-centric metagenomic analyses can reveal important insights into microbiome functionality in human diseases.

• Mirko Treccani, Mirko Treccani, Department of Food and Drug
• Cristiana Mignogna, Cristiana Mignogna, Department of Food and Drug
• Lucia Ghiretti, Lucia Ghiretti, Department of Food and Drug
• José Fernando Rinaldi de Alvarenga, José Fernando Rinaldi de Alvarenga, Department of Food and Drug
• Claudia Favari, Claudia Favari, Department of Food and Drug
• Nicola Luigi Bragazzi, Nicola Luigi Bragazzi, Department of Food and Drug
• Maria Sole Morandini, Maria Sole Morandini, Department of Food and Drug
• Cristina Del Burgo-Gutiérrez, Cristina Del Burgo-Gutiérrez, Department of Food and Drug
• Alice Rosi, Alice Rosi, Department of Food and Drug
• Cristiano Negro, Cristiano Negro, Department of Food and Drug
• Feder

(Poly)phenols (PPs) are a group of bioactive compounds found in plant-based food, widely consumed within diet. Several studies have reported the beneficial effects of PPs in preventing chronic diseases through a myriad of mechanisms of action. However, the bioavailability and effects of these compounds greatly differ across individuals, causing uneven physiological responses. To understand their inter-individual variability, we present a multi-omics investigation comprising genomics, metagenomics and metabolomics. We recruited 300 healthy individuals and collected biological samples (blood, urine, and faeces), anthropometric measurements, health status and lifestyle/dietary information. After identification by UPLC-IMS-HRMS and quantification by UPLC-QqQ-MS/MS, the large set of phenolic metabolites underwent dimensionality reduction and clustering to identify individuals with similar metabolic profiles (metabotypes), identifying high and low PP producers. Then, genomics and metagenomics investigations were performed to gain insights on inter-individual differences and unravel the potential pathophysiological impact of these molecules, with particular regards to cardiometabolic diseases. In details, genome-wide association studies followed by computational functional analyses on genetic variants, and taxonomic and functional investigations of gut microbiome were performed, showing hints for associations in genes and microbial species related to PP metabolism, together with unprecedented genetic associations. Genomics were further investigated in terms of gene networks and computational functional analyses, identifying differentially expressed genes, gene sets enrichments, candidate regulatory regions, and interacting loci and chromatin states, and associations with metabolic traits and diseases. Overall, we demonstrated the benefits of omics research in nutrition, advancing the field of personalised nutrition and health.

• Mehdi Layeghifard, Mehdi Layeghifard, The Hospital for Sick Children
• Marcos Díaz-Gay, Marcos Díaz-Gay, Spanish National Cancer Research Center
• Erik N. Bergstrom, Erik N. Bergstrom, University of California San Diego
• Mathepan J. Mahendralingam, Mathepan J. Mahendralingam, The Hospital for Sick Children
• Sasha Blay, Sasha Blay, The Hospital for Sick Children
• Pedro L. Ballester, Pedro L. Ballester, The Hospital for Sick Children
• Elli Papaemmanuil, Elli Papaemmanuil, Memorial Sloan Kettering Cancer Center
• Mark J. Cowley, Mark J. Cowley, The University of New South Wales
• Anita Villani, Anita Villani, The Hospital for Sick Children
• Ludmil B. Alexandrov, Ludmil B. Alexandrov, University of California San Diego
• Adam Shlien, Adam Shlien, The Hospital for Sick Children

