• Tom Toner, Tom Toner, Queen's Univerity Belfast
• Rashi Pancholi, Rashi Pancholi, Queen's Univerity Belfast
• Tanya Sabwa, Tanya Sabwa, Queen's Univerity Belfast
• Paul M, Paul M, Queen's Univerity Belfast
• Thorsten Forster, Thorsten Forster, LifeArc
• Helen Coleman, Helen Coleman, Queen's Univerity Belfast
• Ian Overton, Ian Overton, Queen's University Belfast

Integration of data from multiple domains can greatly enhance the quality and applicability of knowledge generated in analysis workflows. However, working with health data is challenging, requiring careful preparation in order to support meaningful interpretation. We developed an R package for electronic health data preparation, “eHDPrep” (Gigascience 2023;12:giad030, https://cran.r-project.org/package=eHDPrep) demonstrated upon a multimodal colorectal cancer dataset (661 patients, 155 variables; Colo-661); a further demonstrator is taken from The Cancer Genome Atlas (459 patients, 94 variables; TCGA-COAD). eHDPrep offers user-friendly methods for quality control, including internal consistency checking and redundancy removal with information-theoretic variable merging (Figures 1, 2). eHDPrep also facilitates numerical encoding, variable extraction from free text, and completeness analysis. Semantic enrichment functionality can generate new informative “meta-variables” according to ontological common ancestry, demonstrated with SNOMED CT and the Gene Ontology (Figure 3).
We deployed variational autoencoders with a complex loss function evaluating reconstruction and clustering on the above data and whole-slide tumour images to discover phenotypes and candidate cohorts for more effective molecular stratification (Figures 4-6). Phenotypes represent novel combinations of features across tumour pathology, standard clinical parameters, lifestyle and demographic variables. Molecular stratification within these novel phenotypes seeks to develop new clinical tools for precision oncology.

• Mohammad Al Olaimat, Mohammad Al Olaimat, University of North Texas
• Serdar Bozdag, Serdar Bozdag, University of North Texas

Motivation: Electronic Health Records (EHR) represent a comprehensive resource of a patient's medical history. EHR are essential for utilizing advanced technologies such as deep learning (DL), enabling healthcare providers to analyze extensive data, extract valuable insights, and make precise and data-driven clinical decisions. DL methods such as Recurrent Neural Networks (RNN) have been utilized to analyze EHR to model disease progression and predict diagnosis. However, these methods do not address some inherent irregularities in EHR data such as irregular time intervals between clinical visits. Furthermore, most DL models are not interpretable. In this study, we propose two interpretable DL architectures based on RNN, namely Time-Aware RNN (TA-RNN) and TA-RNN-Autoencoder (TA-RNN-AE) to predict patient’s clinical outcome in EHR at next visit and multiple visits ahead, respectively. To mitigate the impact of irregular time intervals, we propose incorporating time embedding of the elapsed times between visits. For interpretability, we propose employing a dual-level attention mechanism that operates between visits and features within each visit.
Results: The results of the experiments conducted on Alzheimer’s Disease Neuroimaging Initiative (ADNI) and National Alzheimer’s Coordinating Center (NACC) datasets indicated superior performance of proposed models for predicting Alzheimer’s Disease (AD) compared to state-of-the-art and baseline approaches based on F2 and sensitivity. Additionally, TA-RNN showed superior performance on Medical Information Mart for Intensive Care (MIMIC-III) dataset for mortality prediction. In our ablation study, we observed enhanced predictive performance by incorporating time embedding and attention mechanisms. Finally, investigating attention weights helped identify influential visits and features in predictions.

• Panagiotis Nikolaos Lalagkas, Panagiotis Nikolaos Lalagkas, University of Massachusetts Lowell
• Rachel Melamed, Rachel Melamed, University of Massachusetts Lowell

The success of drugs targeting disease genes is widely acknowledged. However, identifying causal genes for common complex diseases remains a non-trivial task. This necessitates innovative approaches to accelerate complex disease drug discovery. We have previously shown that clinical associations between Mendelian and complex diseases can inform complex disease drug discovery due to pleiotropic effects of Mendelian genes. Here, we extend our approach to exploit clinical associations between pairs of complex diseases for drug discovery. We hypothesize that pleiotropic genes shared between a complex disease and its predisposing diseases can help us discover new uses for drugs currently approved only for the predisposing diseases. To test our hypothesis, we start with breast cancer, a well-studied and highly prevalent disease. We compile a list of six traits known to increase breast cancer risk (predisposing diseases), such as depression, high LDL, and type 2 diabetes. Using GWAS summary statistics and local genetic correlation analysis, we find a total of 84 genomic loci harboring mutations with positively correlated effects between breast cancer and each predisposing disease. These loci contain 202 protein-coding genes (shared genes). Using a network biology approach, for each disease pair, we connect drugs already indicated for the predisposing disease to its shared genes with breast cancer and identify drug repurposing candidates for breast cancer. Finally, we show that our list of candidate drugs is enriched for currently investigated and indicated drugs for breast cancer. Our findings suggest a novel way to accelerate drug discovery for complex diseases by leveraging shared genetics.

