Tianyu Liu

Motivation: Single-cell foundation models (scFMs) are transforming computational biology by enabling generalizable, task-agnostic representations for versatile single-cell analysis. Despite their progress in facilitating rapid deployment for downstream tasks, off-the-shelf scFMs still have some overlooked concerns: (I) (Pretraining Cost.) Pretrain-based scFMs necessitate pretraining on a vast volume of cells, rendering it draining resources in applications. (II) (Heterogeneous Gap.) Large language model (LLM)-based scFMs ignore the tremendous heterogeneous gap between LLM textual and raw cellular spaces, leading to insufficient capability when facing downstream tasks. Results: To this end, we introduce RAGCell, a versatile single-cell analysis framework that achieves a double-win in both cost-effectiveness and high performance. The success of RAGCell lies in two key aspects: Leveraging LLMs to construct cell-level and feature-level knowledge databases, which serve as supervision signals for training the cell model and significantly reduce the training cost (>pretrain-based scFMs). Aligning cell representations with text embeddings from the bi-level knowledge databases, enabling knowledge transfer from textual spaces to cellular spaces and effectively mitigating the heterogeneous gap (>LLM-based scFMs). Through extensive experiments on five downstream single-cell analysis tasks, we demonstrate that RAGCell achieves outstanding performance compared to state-of-the-art scFMs while operating at less than ∼1/10 the cost of pretrain-based scFMs. Availability and implementation: The source code is available at our supplementary file. Embeddings can be found in https://sites.google.com/yale.edu/scelmolib/home

Background: Alzheimer's disease (AD), the most common age-related neurodegenerative disease, is closely associated with both amyloid-β plaque and neuroinflammation. Two thirds of AD patients are female, and they have a higher disease risk; women with AD have more extensive brain histological changes than men along with more severe cognitive symptoms and neurodegeneration. Objective: This study aimed to determine how sex difference induces structural brain changes and molecular cell vulnerabilities in AD, with a focus on identifying sex-specific transcriptional alterations and genetic risk factors. Methods: We performed single nucleus RNA sequencing on postmortem brains from individuals with AD and age- and sex-matched controls, focusing on the middle temporal gyrus, a cortical brain region strongly affected by the disease, and integrated single nucleus RNA sequencing results with genome-wide association study (GWAS) data using cell type-specific enrichment and generalized gene-set analysis approaches. The analysis pipeline is provided with threshold information. Results: We identified a selectively vulnerable subpopulation of layer 2/3 excitatory neurons that were RORB-negative and CDH9-expressing in both males and females. Disease-associated, but sex-independent, reactive astrocyte signatures were also present. In clear contrast, the microglia signatures of AD brains differed between males and females. Integrating single cell transcriptomic data with results from GWAS, we identified MERTK genetic variation as a candidate novel risk factor for AD selectively in females. Conclusions: Taken together, our single cell atlas of middle temporal gyrus revealed a unique cellular-level view of sex-specific transcriptional changes in AD, illuminating GWAS identification of sex-specific AD genes. These data serve as a rich resource for interrogation of the molecular and cellular basis of AD.

Models are no longer merely trained and then used, but generate experience during use from which future behavior should improve. A deployed assistant may learn user preferences, a coding agent may encounter recurring repository conventions, and an interactive agentic system may receive corrections, tool outcomes, or environmental feedback. The question is therefore not only how to train a capable model, but how a model should keep, update, and reuse what happens during use. We use test-time learning (TTL) to describe a deployment-time learning paradigm that complements pre-training and post-training. TTL occurs when experience encountered during use updates adaptive state that causally shapes a subsequent decision, forming a deployment-time write-read loop. This formulation provides a unifying lens for a fragmented body of work on memory, adaptation, model editing, and self-evolving agents. We formalize the loop and organize qualifying forms of TTL into context-level, representation-level, and parameter-level TTL, according to where adaptive state re-enters the system. This taxonomy makes lines of work that grew up in separate communities comparable as variants of the same problem: how to turn use into durable learning. We close by reviewing how the causal contribution of learned state should be evaluated, how learning accumulates and remains stable over time, and the challenges facing trustworthy learning during deployment.

Agents developed using large language models (LLMs) and multimodal large language models (MLLMs) have brought about revolutionary changes across multiple fields, significantly boost ing productivity and enthusiasm for artificial intelligence (AI). The design of effective AI agents has traditionally been treated as an engineering-centered problem, with most approaches relying heavily on human expertise and empirical trial and error. However, a systematic examination of high-performing AI agents reveals recurring design patterns from which generalizable principles for building agentic systems can be derived. Motivated by this observation, we trace the evo lution of AI agent design, review major categories and applications, identify recurring patterns among prominent AI agents, and synthesize their core architectural components into four func tional contracts: memory and state management, observation-action interface, reasoning, and runtime control. Furthermore, we propose a better agentic AI framework by following five prin ciples: grounding, autonomy, modularity, efficiency, and robustness. Our framework covers the most critical components of the agentic AI ecosystem design and provides advanced guidance for the healthy operation of agents. We conclude by discussing emerging directions and offering recommendations for the development and deployment of future AI agents. In summary, the design of AI agents requires not only engineering ingenuity but also scientific guidance.