Children with cancer endure numerous short- and long-term side effects of treatment, but the extent of DNA damage associated with chemotherapy exposure remains largely unclear. We used mutational signatures to measure this damage directly in 611 whole-genome sequenced tumours from a multi-institutional pediatric cancer cohort enriched with treated tumours. Compared to treatment-naïve tumours, post-therapy cancers harbored nearly three times as many private signatures (p = 0.0001), twice the total burden of somatic mutations (p = 0.023), and 10% more oncogenic drivers (p = 0.016). Our analysis uncovered 15 therapy-associated signatures and revealed patterns of drug-specific and tissue-specific mutagenesis. Platinum-based treatments, for which we more than doubled the number of known associated signatures, contributed the highest mutation burden in most patients. By integrating clinical exposure timelines with signature evolution, we defined a minimum latency of 91 days and a burden threshold of ~1500 mutations for the emergence of platinum signatures. Notably, over one-third of platinum-treated tumours exhibited measurable resistance within 12 months. To further explore therapy-associated genomic alterations beyond signatures, we employed machine learning (ML) models, which identified additional, non-canonical genomic features predictive of platinum, anthracycline, and antimetabolite exposure, suggesting a broader landscape of therapy-induced damage. This comprehensive genomic analysis provides critical insights into the complex mutagenic legacy of chemotherapy in childhood cancer. The identified therapy-specific signatures, their temporal dynamics, and the novel genomic features uncovered by ML offer potential biomarkers for monitoring treatment response, predicting resistance, and informing future interventions aimed at treatment de-escalation and early detection of resistant clones.

• Lena Möbus, Lena Möbus, FHAIVE
• Angela Serra, Angela Serra, FHAIVE
• Stephan Weidinger, Stephan Weidinger, Department of Dermatology
• Dario Greco, Dario Greco, FHAIVE

Atopic dermatitis (AD) presents with high inter-individual variability in both disease progression and treatment response. Understanding how patients evolve over time—clinically and molecularly—is critical for enabling more personalised care. This study aims to stratify AD patients through multi-layered patient similarity analysis by integrating rich longitudinal clinical records with transcriptomic and other omics data. By quantifying patient-to-patient similarity across multiple data types and timepoints, we seek to uncover dynamic patterns linking clinical trajectories with underlying molecular profiles.

We analyse a prospective, observational cohort of 419 moderate-to-severe AD patients receiving routine care, with repeated clinical assessments and optional biosampling at approximately 3, 6, and 12 months post-treatment initiation. We compute patient-pairwise distance matrices within and across data layers: clinical variables (numeric, ordinal, categorical), skin and blood transcriptomics, and genomic profiles (rank-based distance metrics and other custom measures). A key challenge is the incomplete overlap of data layers across patients; thus, distances are computed using only shared features between patient pairs. These are integrated into a composite similarity matrix to support longitudinal stratification and comparative analyses.

Preliminary analyses reveal structured shifts in patient similarity over time, both in clinical scores and transcriptomic profiles. Clinical distance matrices at different timepoints show evolving clustering patterns, while transcriptomic data indicate increased molecular divergence following treatment. These dynamic shifts are reflected in the density distributions of patient-pairwise distances, supporting temporal changes in both clinical and molecular disease states.

This integrative approach supports precision medicine in AD by accommodating data sparsity while capturing meaningful patient-level similarities.

• Nicola Calonaci, Nicola Calonaci, University of Trieste
• Stefano Scalera, Stefano Scalera, IRCCS Regina Elena National Cancer Institute
• Giulio Caravagna, Giulio Caravagna, University of Trieste
• Eriseld Krasniqi, Eriseld Krasniqi, IRCCS Regina Elena National Cancer Institute
• Giorgia Gandolfi, Giorgia Gandolfi, University of Trieste
• Biagio Ricciuti, Biagio Ricciuti, Lowe Center for Thoracic Oncology
• Daniel Colic, Daniel Colic, Institute for Research in Biomedicine
• Marcello Maugeri-Saccà, Marcello Maugeri-Saccà, IRCCS Regina Elena National Cancer Institute
• Salvatore Milite, Salvatore Milite, Centre for Computational Biology

