Joshua D. Welch

Resolving the causal structure of gene regulatory networks is central to understanding cellular regulation. Pooled CRISPR screens with single-cell RNA readouts measure the transcriptional consequences of many interventions in parallel, providing a powerful resource for causal discovery. Combinatorial designs that perturb multiple regulators per cell increase throughput and can reveal responses that single-gene perturbations do not. However, inferring causal gene networks from these experiments remains challenging because gene regulatory interactions are nonlinear, dynamic, and often involve feedback. Thus, attributing the combined effect of several perturbations to individual regulators remains difficult. Additionally, transcriptional regulation is ultimately carried out by proteins, while single-cell perturbation datasets generally measure only RNA levels. Existing methods often simplify regulatory dynamics or rely on assumptions that conflict with gene regulation, such as acyclicity. Moreover, current evaluation practices usually rely on expression prediction error, which is readily computable in real datasets, but does not necessarily reflect graph quality. A complementary metric is needed to assess inferred regulatory connections using observed perturbation responses in real data. To address these challenges, we present CausalODE, a differentiable framework for learning causal graphs from single-cell perturbation data. Our approach incorporates three central insights to improve upon previous methods. First, genes that are differentially expressed when a gene is perturbed should be downstream of the target gene in the regulatory network; we quantify this property with a metric we term causal reach. Second, causal interactions involve transient cellular responses in RNA transcription, splicing, and degradation. Third, gene regulatory interactions are mediated by proteins, whose levels are not directly measured in single-cell RNA-seq experiments. Building on these insights, CausalODE learns a graph that optimizes the causal reach metric, captures transient cellular responses to perturbation, and explicitly models the latent dynamics of protein levels. Together, these components allow nonlinear responses, feedback loops, and time-lagged regulation. Across simulated Perturb-seq datasets with different sparsities and graph structures, CausalODE improves AUPRC and precision among top-ranked edges compared with baselines and competing methods, while achieving competitive AUROC. We introduce a metric to evaluate the causal reach of inferred graphs on real data by measuring how well the graphs explain differentially expressed genes in single-cell perturbation experiments. Using this metric, CausalODE outperforms the strongest competing method or baseline by approximately 60% on the CaRPool-seq knockdown dataset and 80% on the Norman et al. activation dataset. Its top-ranked edges generally show greater enrichment for high-confidence evidence from cell-type-matched ChIP-seq, CollecTRI, and regulatory STRING than those of other methods. The inferred graphs also distin-guish non-additive from approximately additive genetic interactions in held-out perturbation pairs. Finally, ablation analyses show that optimizing causal reach improves performance on held-out perturbations, and the learned latent regulator activities track simulated protein dynamics, with median temporal correlations ranging from 0.79 to 0.99.

Skeletal muscle tissue underlies several polygenic metabolic traits and diseases. Identifying their genetic mechanisms requires pinpointing cell types, regulatory elements, target genes and causal variants. Here we use genetic multiplexing to generate population-scale single-nucleus chromatin accessibility (single-nucleus assay for transposase-accessible chromatin using sequencing (snATAC–seq)) and transcriptome (single-nucleus RNA sequencing (snRNA-seq)) maps across 287 frozen human skeletal muscle biopsies. We identify 13 cell types and integrate genetic variation to discover >7,000 expression quantitative trait loci (eQTLs) and >100,000 chromatin accessibility QTLs (caQTLs) across cell types, including context-specific eQTL/caQTL with significant G × C interaction. We identify nearly 2,000 eGenes colocalized with caQTL and construct causal directional maps for chromatin accessibility and gene expression. Almost 3,500 genome-wide association study (GWAS) signals across 38 relevant traits colocalize with sn-eQTL/caQTL, most in a cell-specific manner. These signals typically colocalize with caQTL and not eQTL, highlighting the importance of population-scale chromatin profiling for GWAS functional studies. The results of this study illuminate the genetic regulatory architecture of human skeletal muscle.