Background: Genome-wide polygenic risk scores (PRSs) for coronary artery disease (CAD) aggregate genetic effects across the genome and may obscure biologically distinct mechanisms. We aimed to develop cell-type-specific PRSs (csPRSs) using single-cell RNA sequencing (scRNA-seq) data and investigate their interactions with lipids on CAD risk. Methods: Using publicly available scRNA-seq data from human heart tissue, we identified cell-type-specific genes across 13 major cell types and 64 subpopulations and grouped them into 10 cell clusters. Variants from a CAD genome-wide association study (GWAS) were mapped to cluster-specific genes to construct csPRSs for European-ancestry participants from the UK Biobank (UKB). Interactions between csPRSs and lipid-related phenotypes were evaluated using Cox proportional hazards models and stratified analyses, with significant findings further assessed in an internal validation dataset. Results: Distinct interaction patterns with lipid phenotypes were observed across csPRSs. Low-density lipoprotein (LDL)-related lipid traits, including apolipoprotein B (ApoB), low-density lipoprotein cholesterol (LDL-C), and total cholesterol (cholesterol), primarily interacted with adipocytes (Adip), whereas high-density lipoprotein (HDL) traits interacted with endothelial-mesothelial (EC-Meso), fibroblast (FB), and immune-cell csPRSs. Notably, interactions for Adip csPRSs were replicated in internal validation analyses. Conclusions: Cell-type-specific decomposition of genome-wide PRSs for CAD identified biologically distinct lipid interactions that were not captured by the genome-wide PRS. Adipocyte genetic factors may influence how LDL lipids affect CAD risk. These findings highlight the potential of cell-type-informed PRSs to improve the biological interpretation of PRSs and provide insights into the heterogeneous mechanisms underlying CAD.

AI Scientists are systems that pursue scientific questions by generating, evaluating, and refining evidence across extended research cycles, and they are beginning to move from proof-of-concept demonstrations toward early translational use. Recent systems have autonomously proposed biomedical hypotheses, designed computational pipelines, and produced findings later validated in wet-lab experiments. Health is both the most compelling and the most demanding domain for this paradigm: compelling because the need for faster, more adaptive scientific inquiry is acute across biomedical discovery, clinical research, healthcare delivery, and population health; demanding because failures do not remain confined to model outputs, but can propagate across scientific trajectories, clinical workflows, and the evidence base itself. We therefore argue that AI Scientists in health should be understood not simply as a new application of generative AI, but as an emerging biomedical engineering challenge. Their central risk is trajectory-level failure, a long-horizon failure mode in which errors self-reinforce across multi-stage reasoning cycles rather than appearing as isolated incorrect outputs, and can become embedded in the scientific record and clinical decisions before they are visible enough to correct. Existing frameworks for large language models and AI agents were not designed to detect or contain it. Progress should therefore be judged not primarily by autonomy, but by whether these systems can be engineered, evaluated, and governed as trustworthy collaborators in health. We characterize the emerging landscape of AI Scientists in health, identify five directions of opportunity and six major challenges, and propose the Workflow-driven, Evidence-grounded, Cross-modal, Accountable, Reproducible, and Embodied (WE CARE) framework as a practical roadmap for engineering, evaluating, and governing them.

Recent advances in spatial transcriptomics (ST) have generated an expanding collection of heterogeneous datasets, offering unprecedented opportunities to investigate tissue organizations and functions. However, effective interpretation and integration of data originating from diverse sources and conditions remain a major challenge. We present INSPIRE, a deep-learning method for interpretable, integrative analysis of multiple ST datasets. INSPIRE adopts an adversarial learning strategy with graph neural networks to achieve spatially informed and adaptive data integration. By incorporating non-negative matrix factorization, INSPIRE identifies interpretable spatial factors and associated gene programs that characterize tissue architecture, cell-type organization and biological processes. Across a broad range of applications, INSPIRE demonstrates superior performance in resolving fine-grained biological signals, integrating complementary strengths across technologies, capturing condition-specific variation, uncovering tumor microenvironment heterogeneity, elucidating developmental dynamics and facilitating three-dimensional tissue reconstruction. INSPIRE also scales to extremely large datasets, as demonstrated by applications to Xenium-profiled human breast cancer and Stereo-seq mouse organogenesis datasets.

Foundation Models (FMs) have made significant strides in both industrial and scientific domains. In this paper, we evaluate the performance of FMs for single‐cell sequencing data analysis through comprehensive experiments across eight downstream tasks pertinent to single‐cell data. Overall, the top FMs include scGPT, Geneformer, and CellFM by considering model performances and user accessibility among ten single‐cell FMs. However, by comparing these FMs with task‐specific methods, we found that single‐cell FMs may not consistently excel than task‐specific methods in all tasks, which challenges the necessity of developing foundation models for single‐cell analysis. In addition, we evaluated the effects of hyperparameters, initial settings, and stability for training single‐cell FMs based on a proposed scEval framework, and provide guidelines for pre‐training and fine‐tuning to enhance the performances of single‐cell FMs. Our work summarizes the current state of single‐cell FMs, points to their constraints and avenues for future development, and offers a freely available evaluation pipeline to benchmark new models and improve method development.