The intricate interplay between somatic mutations and copy number alterations critically influences tumour evolution and clinical outcomes. Yet, conventional genomic analyses often treat these biomarkers independently, overlooking the role of mutant gene dosage—a key mechanistic consequence of their interaction. We developed INCOMMON, an innovative computational model for rapidly inferring allele-specific copy numbers directly from read count data. We applied INCOMMON to 500,000 mutations in 60,000 publicly available clinical samples spanning 39 cancer types. We found 11 genes and 3 mutational hotspots exhibiting recurrent tumour-specific patterns associated with high mutant dosage across 17 tumours. By stratifying more than 24,000 patients based on mutant dosage across actionable oncogenes and tumour suppressors, we identified 6 groups with distinct prognostic significance across mutant dosage classes, and 4 novel biomarkers not detectable in a standard mutation-centric stratification. Additionally, 11 mutant dosage-defined subgroups showed increased metastatic propensity, with 6 enriched for site-specific dissemination patterns. By eliminating reliance on controlled-access raw sequencing data, our method offers a practical and scalable path for integrating dosage-aware biomarkers into clinical research. This augmented insight into genomic drivers enhances our understanding of cancer progression and metastasis, holding the potential to significantly foster biomarker discovery.

• Muhammed Talo, Muhammed Talo, University of North Texas
• Serdar Bozdag, Serdar Bozdag, University of North Texas

Motivation: The accurate prediction of drug–target interactions (DTI) is a crucial step in drug discovery, providing a foundation for identifying novel therapeutics. Traditional drug development is both costly and time-consuming, often spanning over a decade. Computational approaches help narrow the pool of compound candidates, offering significant starting points for experimental validation. In this study, we propose Top-DTI framework for predicting DTI by integrating topological data analysis (TDA) with large language models (LLMs). Top-DTI leverages persistent homology to extract topological features from protein contact maps and drug molecular images. Simultaneously, protein and drug LLMs generate semantically rich embeddings that capture sequential and contextual information from protein sequences and drug SMILES strings. By combining these complementary features, Top-DTI enhances predictive performance and robustness.

Results: Experimental results on the public BioSNAP and Human DTI benchmark datasets demonstrate that the proposed Top-DTI model outperforms state-of-the-art approaches across multiple evaluation metrics, including AUROC, AUPRC, sensitivity, and specificity. Furthermore, the Top-DTI model achieves superior performance in the challenging cold-split scenario, where the test and validation sets contain drugs or targets absent from the training set. This setting simulates real-world scenarios and highlights the robustness of the model. Notably, incorporating topological features alongside LLM embeddings significantly improves predictive performance, underscoring the value of integrating structural and sequence-based representations.

Availability: The data and source code of Top-DTI is available at https://github.com/bozdaglab/Top DTI under Creative Commons Attribution Non-Commercial 4.0 International Public License

• Sophia Krix, Sophia Krix, Fraunhofer Institute for Algorithms and Scientific Computing (SCAI)
• Neus Falgàs, Neus Falgàs, Hospital Clínic de Barcelona
• Andrea del Val-Guardiola, Andrea del Val-Guardiola, Hospital Clínic de Barcelona
• Sarah Hücker, Sarah Hücker, Fraunhofer Institute for Toxicology and Experimental Medicine ITEM
• Raquel Sanchez-Valle, Raquel Sanchez-Valle, Hospital Clínic de Barcelona
• Kuti Baruch, Kuti Baruch, ImmunoBrain Checkpoint Ltd.
• Stefan Kirsch, Stefan Kirsch, Fraunhofer Institute for Toxicology and Experimental Medicine ITEM
• Holger Fröhlich, Holger Fröhlich, Fraunhofer Institute for Algorithms and Scientific Computing (SCAI)

In the early stages of Alzheimer’s Disease (AD), significant changes in sleep behavior and immune system dynamics occur prior to the onset of pathological alterations in the brain, ultimately leading to cognitive decline. To investigate the intricate relationships between these changes, we employ multimodal data fusion and modern nonlinear gradient-based Bayesian Network structure learning techniques.

