Jinze Liu

Bioinformatics stands at a precarious intersection: while biological questions are cutting-edge, the software infrastructure supporting them is often fragile, "ad hoc", and ephemeral. In this paper, we reflect on the state of the practice through a qualitative study of bioinformatics researchers. Our findings reveal a systemic failure: researchers with Computer Science backgrounds struggle to bridge the biological "knowledge gap", while those with Biology backgrounds are bogged down by "installation hell" and fragmented documentation. Furthermore, we observe that while AI tools are increasingly adopted, their usage remains superficial—restricted to generating isolated code snippets rather than addressing structural engineering debt. We argue that the current solution—expecting scientists to act as expert systems engineers—is unscalable. We propose a vision of Bio-Aware Software Engineering (BASE): an AI-driven paradigm where "invisible" agents observe exploratory analysis and automatically synthesize the necessary software infrastructure (containers, regression tests, and documentation) by leveraging domain-specific knowledge graphs. This vision points toward more reproducible and maintainable bioinformatics workflows.

Accurate cell segmentation in high-plex imaging remains challenging because no single biomarker channel captures the full structural diversity of cells, and many practical segmentation backbones are trained on a small, fixed-channel input convention. In this work, we study channel complementarity under the standard three-channel input convention of Cellpose-SAM using a systematic fused-evaluation framework with combination-specific operating-point calibration. On a primary healthy salivary gland region, we show that multi-channel gains are selective but meaningful: the strongest pair improves over the single-channel baseline, and the strongest triplet further improves over the best pair. On a second held-out diseased tissue region, we identify strong auxiliary combinations centered on EPCAM, and the triplet combination (DAPI, EPCAM, CD31) remains beneficial across both regions. To test whether these gains are artifacts of conservative post-processing, we additionally perform controlled-merge sensitivity analysis across the merge hyperparameter grid. These analyses do not reproduce the fused triplet gains, suggesting that the observed improvements arise from joint multi-channel inference rather than from the merge heuristic itself. Together, our results show that higher-order channel complementarity is real but selective under the standard three-channel input convention of Cellpose-SAM, and that reproducible triplet benefits can be identified through systematic search.

Oral inflammatory diseases affect nearly half of all humans, yet mechanisms underlying rapidly-destructive inflammation remain poorly understood. We compared peri-implantitis with moderate- and high-grade periodontitis using integrated microbial and single-cell sequencing (>967,169-cells; single-cell RNA-seq, spatial proteotranscriptomics). Laser capture microdissection with compartmental microbiome analysis revealed reduced bacterial load and diversity in peri-implantitis. Expansion of the Human Periodontal Atlas with peri-implantitis single-cell RNA-seq data (36-samples; 121,395 cells) identified CD34⁺ vascular endothelial cell (VEC) rarefaction and oxidative stress, hypoxia, and NAD⁺ metabolism-associated transcriptional programs enriched in a TNFRSF6B⁺/ICAM1⁺ post-capillary venule (PC-VEC) subpopulation. NAD⁺-consuming ectoenzyme CD38 was selectively enriched and orthogonally confirmed by spatial transcriptomics (6-samples; 283,377-cells) and proteomics (23-samples; 562,397-cells). Spatial neighborhood analyses demonstrated CD38⁺-high PC-VEC expansion, closer proximity, and higher IL16–CD4 T cell signaling in peri-implantitis. Matched high-grade periodontitis biopsies confirmed spatially restricted CD38⁺-VECs despite similar microbial burden, identifying endothelial vasculopathy underlying rapidly advancing oral inflammation and a potential therapeutic axis.

