Luyi Tian

Predicting single-cell transcriptional responses to perturbations is central to dissecting gene regulation and accelerating therapeutic design, yet the field lacks a rigorous, task-spanning assessment of model behavior. We present a large-scale benchmark of 13 representative methods and baselines across 25 datasets spanning diverse perturbation modalities and species, including two primary immune-cell drug-response resources. We evaluated three core tasks—generalization to unseen single-gene perturbations, prediction of combinatorial interactions, and transfer across cell types—using 24 metrics covering expression-level accuracy, relative changes, differential expression (DE) recovery, and distributional similarity. Across tasks, performance depended strongly on perturbation effect size and evaluation perspective: Expression-level agreement was the highest for small-effect perturbations resembling controls, whereas delta- and DE-based metrics improved with larger effects, providing clearer signals. Models shared a conservative bias, with fine-tuned foundation models compressing variance and underestimating synergistic effects in combinations. PerturbNet showed superior recovery of DE signatures in Tasks 1 and 2, while no method consistently generalized across cell types in Task 3, where biological consistency dominated outcomes. This benchmark establishes current methodological limits, clarifies that different metrics probe distinct biological signals rather than redundant summaries of the same prediction problem, and provides a foundation for developing virtual-cell models that more faithfully capture heterogeneous perturbation responses.

Artificial-intelligence agents propose drug-discovery hypotheses faster than experiments can test them, yet their conclusions are rarely verified, against the underlying biology, the predicted perturbation, or the agent's own scoring logic. We close this verification gap with an agentic framework built on three verifiers. First, PACE, a phenotype verifier, resolves immune aging into ten directionally scored, cell-type-resolved gene-set modules, selected for cross-cohort stability across four PBMC cohorts, and outperforms five established aging clocks in an independent in-house aging cohort of 434 elderly donors. Second, CellQ, a virtual-cell verifier built with multi-modal LLM, compresses each single-cell transcriptome into eight discrete tokens aligned to a language model's vocabulary through residual vector quantization; it attains state-of-the-art perturbation prediction and uniquely resolves the weak, module-level shifts that differential-expression recovery misses. Third, an Analyzer-Planner-Auditor agent verifies its own scoring logic: screening 110 compounds in primary human PBMCs, it found aged-down modules more reversible than aged-up modules and revised its objective from an equal-weight mean to a balance-constrained minimum, a self-correction that generalized to an independent 13-compound T-cell assay. By verifying its predictions and its own objective against experiment, the framework points beyond hypothesis-generating AI toward self-correcting AI scientists whose objectives could continuously evolve.

A major challenge in single-cell RNA sequencing (scRNA-seq) analysis is recovering biologically meaningful cell ontology trees and conserved gene modules across datasets. Data integration and batch-effect correction methods have enabled effective analyses of multiple datasets but often fail to disentangle cell states in heterogeneous samples, such as cancer and the immune system. Here, we present Super Single-Cell Clustering (SuperSCC), a computational framework that utilizes machine learning models to discover cell identities and gene modules across multiple datasets without the need for data integration. Notably, SuperSCC can be implemented at both the cell lineage and cell state levels, thereby allowing the creation of hierarchies of cell programs with specific cell identities and gene modules. This information can be used to identify shared rare populations across datasets regardless of batch effects and has advantages for mapping cell labels from reference to query datasets. We used SuperSCC to perform atlas-level data analysis with more than 90 datasets and built cell state maps of complex tissues, such as the human lung, in healthy and diseased states. SuperSCC outperforms existing approaches in identifying cellular contexts, achieves higher annotation accuracy, and identifies gene modules that indicate conserved immune cell statuses in the lung microenvironment.

The early organogenesis stage is a critical phase of embryogenesis that lays the foundation for organ development, and is characterized by dynamic and spatially organized transcriptional programs. However, limited spatial transcriptomic information has constrained our understanding of early primate organogenesis. Here we present a comprehensive three-dimensional (3D) spatial transcriptomic atlas of cynomolgus monkey embryos at Carnegie stages (CS) 9 and 10, capturing key morphogenetic events including cardiogenesis, gut tube regionalization, neurulation, axial mesendoderm patterning and early somitogenesis. Using high-resolution spatial transcriptomics and 3D reconstruction, we identify spatially defined lineage domains across germ layers and resolve regionally restricted gene expression, transcription factor activity, and signalling landscapes along major embryonic axes, exemplified by the emergence of dorsoventrally patterned spinal cord subpopulations during neurulation. Cross-species comparisons with human and mouse datasets reveal conserved and species-biased transcriptional programs. Together, this atlas provides a foundational reference for studying early primate development.

Integrating continuous single-cell gene expression profiles with discrete-token large language models (LLMs) remains an open challenge: text-based methods are token-inefficient and discard quantitative precision, continuous embeddings preclude autoregressive generation, and single-codebook vector quantization cannot stratify biological scales. We present RVQ-Alpha , an end-to-end framework that bridges this modality gap through three contributions: a Residual Vector Quantization (RVQ) tokenizer that compresses each cell into a fixed 10-token sequence via eight residual codebooks embedded directly in the LLM vocabulary, requiring 3.4× fewer tokens than prior discrete methods and enabling bidirectional unification of cell interpretation and generation; scCoT-Synth , a teacher–student engine that grounds newly added biological tokens through evidence-before-conclusion reasoning, using the language modeling objective as the cross-modal alignment signal without a separate projection network; and (iii) a Fact-Aware RLVR system combining an ontology-grounded answer judge with saliency-weighted verification of biological claims against actual expression data, under dynamic gating that conditions hallucination suppression on task competence. Built on Qwen3-4B and trained via continued pretraining, supervised fine-tuning, and reinforcement learning, RVQ-Alpha substantially improves out-of-distribution generalization and rare-cell recognition across eight held-out datasets; ablations confirm that evidence-first grounding reduces hallucination more than fivefold.

