Model-agnostic confidence measurement for aggregating multimodal ensemble models in automatic diagnostic systems.

Automatic diagnostic systems (ADSs) have been garnering increased attention because they can alleviate the workload of clinicians by assisting in diagnosis and offering low-cost access to healthcare for people in medically underserved areas. ADS can suggest potential diseases by analyzing a patient's self-report. Previous research on ADS has leveraged diagnostic case data from various patients and medical knowledge to diagnose diseases, with multimodal ensemble methods proving particularly effective. However, the existing multimodal ensemble method combines the probabilities of different models in the aggregating process, which can not properly combine the probabilities that are produced by different criteria. To address these issues, we propose an effective aggregation framework for multimodal ensembles that can properly aggregate model-agnostic confidence scores and predictions from each model. Our framework transforms probability scores from different criteria into unified aggregation rule-based scores and reflects the gap between the probabilities that may be blurred in the aggregation process through the confidence score. In particular, The proposed confidence measurement method employs a post-analysis approach with the developed model or algorithm, making it adaptable in a model-agnostic manner and suitable for multimodal ensemble learning that utilizes heterogeneous prediction results. Our experimental results demonstrate that our framework outperforms existing approaches by more effectively leveraging the strengths of each ensemble member.

Single-molecule analysis of combinatorial epigenomic states in normal and tumor cells.

Proper placement of epigenetic marks on DNA and histones is fundamental to normal development, and perturbations contribute to a variety of disease states. Combinations of marks act together to control gene expression; therefore, detecting their colocalization is important, but because of technical challenges, such measurements are rarely reported. Instead, measurements of epigenetic marks are typically performed one at a time in a population of cells, and their colocalization is inferred by association. Here, we describe a single-molecule analytical approach that can perform direct detection of multiple epigenetic marks simultaneously and use it to identify mechanisms coordinating placement of three gene silencing marks, trimethylated histone H3 lysine 9, lysine 27 (H3K9me3, H3K27me3), and cytosine methylation (mC), in the normal and cancer genome. We show that H3K9me3 and mC are present together on individual chromatin fragments in mouse embryonic stem cells and that half of the H3K9me3 marks require mC for their placement. In contrast, mC and H3K27me3 coincidence is rare, and in fact, mC antagonizes H3K27me3 in both embryonic stem cells and primary mouse fibroblasts, indicating this antagonism is shared among primary cells. However, upon immortalization or tumorigenic transformation of mouse fibroblasts, mC is required for complete H3K27me3 placement. Importantly, in human promyelocytic cells, H3K27me3 is also dependent on mC. Because aberrant placement of gene silencing marks at tumor suppressor genes contributes to tumor progression, the improper dependency of H3K27me3 by mC in immortalized cells is likely to be fundamental to cancer. Our platform can enable other studies involving coordination of epigenetic marks and leverage efforts to discover disease biomarkers and epigenome-modifying drugs.