Multimodal AI/ML for discovering novel biomarkers and predicting disease using multi-omics profiles of patients with cardiovascular diseases

Cardiovascular diseases (CVDs) are multifactorial diseases, requiring personalized assessment and treatment. The advancements in multi-omics technologies, namely RNA-seq and whole genome sequencing, have offered translational researchers a comprehensive view of the human genome; utilizing this data, we can reveal novel biomarkers and segment patient populations based on personalized risk factors. Limitations in these technologies in failing to capture disease complexity can be accounted for by using an integrated approach, characterizing variants alongside expression related to emerging phenotypes. Designed and implemented data analytics methodology is based on a nexus of orthodox bioinformatics, classical statistics, and multimodal artificial intelligence and machine learning techniques. Our approach has the potential to reveal the intricate mechanisms of CVD that can facilitate patient-specific disease risk and response profiling. We sourced transcriptomic expression and variants from CVD and control subjects. By integrating these multi-omics datasets with clinical demographics, we generated patient-specific profiles. Utilizing a robust feature selection approach, we reported a signature of 27 transcripts and variants efficient at predicting CVD. Here, differential expression analysis and minimum redundancy maximum relevance feature selection elucidated biomarkers explanatory of the disease phenotype. We used Combination Annotation Dependent Depletion and allele frequencies to identify variants with pathogenic characteristics in CVD patients. Classification models trained on this signature demonstrated high-accuracy predictions for CVDs. Overall, we observed an XGBoost model hyperparameterized using Bayesian optimization perform the best (AUC 1.0). Using SHapley Additive exPlanations, we compiled risk assessments for patients capable of further contextualizing these predictions in a clinical setting. We discovered a 27-component signature explanatory of phenotypic differences in CVD patients and healthy controls using a feature selection approach prioritizing both biological relevance and efficiency in machine learning. Literature review revealed previous CVD associations in a majority of these diagnostic biomarkers. Classification models trained on this signature were able to predict CVD in patients with high accuracy. Here, we propose a framework generalizable to other diseases and disorders.