ALS shows marked clinical heterogeneity, yet much real-world evidence remains trapped in unstructured reports. Here we introduce MEDSTREM, a large-language-model (LLM)–based agent that converts patient-sourced document images into standardized longitudinal electronic health records, enabling bottom-up cohort building and linkage to trials and multi-omics. By applying MEDSTREM to clinical report images from 8,298 individuals collected via AskHelpU and harmonizing with PRO-ACT and Answer ALS, we generated 17,602 standardized records and multi-omics profiles from 940 induced motor neuron lines. Progression modelling resolved five subtypes and a continuous degeneration score with interpretable anchors: hand-grip strength and forced vital capacity tracked functional loss, and malnutrition emerged as a modifiable correlate. Across RNA-seq and ATAC-seq, clinical severity is aligned with suppression of cell-cycle programmes, declining histone-gene activity and genome-wide chromatin opening, suggesting distinct epigenetic trajectories. These findings establish an agentic AI framework that turns unstructured clinical records into mechanistic insight and links them to multi-omics, reframing ALS studies from top-down, trial-centric analyses to a bottom-up, patient-sourced approach that reveals actionable heterogeneity.

Long-read single-cell RNA-sequencing enables the profiling of RNA isoform expression and alternative splicing at single cell resolution. However, diverse single-cell technologies and sparse isoform data demand flexible and accurate analysis tools. We introduce FLAMESv2 , a highly modular and protocol-agnostic R/Bioconductor package for long-read single-cell RNA-seq data analysis. FLAMESv2 supports a wide range of single-cell and spatial protocols, is highly configurable, scales to allow multi-sample analysis and provides versatile visualisation and analysis outputs. We demonstrate its compatibility with both droplet-based and combinatorial barcoding single-cell methods, as well as spatial transcriptomics workflows. Benchmarking confirms FLAMESv2 achieves field-leading performance across key analysis tasks. Applying FLAMESv2 to in vitro differentiation of stem cells into neurons, we identify cell-types, differentiation trajectories, expression of annotated and novel isoforms and isoform expression diversity and heterogeneity within individual cells. FLAMESv2 provides a comprehensive, flexible approach to analysing long-read single-cell RNA-sequencing, unlocking this powerful methodology for RNA isoform characterisation.

Spatial transcriptomics technology has developed rapidly in recent years, with various sequencing-based platforms such as 10x Visium, Slide-seq and Stereo-seq becoming widely used by researchers. Each platform brings its own set of protocols and customised data analysis pipelines which presents challenges when the goal is to obtain uniformly preprocessed data that is conveniently formatted for downstream analysis. To address the need for simpler, open-source solutions that deal with sequencing-based spatial transcriptomics (sST) data from different platforms, we present stPipe , a comprehensive and modular pipeline for analysing sST data. stPipe handles various analysis steps including (i) data processing from raw paired end FASTQ files to create a spatially resolved gene count matrix; (ii) the collation of relevant quality control metrics during preprocessing to ensure unwanted artefacts can be filtered from further analysis; and (iii) the adoption of standardised data storage containers to allow results to be easily passed on to a wide range of downstream analysis packages tailored to different goals (such as clustering, cell-cell communication analysis and differential expression analysis). stPipe is implemented as an R/Bioconductor package that builds upon functionality in the scPipe software, and offers a flexible preprocessing pipeline that can manage data from all current main-stream sST plaforms. A key use case for stPipe is in methods benchmarking, and we demonstrate how the uniform processing of sST data collected on reference tissue samples from the cadasSTre and SpatialBenchVisium projects is made easier, allowing comparisons between different technology platforms and downstream analysis tools. Our framework thus aims to advance the standardization and optimization of spatial transcriptomics analyses, fostering collaboration and innovation within the research community.

A major challenge in scRNAseq analysis is how to recover the biologically meaningful cell ontology tree and conserved gene modules across datasets. Data integration and batch-effect correction have been the key to effectively analyze multiple datasets, but often fail to disentangle cell states in heterogeneous samples, such as in cancer and the immune system. Here we present super single cell clustering (SuperSCC), a novel algorithm that utilizes machine-learning models to discover cell identities and gene modules from multiple datasets without the need of data integration. Of note, SuperSCC can be implemented both in cell lineage and cell state level, thereby building the hierarchy of cell programs with specific cell identity and gene modules. Such information has the great potential to identify the shared rare populations across datasets regardless of batch effect and benefits label transfer for mapping cell labels from reference to query. We used SuperSCC to perform atlas level data analysis on more than 90 datasets and build a cell state map of complex tissue in healthy and diseased stages, such as human lung. We show that SuperSCC outperforms existing approaches in identifying cellular context, has better annotation accuracy, and outlines gene modules that indicate conserved immune cell status in lung microenvironments.