Bipolar disorder (BD) is a debilitating neuropsychiatric disorder characterized by recurrent mood perturbations, cognitive and functional impairment, and increased risk of suicide. Genome-wide association studies (GWAS) have repeatedly identified the gene ANK3 , which encodes the neuronal scaffolding protein ankyrin-G, as a risk locus for BD. Ankyrin-G interacts directly with the GABA _A receptor-associated protein (GABARAP) to stabilize inhibitory synapses, providing a potential mechanistic link between ANK3 variation and GABAergic dysfunction observed in BD. Previously, we identified an ANK3 p.W1989R variant in a family with BD and demonstrated that this variant abolishes the ankyrin-G/GABARAP interaction, resulting in inhibitory signaling deficits in mice. Here, we investigated the effects of this variant in human induced pluripotent stem cell (hiPSC)-derived neurons generated from an affected individual heterozygous for the ANK3 p.W1989R variant. CRISPR correction of the variant increased inhibitory synapse density on both glutamatergic and GABAergic patient-derived neurons. Integrated spatial transcriptomics and calcium imaging identified cell-type-dependent changes in gene expression and calcium dynamics. ANK3 p.W1989R somatostatin (SST) positive interneurons displayed altered calcium dynamics and reduced SST expression. To determine whether the variant impacts SST interneuron-mediated inhibition of cortical circuits, we selectively stimulated prefrontal cortical SST interneurons in heterozygous Ank3 p.W1989R mice and recorded optically evoked inhibitory postsynaptic currents (oIPSCs) from cortical pyramidal neurons. SST interneuron-mediated oIPSC peak amplitude, charge transfer, and decay times were reduced in variant mice, without a significant change in the paired-pulse ratio. These findings, using complementary human and mouse models, identify novel roles for ankyrin-G in SST interneuron-mediated inhibitory signaling and demonstrate that altered inhibitory connectivity and SST interneuron function are associated with ANK3 p.W1989R. This work provides mechanistic insight into how disrupted ankyrin-G-dependent synaptic function may contribute to inhibitory signaling deficits relevant to BD.

Mutant isocitrate dehydrogenase 1 (mIDH1) catalyzes 2-hydroxyglutarate (2HG) production which leads to epigenetic reprogramming in astrocytomas with tumor protein p53 (TP53)/α-thalassemia/mental retardation, X-linked (ATRX) loss. RNA-sequencing, single-cell RNA-sequencing, and Chromatin Immunoprecipitation sequencing (ChIP-seq) followed by bioinformatics analysis shows that human and mouse mIDH1 gliomas exhibit downregulated gene ontologies (GOs) related to mitochondrial metabolism and upregulated autophagy-related GOs. Decreased mitochondrial metabolism is accompanied by decreased glycolysis, rendering autophagy as a source of energy in mIDH1 gliomas. Mutant IDH1 glioma cells exhibit increased expression of autophagy-related proteins and enhanced microtubule-associated protein 1 light chain 3 (LC3) I/II conversion, indicating augmented autophagy. Inhibiting autophagy in vivo by administration of synthetic protein nanoparticles (SPNPs) encapsulating autophagy related gene 7 (ATG7) silencing RNA sensitizes mIDH1 glioma cells to radiation, resulting in tumor regression, long-term survival, and immunological memory. This work uncovers autophagy as a critical pathway for survival in mIDH1 gliomas and its inhibition elicits radiosensitivity in vitro in human and mouse mIDH1 glioma cells, and in vivo in mIDH1 models.

Droplet-based microfluidics and combinatorial indexing (scifi-ATAC and scifi-RNA) have made single-cell experiments massively scalable. However, higher-order multiplexing complicates data quality, introduces noise, and affects the potential for biological discovery. Here, we show that ambient chromatin accumulates through the experimental workflow and distorts chromatin profiles, most drastically in low-depth nuclei and in minority populations. Standard cell calling relies heavily on read count thresholds, while existing decontamination methods generally operate on aggregated count matrices rather than the underlying reads. We introduce scifi-demux, for preprocessing scifi-ATAC libraries, and AmbientMapper, a generative model that maps reads competitively against multiple references, learns the ambient profile from empty and low-complexity barcodes, and separates nuclei from background and singlets from doublets by Bayesian Information Criterion. Using interspecies ground truth experiments, published multi-genotype libraries, and simulated read-level synthetic barcodes in which every contaminating read is traceable, we show that calls are robust to parameter variation and stable across designs, achieving a wrong-genome rate of 0.19% on a 26-genome reference panel. Finally, we evaluated the impact of removing contaminants, showcasing how AmbientMapper rescues low-depth nuclei discarded by standard pipelines and restores biological structure obscured by contamination.

While human genetic evidence improves drug program success, target selection largely relies on associational features. Even functional features are mainly observational, capturing disease associations rather than the consequence of perturbing genes in the disease-relevant human cell type. To move beyond association, we present IGNITE (Immune Genomics and fuNctional Integration for Target Enrichment), a framework that prioritizes immune drug targets by integrating human genetic priors with features from genome-scale perturb-seq in 22 million primary human CD4 ⁺ T cells and polarized T-helper subset differential expression. IGNITE is a semi-supervised machine learning model that applies positive-unlabeled learning to approved immune targets, then ranks 19,502 protein-coding genes to prioritize new candidates. In 727 in-trial genes held out from training, functional genomics increased target enrichment among the top 50 nominations from 2.7- to 4.8-fold for immune trial targets at any phase, and from 4.5- to 5.9-fold for targets in Phase III. These performance gains were immune-specific, as IGNITE outperformed genetic comparators on predicting immune-exclusive targets but not on cardiac-exclusive targets. Using temporal validation with labels frozen in 2014, IGNITE outperformed the genetics-only model in ranking 134 genes that subsequently entered immune trials. The functional genomics layer also surfaced two novel, pharmacologically tractable candidates, ELOVL6 and RUVBL1 , elevating their rankings from outside the top 1,000 into the top 50. These findings highlight that perturbational signatures in the disease-relevant human cell type harbor target-relevant signals beyond human genetics, offering a generalizable route to indication-specific prioritization as perturbation atlases expand. Full IGNITE scores are publicly available at https://ignite.eecs.umich.edu/