Oral inflammatory diseases affect nearly half of the global population. Among them, newly defined peri-implantitis and high-grade periodontitis represent rapidly advancing inflammatory disease types, marked by relatively rapid tissue destruction. Despite their prevalence, the cell mechanisms and spatial architecture driving this severity remain poorly understood. Focusing first on peri-implantitis versus low- and moderate-grade periodontitis, we applied microbial profiling, single-cell RNA sequencing (scRNA-seq), and spatial proteomics (sp-proteomics) to uncover shared pathogenic programs linked to accelerated niche breakdown. Furthermore, to preserve spatial fidelity, each tissue was anatomically orientated along the tooth- or implant-epithelial interface, analogous sites of disease origination. Laser capture microdissection followed by microbiome analysis of unique tissue compartments revealed reduced bacterial load and diversity in peri-implantitis stroma. We then expanded our version-1 Human Periodontal Atlas by integrating newly generated peri-implantitis scRNAseq data (36-total samples; 121395-cells), revealing widespread transcriptional alterations, including oxidative stress, hypoxic, and NAD+ metabolism-associated signatures, primarily in a subpopulation of TNFRSF6B+/ICAM1+ post-capillary venules. We then performed high-resolution sp-proteomics (15-total samples; 337260-cells) and analyzed VEC states and associated neighborhoods via AstroSuite using newly developed tri-wise spatial analysis. This revealed CD34+-VEC loss and CD38+-VEC expansion almost exclusively in peri-implantitis. We extended this analysis to high-grade periodontitis. Mucosal biopsies from four lesion-affected and four unaffected sites within the same individuals (1:1 matched; 8 samples; 225137-cells) again demonstrated spatially restricted CD38+-VEC remodeling exclusively in affected tissues, with similar vasculopathy front patterning. The findings nominate spatially distinct vasculopathy patterning as a hallmark of rapidly advancing oral inflammation and a targetable therapeutic axis.

Lung cancer, the leading cause of cancer-related death in the United States, remains a significant public health burden. One of the most personalized treatments to date uses a patient’s own tumor-infiltrating lymphocytes (TILs) as a cellular therapy to target the tumor. Endogenous TIL activity is also essential for immune checkpoint inhibitor (ICI) effectiveness. However, responses to immunotherapies vary due to immune-suppressive tumor microenvironments (TMEs) and limited antigen presentation. In this study, we computationally examine cell-cell signaling and transcriptional activity using single-cell RNA sequencing of lung cancer treated by inhibiting methyltransferase EZH2. We show that EZH2 inhibition shifts the TME to immunogenic signaling patterns conducive to increased T cell response, including antigen presentation and homing. T cells also showed more stem-like phenotypes. Importantly, inferred EZH2 activity showed it can still perform non-methyltransferase duties that are vital for T cell differentiation. Lastly, we tested our findings with patient-derived NSCLC TILs and found that transient EZH1/2 inhibition with valemetostat prior to harvest increases the CD8:CD4 ratio and T cell reactivity, but preferentially expands effector memory rather than central memory populations. These results indicate that transitory EZH2 inhibition could improve immunotherapies for lung cancer patients, with additional optimization needed to maximize TIL therapeutic benefit.

Identifying cell types and states remains a time-consuming, error-prone challenge for spatial biology. While deep learning increasingly plays a role, it is difficult to generalize due to variability at the level of cells, neighborhoods, and niches in health and disease. To address this, we develop TACIT, an unsupervised algorithm for cell annotation using predefined signatures that operates without training data. TACIT uses unbiased thresholding to distinguish positive cells from background, focusing on relevant markers to identify ambiguous cells in multiomic assays. Using five datasets (5,000,000 cells; 51 cell types) from three niches (brain, intestine, gland), TACIT outperforms existing unsupervised methods in accuracy and scalability. Integrating TACIT-identified cell types reveals new phenotypes in two inflammatory gland diseases. Finally, using combined spatial transcriptomics and proteomics, we discover under- and overrepresented immune cell types and states in regions of interest, suggesting multimodality is essential for translating spatial biology to clinical applications.