On a molecular biology level our ADIS study (https://adis-project.eu/) employs single-cell RNA sequencing data from peripheral blood mononuclear cells of 75 patients (25 cognitively unimpaired, 25 mild cognitively impaired, 25 with Alzheimer’s dementia), representing cell-type-specific immune pathway alterations, which we embed into a lower dimensional space via a conditional variational autoencoder. In addition, we incorporate neuroinflammatory cytokine levels measured in the cerebral spinal fluid (CSF). We combine these molecular data with questionnaire-based sleep assessments, standardized and app-based tests of different cognitive functions and MRI-derived brain volume measures.

To uncover the interdependencies among these different data modalities, we utilize a recent neural-network based Bayesian Network structure learning method, DAGMA (Directed Acyclic Graphs via M-matrices for Acyclicity). Our approach allows us to observe that in different stages of AD there exist different dependencies between immune system dysregulation, brain volume changes, cognitive function and sleep. Altogether and in agreement with earlier findings, our results implicate that the peripheral immune system plays a pivotal role during the development of AD pathology, opening perspectives for innovative treatment options, which are currently tested in clinical trials by one of our project partners.

• Barry Ryan, Barry Ryan, University of Edinburgh

This study demonstrates the integration of Large Language Model (LLM)-derived clinical text embeddings from the Movement Disorder Society Unified Parkinson’s Disease Rating Scale (MDS-UPDRS) questionnaire with molecular genomics data to enhance interpretability in Parkinson’s disease (PD) classification. By combining genomic modalities encoded using an interpretable biological architecture with a patient similarity network constructed from clinical text embeddings, our approach leverages both clinical and genomic information to provide a robust, interpretable model for disease classification and molecular insights. We benchmarked our approach using the baseline time point from the Parkinson’s Progression Markers Initiative (PPMI) dataset, identifying the Llama-3.2-1B text embedding model on Part III of the MDS-UPDRS as the most informative. We further validated the framework at years 1, 2, and 3 post-baseline, achieving significance in identifying PD associated genes from a random null set by year 2 and replicating the association of MAPK with PD in a heterogeneous cohort. Our findings demonstrate that LLM text embeddings enable robust interpretable genomic analysis, revealing molecular signatures associated with PD progression.

• Philip Kenneweg, Philip Kenneweg, Bielefeld University
• Raghuram Dandinasivara, Raghuram Dandinasivara, Bielefeld University
• Xiao Luo, Xiao Luo, Hunan University
• Barbara Hammer, Barbara Hammer, Bielefeld University
• Alexander Schönhuth, Alexander Schönhuth, Bielefeld University

In this paper, we introduce the first diffusion model designed to generate complete synthetic human genotypes, which, by standard protocols, one can straightforwardly expand into full-length, DNA-level genomes.
The synthetic genotypes mimic real human genotypes without just reproducing known genotypes, in terms of approved metrics. When training biomedically relevant classifiers with synthetic genotypes, accuracy is near-identical to the accuracy achieved when training classifiers with real data. We further demonstrate that augmenting small amounts of real with synthetically generated genotypes drastically improves performance rates. This addresses a significant challenge in translational human genetics: real human genotypes, although emerging in large volumes from genome wide association studies, are sensitive private data, which limits their public availability. Therefore, the integration of additional, insensitive data when striving for rapid sharing of biomedical knowledge of public interest appears imperative.

• Chaoyang Yan, Chaoyang Yan, Centre for Bioinformatics and Intelligent Medicine
• Zhihan Ruan, Zhihan Ruan, Centre for Bioinformatics and Intelligent Medicine
• Songkang Chen, Songkang Chen, Centre for Bioinformatics and Intelligent Medicine
• Yichen Pan, Yichen Pan, Centre for Bioinformatics and Intelligent Medicine
• Xue Han, Xue Han, Centre for Bioinformatics and Intelligent Medicine
• Yuanyu Li, Yuanyu Li, Centre for Bioinformatics and Intelligent Medicine
• Jian Liu, Jian Liu, Centre for Bioinformatics and Intelligent Medicine

