The standard analysis pipeline for single-cell RNA-seq data consists of sequential steps initiated by clustering the cells. An innate limitation of this pipeline is that an imperfect clustering result can irreversibly affect the succeeding steps. For example, there can be cell types not well distinguished by clustering because they largely share the global structure, such as the anterior primitive streak and mid primitive streak cells. If one searches differentially expressed genes (DEGs) solely based on clustering, marker genes for distinguishing these types will be missed. Moreover, clustering depends on many parameters and can often be subjective to manual decisions. To overcome these limitations, we propose MarcoPolo, a method that identifies informative DEGs independently of prior clustering. MarcoPolo sorts out genes by evaluating if the distributions are bimodal, if similar expression patterns are observed in other genes, and if the expressing cells are proximal in a low-dimensional space. Using real datasets with FACS-purified cell labels, we demonstrate that MarcoPolo recovers marker genes better than competing methods. Notably, MarcoPolo finds key genes that can distinguish cell types that are not distinguishable by the standard clustering. MarcoPolo is built in a convenient software package that provides analysis results in an HTML file.

Building trustworthy and transparent image-based medical AI systems requires the ability to interrogate data and models at all stages of the development pipeline: from training models to post-deployment monitoring. Ideally, the data and associated AI systems could be described using terms already familiar to physicians, but this requires medical datasets densely annotated with semantically meaningful concepts. Here, we present a foundation model approach, named MONET (Medical cONcept rETriever), which learns how to connect medical images with text and generates dense concept annotations to enable tasks in AI transparency from model auditing to model interpretation. Dermatology provides a demanding use case for the versatility of MONET, due to the heterogeneity in diseases, skin tones, and imaging modalities. We trained MONET on the basis of 105,550 dermatological images paired with natural language descriptions from a large collection of medical literature. MONET can accurately annotate concepts across dermatology images as verified by board-certified dermatologists, outperforming supervised models built on previously concept-annotated dermatology datasets. We demonstrate how MONET enables AI transparency across the entire AI development pipeline from dataset auditing to model auditing to building inherently interpretable models.

Human leukocyte antigen (HLA) gene variants in the major histocompatibility complex (MHC) region are associated with numerous complex human diseases and quantitative traits. Previous phenome-wide association studies (PheWAS) for this region demonstrated that HLA association patterns to the phenome have both population-specific and population-shared components. We performed MHC PheWAS in the Korean population by analyzing associations between phenotypes and genetic variants in the MHC region using the Korea Biobank Array project data samples from the Korean Genome and Epidemiology Study (KoGES) cohorts. Using this single-population dataset, we curated and analyzed 82 phenotypes for 125 673 Korean individuals after imputing HLA using CookHLA, a recently developed imputation framework. More than one-third of these phenotypes showed significant associations, confirming 56 known associations and discovering 13 novel association signals that were not reported previously. In addition, we analyzed heritability explained by the variants in the MHC region and genetic correlations among phenotypes based on the MHC variants.