Background: The performance of a classification algorithm eventually reaches a point of diminishing returns, where the additional sample added does not improve the results. Thus, there is a need to determine an optimal sample size that maximizes performance while accounting for computational burden or budgetary concerns. Objective: This study aimed to determine optimal sample sizes and the relationships between sample size and dataset-level characteristics over a variety of binary classification algorithms. Methods: A total of 16 large open-source datasets were collected, each containing a binary clinical outcome. Furthermore, 4 machine learning algorithms were assessed: XGBoost (XGB), random forest (RF), logistic regression (LR), and neural networks (NNs). For each dataset, the cross-validated area under the curve (AUC) was calculated at increasing sample sizes, and learning curves were fit. Sample sizes needed to reach the observed full–dataset AUC minus 2 points (0.02) were calculated from the fitted learning curves and compared across the datasets and algorithms. Dataset–level characteristics, minority class proportion, full–dataset AUC, number of features, type of features, and degree of nonlinearity were examined. Negative binomial regression models were used to quantify relationships between these characteristics and expected sample sizes within each algorithm. A total of 4 multivariable models were constructed, which selected the best-fitting combination of dataset–level characteristics. Results: Among the 16 datasets (full-dataset sample sizes ranging from 70,000-1,000,000), median sample sizes were 9960 (XGB), 3404 (RF), 696 (LR), and 12,298 (NN) to reach AUC stability. For all 4 algorithms, more balanced classes (multiplier: 0.93-0.96 for a 1% increase in minority class proportion) were associated with decreased sample size. Other characteristics varied in importance across algorithms—in general, more features, weaker features, and more complex relationships between the predictors and the response increased expected sample sizes. In multivariable analysis, the top selected predictors were minority class proportion among all 4 algorithms assessed, full–dataset AUC (XGB, RF, and NN), and dataset nonlinearity (XGB, RF, and NN). For LR, the top predictors were minority class proportion, percentage of strong linear features, and number of features. Final multivariable sample size models had high goodness-of-fit, with dataset–level predictors explaining a majority (66.5%-84.5%) of the total deviance in the data among all 4 models. Conclusions: The sample sizes needed to reach AUC stability among 4 popular classification algorithms vary by dataset and method and are associated with dataset–level characteristics that can be influenced or estimated before the start of a research study.

The oral cavity is a critical barrier with immunosurveillance capabilities. A detailed understanding of its cellular, molecular, and spatial architecture is essential for advancing precision medicine across aerodigestive tissues. Here, we present the first integrated atlas of human adult oral and craniofacial tissues, derived from single-cell RNA sequencing of ∼250,000 cells from 70 samples across 13 niches, including salivary glands and oral mucosae. Complementary spatial atlases using a 40-plex protein and 300-plex transcriptomics panels were developed, covering >2 million cells. Using our new AstroSuite package, we identified 8 tissue cellular neighborhoods (TCNs) and 17 multicellular interaction modules (MCIMs) across gland and mucosae in health. Diverse fibrovascular TCNs emerged as centralized hubs, coordinating immune cell migration and activation. Mucosal hubs within stroma displayed higher immune cell density, diversity, and activation than glands, underscoring differences in functional architecture. In keratinized and nonkeratinized mucosae, common fibrovascular hubs, and MCIMs revealed a specialized biogeography of immunoregulation. Under chronic inflammation, mucosal TCNs were fragmented, and MCIMs involving fibroblasts and immune cells were reorganized and expanded. Focused analysis of tertiary lymphoid structures (TLS) and fibroblasts identified TLS-associated MCIMs and drug candidates targeting these spatial interactions, offering a foundational framework for broadly dissecting mucosal immunity to guide precision medicine approaches. KEY TERMS AstroSuite: A suite of bioinformatics tools, including TACIT, Constellation, STARComm, and Astrograph, used for identifying and analyzing spatial features, cell types, and interactions in tissues. Astrograph: A visualization tool within AstroSuite for exploring cellular neighborhoods and spatial relationships within tissues. Cell Ontology: A structured, hierarchical classification system that provides a controlled vocabulary for cell types based on distinct properties like morphology, lineage, function, and molecular markers. Cell Recipe: A pre-defined set of markers used to specify cell identities and states by combining proteins or transcripts. Constellation: An integration and clustering algorithm within AstroSuite that identifies Tissue Cellular Neighborhoods by dividing tissue into subregions, aiding in scalable whole-slide imaging analyses. Fibrovascular Hubs: Centralized regions where fibroblasts and vascular endothelial cells within defined Tissue Cellular Neighborhoods coordinate with immune cells. Heterotypic Interactions: Predicted receptor-ligand interactions occurring between different cell types, such as immune cells and structural cells. Homotypic Interactions: Predicted receptor-ligand interactions between cells of the same type. Multicellular Interaction Modules (MCIMs): Spatial clusters where multiple cell types engage in receptor-ligand interactions—conceptually akin to Tissue Cellular Neighborhoods. Receptor-Ligand Density Scoring: A method in STARComm used to measure the density of receptor-ligand interactions within tissue regions, predicting areas of high cell-cell communication. Spatial Drug Targeting: The approach of identifying druggable targets within specific tissue regions, guided by spatial analysis of receptor-ligand interactions within identified regions of interest. STARComm: A computational tool in AstroSuite for analyzing spatial data, focusing on quantifying receptor-ligand interactions and communication densities in tissues to discover single or Multicellular Interaction Modules. Structural Immunity: The concept that structural cells, like fibroblasts, contribute to immune regulation through interactions with immune cells to maintain homeostasis. TACIT: A tool within AstroSuite used to identify cell types and interactions in spatial data, supporting neighborhood analysis and visualization. Tiered Cell Annotation: A classification system that organizes cells into broad groups, refining to specific cell types for in-depth tissue analysis—i.e., higher tiered ontologies are more generally labeled. Tissue Cellular Neighborhoods (TCNs): Spatially organized groups of cells within a tissue that are predicted to interact with one another in specific patterns or proximities. These neighborhoods represent the local microenvironments where different cell types coordinate biological processes— conceptually akin to or Multicellular Interaction Modules.