Motivation: Fine-grained cellular characterization provides critical insights into biological processes, including tissue development, disease progression, and treatment responses. The spatial organization of cells and the interactions among distinct cell types play a pivotal role in shaping the tumor micro-environment, driving heterogeneity, and significantly influencing patient prognosis. While computational pathology can uncover morphological structures from tissue images, conventional methods are often restricted to identifying coarse-grained and limited cell types. In contrast, spatial transcriptomics-based approaches hold promise for pinpointing fine-grained transcriptional cell types using histology data. However, these methods tend to overlook key molecular signatures inherent in gene expression data.
Results: To this end, we propose a cross-modal unified representation learning framework (CUCA) for identifying fine-grained cell types from histology images. CUCA is trained on paired morphology-molecule spatial transcriptomics data, enabling it to infer fine-grained cell types solely from pathology images. Our model aims to harness the cross-modal embedding alignment paradigm to harmonize the embedding spaces of morphological and molecular modalities, bridging the gap between image patterns and molecular expression signatures. Extensive results across three datasets show that CUCA captures molecule-enhanced cross-modal representations and improves the prediction of fine-grained transcriptional cell abundances. Downstream analyses of cellular spatial architectures and intercellular co-localization reveal that CUCA provides insights into tumor biology, offering potential advancements in cancer research.

• Linan Zhang, Linan Zhang, Ningbo University
• Matthew Lu, Matthew Lu, University of Pittsburgh
• Hatice Osmanbeyoglu, Hatice Osmanbeyoglu, University of Pittsburgh

Glioblastoma (GBM) is the most aggressive primary brain tumor, with poor prognosis and limited treatment options. Its tumor microenvironment (TME) plays a critical role in driving cancer progression, immune evasion, and therapy resistance. Spatial transcriptomics (ST) technologies now allow for the investigation of gene regulation and cell-cell interactions in their spatial context. We developed STAN (Spatially informed Transcription Factor Activity Network), a computational method to infer spot-specific transcription factor (TF) activity from ST data and cis-regulatory information. STAN enables identification of TFs associated with specific cell types, spatial domains, and ligand-receptor signaling events. We further extend this with SPAN (Spatially informed Pathway Activity Network) to predict localized pathway activity. Applying STAN and SPAN to GBM ST datasets (n=26), we uncovered spatial regulatory networks and key ligand-receptor interactions in the TME. Notably, we observed strong correlation between STAN-predicted SOX2 activity and expression of CD44 and VIM, which we validated at the protein level in independent GBM specimens. To identify therapeutic opportunities, we integrated our regulatory maps with Drug2Cell, which links drugs to targetable expression patterns at cellular resolution. This revealed cell type- and region-specific drug-target relationships, nominating compounds with potential to modulate malignant or immunosuppressive cell populations in GBM. Together, STAN, SPAN, and Drug2Cell form a comprehensive framework to connect spatial gene regulation with therapeutic insights. This approach not only advances understanding of GBM biology but is also broadly applicable to other diseases and tissue contexts.

• Yu Wang, Yu Wang, Vanderbilt University Medical Center
• Yuechen Yang, Yuechen Yang, Vanderbilt University
• Jiayuan Chen, Jiayuan Chen, Vanderbilt University
• Ruining Deng, Ruining Deng, Vanderbilt University
• Mengmeng Yin, Mengmeng Yin, Vanderbilt University Medical Center
• Haichun Yang, Haichun Yang, Vanderbilt University Medical Center
• Yuankai Huo, Yuankai Huo, Vanderbilt University
• Shilin Zhao, Shilin Zhao, Vanderbilt University Medical Center

Abstract: We present SpatialPathomicsToolkit, a modular and platform-agnostic toolkit for comprehensive analysis of pathomics features from whole slides image and their integration with spatial transcriptomics data. The toolkit supports diverse data sources including CellProfiler, PySpatial, and AI-based feature extractors, and enables feature summarization, group comparison, dimensionality reduction, and correlation with transcriptomic and clinical data. We evaluated its utility across multiple renal pathology studies involving both human and mouse samples. Our results demonstrate the toolkit’s generalizability and value for spatially resolved, multimodal pathology analysis.