Correcting batch effects and integrating single-cell sequencing datasets has been a crucial step in large-scale biological studies. Many methods have been published for this task, and excel in various scenarios. Our previous work, LIGER, leveraging integrative non-negative matrix factorization (iNMF), stands out in providing an interpretable low-dimensional representation. To adapt to the modern need for integrating millions of cells, we developed a highly-optimized parallel factorization solution with on-demand loading from disk. The upgraded LIGER algorithm shows significant improvements in time and memory efficiency for single-cell data integration. We also developed a new downstream embedding alignment method significantly improved performance in conserving biological variation while still aligning corresponding cell types across datasets.

Glioblastoma is among the most lethal human malignancies. Immune-based therapies have failed due to a strong immunosuppressive tumor microenvironment (TME)1. We uncovered that LAIR-12 expressed by tumor cells simultaneously drives TME fibrosis and inhibits migration of immune cells to brain tumors, thus achieving powerful immune exclusion3. We demonstrate that glioma-cell-specific LAIR-1 knockdown (KD), but not LAIR-1 KD from host cells, significantly extends survival in an immune-dependent manner. Glioma-cell-specific LAIR-1 signals through SHP2 to activate JNK –which on one hand sustains high levels of Lysyl Oxidase-Like 1 and collagen I to block immune cells’ entry– and on the other hand suppresses STAT3 signaling. In the absence of LAIR-1, gliomas’ collagen-dense ECM becomes disassembled and, through STAT3-driven upregulation of ADAM10 and ADAM17, promotes release of CXCL16, and recruitment of NK cells and cytotoxic T cells. Combining LAIR-1 KD with immune-stimulatory gene therapy4 achieved 100% long-term survival with durable immunological memory in immunocompetent mice. Pharmacological SHP2 inhibition in LAIR-1 WT mouse and human glioma cells recapitulated the LAIR-1 KD molecular phenotype and similarly potentiated gene therapy. These findings define LAIR-1 as a tumor-cell-intrinsic pro-fibrotic and immunosuppressive checkpoint and identify the LAIR-1>SHP2>JNK>CXCL16 and/or LOXL1 axis as a therapeutic target for sensitizing glioma to immunotherapy.

Primary cutaneous T-cell lymphomas (CTCL) are a heterogeneous group of extranodal non-Hodgkin lymphomas. Outcomes for patients with advanced-stage disease are suboptimal, as few complete and durable responses are achieved with currently available therapeutic agents. However, improved understanding of oncogenic transcriptional programs in malignant T cells and their engagement with the tumor microenvironment (TME) may unveil therapeutic vulnerabilities. Therefore, we have compiled the largest available scRNA-seq CTCL atlas that includes >2 million skin and blood cells from 116 individual patients. We identified recurrent transcriptional programs in malignant T cells, a subset of which are associated with GATA-3 dependent transcriptional programs, especially in the setting of large cell transformation and advanced-stage disease. Many of the transcriptional programs identified are therapeutically targetable with clinically available agents, including HDAC, XPO1, CDK9, JAK/CSF1R, and IKZF1/IKZF3 antagonists. The CTCL TME is dominated by infiltrating and exhausted effector and cytotoxic T cells that are restrained by a robust infiltrate of regulatory T cells, transcriptionally polarized monocytes/macrophages, and cancer-associated fibroblasts. Collectively, these findings have significant implications for the rationale design of combinatorial strategies targeting immune checkpoints, including PD-1. We have identified transcriptional programs, driven by oncogenic transcription factors, and constituents of the TME as therapeutic vulnerabilities in CTCL, and hope that the CTCL atlas constructed will provide a valuable resource for future studies exploiting these therapeutic vulnerabilities.

Cells communicate through ligand–receptor (LR) signaling interactions, but identifying when and where these interactions are active remains challenging. We developed CytoSignal to infer the locations and dynamics of cell–cell communication at cellular resolution from spatial transcriptomic data. Here we show that our cellular resolution, spatially resolved signaling scores enable several important analyses—identifying spatial gradients in signaling strength, quantifying the locations of contact-dependent and diffusible interactions, detecting signaling-associated genes and identifying differential signaling across multisample data. Additionally, we can predict the temporal dynamics of a signaling interaction at each spatial location. We experimentally validate our results in situ by proximity ligation assay, confirming that CytoSignal predicts the locations of LR interactions more accurately than previous approaches. This study addresses the field’s current need for a robust and scalable tool to detect cell–cell signaling interactions and their dynamics at cellular resolution from spatial transcriptomic data.