BACKGROUND: Dysregulated metabolism of bioactive sphingolipids, including ceramides and sphingosine-1-phosphate, has been implicated in cardiovascular disease, although the specific species, disease contexts, and cellular roles are not completely understood. Sphingolipids are produced by the serine palmitoyltransferase enzyme, canonically composed of 2 subunits, SPTLC1 (serine palmitoyltransferase long chain base subunit 1) and SPTLC2 (serine palmitoyltransferase long chain base subunit 2). Noncanonical sphingolipids are produced by a more recently described subunit, SPTLC3 (serine palmitoyltransferase long chain base subunit 3). METHODS: The noncanonical (d16) and canonical (d18) sphingolipidome profiles in cardiac tissues of patients with end-stage ischemic cardiomyopathy and in mice with ischemic cardiomyopathy were analyzed by targeted lipidomics. Regulation of SPTLC3 by HIF1α under ischemic conditions was determined with chromatin immunoprecipitation. Transcriptomics, lipidomics, metabolomics, echocardiography, mitochondrial electron transport chain, mitochondrial membrane fluidity, and mitochondrial membrane potential were assessed in the cSPTLC3 ^KO transgenic mice we generated. Furthermore, morphological and functional studies were performed on cSPTLC3 ^KO mice subjected to permanent nonreperfused myocardial infarction. RESULTS: Herein, we report that SPTLC3 is induced in both human and mouse models of ischemic cardiomyopathy and leads to production of atypical sphingolipids bearing 16-carbon sphingoid bases, resulting in broad changes in cell sphingolipid composition. This induction is in part attributable to transcriptional regulation by HIF1α under ischemic conditions. Furthermore, cardiomyocyte-specific depletion of SPTLC3 in mice attenuates oxidative stress, fibrosis, and hypertrophy in chronic ischemia, and mice demonstrate improved cardiac function and increased survival along with increased ketone and glucose substrate metabolism utilization. Depletion of SPTLC3 mechanistically alters the membrane environment and subunit composition of mitochondrial complex I of the electron transport chain, decreasing its activity. CONCLUSIONS: Our findings suggest a novel essential role for SPTLC3 in electron transport chain function and a contribution to ischemic injury by regulating complex I activity.