• Marcelo Hurtado, Marcelo Hurtado, Cancer Research Center of Toulouse
• Abdelmounim Essabbar, Abdelmounim Essabbar, Cancer Research Center of Toulouse
• Leila Khajavi, Leila Khajavi, Bioinformatics Department
• Vera Pancaldi, Vera Pancaldi, Cancer Research Center of Toulouse

The tumor microenvironment (TME) plays a key role in cancer development by influencing physiopathological processes. Despite significant progress in understanding this complex system, it remains unclear why some patients respond to specific therapies while others experience recurrence. Computational methods for cell type deconvolution from bulk RNA-seq data have been developed, yet their high feature complexity and variability limit their effectiveness for patient stratification. This project introduces CellTFusion, a novel framework for characterizing TME patient profiles by constructing transcriptional regulatory networks (TRNs) based on inferred transcription factor (TF) activity and cell type deconvolution from bulk RNA-seq data. This approach is able to capture multiple possible cell phenotypes and states within patient samples. We applied CellTFusion to several publicly available cancer datasets, including melanoma, neuroblastoma, lung cancer, and bladder cancer. Using existing algorithms, we inferred TF activity from gene expression data and constructed TF networks from highly correlated features. These networks were integrated with cell type proportion estimates derived from both bulk and single-cell reference signatures to generate cell group scores that reflect their association. Additionally, we incorporated a robust machine learning pipeline to identify if these potential cell states are significantly associated with clinical outcomes, including survival, recurrence, and response to immunotherapy. CellTFusion provides a novel framework to identify clusters of deconvolution features based on regulatory activities that can be used as TME profiles to allow a better patient stratification. It also overcomes the problem of heterogeneity and high-dimensionality of current deconvolution methods by the integration of prior-knowledge networks.

• Satoko Namba, Satoko Namba, Nagoya University
• Michio Iwata, Michio Iwata, Kyushu Institute of Technology
• Shin-Ichi Nureki, Shin-Ichi Nureki, Oita University
• Noriko Yuyama Otani, Noriko Yuyama Otani, Nagoya University
• Yoshihiro Yamanishi, Yoshihiro Yamanishi, Nagoya University

Identifying therapeutic targets for diseases is important in drug discovery. However, the depletion of viable therapeutic targets has a major bottleneck, contributing to the recent stagnation in drug development, especially for rare and orphan diseases.
　Here, we propose a disease signature, TRESOR, which characterizes the functional mechanisms of each disease through genome-wide association study (GWAS) and transcriptome-wide association study (TWAS) data. Based on TRESOR and target perturbation signatures (TGPs)—i.e., gene knockdown and overexpression profiles of target-coding genes—we develop machine learning methods for predicting inhibitory and activatory therapeutic targets for various diseases. TRESOR enables highly accurate identification of target candidate proteins based on inverse correlations between TRESOR and TGPs. Furthermore, Bayesian integrative method combines TRESOR-based inverse correlations and omics-based disease similarities, providing more reliable predictions for rare and orphan diseases with few or no known targets. We make comprehensive predictions for 284 diseases with 4,345 inhibitory target candidates and 151 diseases with 4,040 activatory target candidates. These predictions were validated through literature-based cross-referencing and independently assessed using human cohort data, supporting their therapeutic potential. For instance, multiple endocrine neoplasia, a rare disease with only one known inhibitory target, was predicted to have new candidate targets such as FHL2, whose downregulation correlated with improved survival in independent cohorts. The proposed method was also applied to orphan diseases lacking any known targets, such as tauopathies, identifying promising candidates including RAB1B. Proposed methods are expected to be useful for understanding disease–disease relationships and identifying therapeutic targets for rare and orphan diseases.