• Emily Hoskins, Emily Hoskins, Comprehensive Cancer Center and James Cancer Hospital
• Raven Vella, Raven Vella, Comprehensive Cancer Center and James Cancer Hospital
• Julie Reeser, Julie Reeser, Comprehensive Cancer Center and James Cancer Hospital
• Michele Wing, Michele Wing, Comprehensive Cancer Center and James Cancer Hospital
• Eric Samorodnitsky, Eric Samorodnitsky, Comprehensive Cancer Center and James Cancer Hospital
• Altan Turkoglu, Altan Turkoglu, Comprehensive Cancer Center and James Cancer Hospital
• Leah Stein, Leah Stein, Comprehensive Cancer Center and James Cancer Hospital
• Elizabeth Breuning, Elizabeth Breuning, The Bioinformatics Program

Objective: Head and neck cancer (HNC) is the seventh most common cancer worldwide, with a 5-year survival rate of ~50%. The only existing genomic biomarker that guides targeted therapies in HNC is oncogenic HRAS mutations. Gene fusions are clinically targetable, genomic events that involve chromosomal rearrangement, resulting in aberrant function. Here we describe the biological and clinical impact of oncogenic fusions in a combined dataset of HNC. Methods: We evaluated RNA sequencing data from HNCs from the Oncology Research Information Exchange Network (ORIEN, n=1,540), The Cancer Genome Atlas (TCGA, n=528), and other published studies (n=588). We utilized STAR-Fusion and Arriba to detect gene fusions from RNAseq data. Results: Leveraging our combined cohort of 2,666 tumors with RNAseq, we identified 74 cases (2.8%) harboring a clinically relevant gene fusion. The most common fusions involved FGFR3 (N=19), EGFR (n=10), and FGFR2 (n=5). We observed significant gene overexpression in fusion-positive samples with respect to their gene fusion partner (p<0.001). Intrigued by the EGFR fusions that we uncovered, which have not previously been described in head and neck cancers, we further assessed the structure and breakpoints in these fusions. In ORIEN, 4/5 gene fusions harbored the same breakpoint in EGFR with a gene fusion structure found to be successfully clinically targetable in lung cancer. Conclusions: Our results demonstrate that oncogenic gene fusions are prevalent in HNC, often lead to overexpression of the oncogene fusion partner, and are clinically relevant. Our results provide expanded therapeutic opportunities for patients with HNC.

• Chaoyang Yan, Chaoyang Yan, College of Computer Science
• Jialiang Sun, Jialiang Sun, College of Computer Science
• Yiming Guan, Yiming Guan, College of Computer Science
• Jiuxin Feng, Jiuxin Feng, College of Computer Science
• Hong Liu, Hong Liu, The Second Surgical Department of Breast Cancer
• Jian Liu, Jian Liu, College of Computer Science

Motivation: HER2 status identification enables physicians to assess the prognosis risk and determine the treatment schedule for patients. In clinical practice, pathological slides serve as the gold standard, offering morphological information on cellular structure and tumoral regions. Computational analysis of pathological images has the potential to discover morphological patterns associated with HER2 molecular targets and achieve precise status prediction. However, pathological images are typically equipped with high-resolution attributes, and HER2 expression in breast cancer images often manifests the intratumoral heterogeneity.
Results: We present a phenotype-informed weakly-supervised multiple instance learning architecture (PhiHER2) for the prediction of the HER2 status from pathological images of breast cancer. Specifically, a hierarchical prototype clustering module is designed to identify representative phenotypes across whole slide images. These phenotype embeddings are then integrated into a cross-attention module, enhancing feature interaction and aggregation on instances. This yields a prototype-based feature space that leverages the intratumoral morphological heterogeneity for HER2 status prediction. Extensive results demonstrate that PhiHER2 captures a better WSI-level representation by the typical phenotype guidance and significantly outperforms existing methods on real-world datasets. Additionally, interpretability analyses of both phenotypes and WSIs provide explicit insights into the heterogeneity of morphological patterns associated with molecular HER2 status.