• Peter White, Peter White, The Abigail Wexner Research Institute at Nationwide Children's Hospital
• Bimal Chaudhari, Bimal Chaudhari, The Abigail Wexner Research Institute at Nationwide Children's Hospital
• Austin Antoniou, Austin Antoniou, The Abigail Wexner Research Institute at Nationwide Children's Hospital
• David Gordon, David Gordon, The Abigail Wexner Research Institute at Nationwide Children's Hospital
• Ashley Kubatko, Ashley Kubatko, The Abigail Wexner Research Institute at Nationwide Children's Hospital
• Ben Knutson, Ben Knutson, The Abigail Wexner Research Institute at Nationwide Children's Hospital

More than 350 million individuals globally suffer from approximately 10,000 known rare diseases. Despite genomic advances, patients today experience prolonged diagnostic journeys, averaging six years, and roughly 50% remain undiagnosed. This delay often results in inappropriate care, irreversible disease progression, and increased medical costs. To address these challenges, we developed GENIUS, a comprehensive framework targeting patient identification, variant prioritization, and continuous genomic data reanalysis to accelerate diagnoses and improve patient outcomes.

GENIUS integrates three innovative machine learning algorithms: NeoGX identifies undiagnosed patients through phenotypic features extracted via NLP from electronic health records, facilitating timely genetic testing referrals, particularly in NICU settings. CAVaLRi employs an advanced likelihood-ratio framework incorporating variant pathogenicity, phenotype overlap, parental genotypes, and segregation data, effectively prioritizing diagnostic genetic variants amidst noisy phenotype data. PARDIGM automates genomic data reanalysis by continuously updating clinicians as new gene-disease associations emerge.

GENIUS demonstrated remarkable performance, with NeoGX accurately predicting the need for genetic testing (ROC AUC = 0.855), halving testing initiation time from 62 to 31 days. CAVaLRi significantly outperformed existing methods (PR AUC = 0.701), ranking diagnostic variants first in over 70% of cases. PARADIGM, the automated genomic reanalysis component, achieved a 40% diagnostic yield, substantially surpassing conventional methods.

GENIUS exemplifies a scalable computational framework integrating predictive analytics, precise variant prioritization, and dynamic genomic reanalysis. By automating complex diagnostic workflows, GENIUS substantially accelerates diagnosis, optimizes clinical decision-making, and demonstrates the transformative potential of machine learning to advance personalized genomic medicine in rare genetic disorders.

• Yasmin Jolasun, Yasmin Jolasun, McGill University
• Yumin Zheng, Yumin Zheng, McGill University
• Kailu Song, Kailu Song, McGill University
• Jingtao Wang, Jingtao Wang, McGill University
• David H. Eidelman, David H. Eidelman, McGill University
• Jun Ding, Jun Ding, McGill University

Single-cell RNA sequencing (scRNA-seq) offers unparalleled resolution for studying cellular heterogeneity but is costly, restricting its use to small cohorts that often lack comprehensive clinical data, limiting translational relevance. In contrast, bulk RNA sequencing is scalable and cost-effective but obscures critical single-cell insights. We introduce SIDISH, a neural network framework that integrates the granularity of scRNA-seq with the scalability of bulk RNA-seq. Using a Variational Autoencoder, deep Cox regression, and transfer learning, SIDISH identifies High-Risk cell populations while enabling robust clinical predictions from large-cohort data. Its in silico perturbation module identifies therapeutic targets by simulating interventions that reduce High-Risk cells associated with adverse outcomes. Applied across diverse diseases, SIDISH establishes the link between cellular dynamics and clinical phenotypes, facilitating biomarker discovery and precision medicine. By unifying single-cell insights with large-scale clinical data, SIDISH advances computational tools for disease risk assessment and therapeutic prioritization, offering a transformative approach to precision medicine.

• Jonathan Carlson

The fundamental breakthrough of generative AI is the ability to automatically learn, extract, express, and reason over concepts. And not just concepts encoded in human language, but increasingly those encoded in the languages of biomedicine: from images to proteins. In this talk, we will explore the emerging frontier of generative AI in the science and practice of medicine.