• Hanzhang Fang, New Jersey Institute of Technology, United States
• Cong Qi, New Jersey Institute of Technology, United States
• Yuanjie Zou, New Jersey Institute of Technology, United States
• Yeqing Chen, New Jersey Institute of Technology, United States
• Zhi Wei, New Jersey Institute of Technology, United States

Motivation: Spatial transcriptomics (ST) measures gene expression in intact tissues, but each spot typically contains mixtures of multiple cell types. Deconvolution is particularly challenging when closely related cell subtypes share highly similar expression profiles and when spatial context is underutilized during proportion estimation.
Results: We present SlotDeconv, a method for ST deconvolution consisting of a single-cell reference module and a spatial inference module. The reference module learns discriminative cell-type signatures using slot based prototype vectors decoded into a reference matrix, trained with a negative binomial reconstruction loss and a max-margin diversity constraint that discourages similar cell-type signatures. Ablation studies confirm that both components are essential: removing the diversity constraint reduces spot-wise Pearson correlation by 49%, and replacing learned prototypes with cell-type mean expression reduces it to near zero. The spatial inference module initializes spot level proportions via
gene weighted nonnegative least squares, then refines them by minimizing KL divergence between observed and reconstructed spot expression under a spatial neighborhood consistency regularizer. Benchmarked against CARD, RCTD, Cell2location, and Spotiphy on a 27 cell type mouse brain dataset, SlotDeconv achieves the highest spot wise Pearson correlation (0.561) and cosine similarity (0.633), outperforming the next best method by 7.5% in spot-wise correlation, with particularly strong gains on transcriptionally similar cortical neuronal subtypes. Biological validation on human pancreatic cancer and mouse olfactory bulb datasets further confirms spatial specificity.

• Md Istiaq Ansari, University of Central Florida, United States
• Muhtasim Noor Alif, University of Central Florida, United States
• Wei Zhang, University of Central Florida, United States

Motivation: Spatial transcriptomics (ST) technologies profile gene expression while preserving tissue architecture, enabling the study of spatial cellular organization and microenvironmental interactions. However, raw ST data are heavily affected by technical noise, sparsity, and dropout events, which obscure true biological signals and hinder downstream analyses. While recent denoising methods incorporate spatial neighborhood information, they lack explicit supervision due to missing cell-type annotations in ST data. To address this challenge, we introduce SPIDER, a semi-supervised framework that leverages independently generated, annotated single-cell RNA-seq (scRNA-seq) references to guide ST denoising. SPIDER synthesizes pseudo-ST data from scRNA-seq to inject cell-type information without requiring paired measurements. The method constructs three graphs capturing spatial proximity, transcriptional similarity in real-ST data, and transcriptional structure in pseudo-ST data. Graph encoders map these representations into a shared latent space, and a domain-alignment module transfers biologically meaningful structure from pseudo-ST to real-ST embeddings. A graph-attention decoder with a zero-inflated negative binomial objective reconstructs denoised ST expression profiles.

Results: We benchmark SPIDER on human dorsolateral prefrontal cortex and breast cancer datasets. SPIDER consistently enhances spatial gene expression patterns, recovers known tissue structures, and achieves superior clustering performance compared to existing approaches. Marker gene analyses demonstrate improved spatial continuity and clearer anatomical organization. By directly producing denoised expression matrices, SPIDER improves both accuracy and interpretability, providing a generalizable solution for robust ST data denoising.

Availability and Implementation: The source code is available at: https://github.com/compbiolabucf/SPIDER

• Peiying Cai, University of Zurich, Switzerland
• Mark Robinson, University of Zurich, Switzerland
• Simone Tiberi, Department of Statistical Sciences, University of Bologna, Bologna, Italy, Italy

Motivation: Spatially resolved transcriptomics (SRT) enables the investigation of mRNA expression in a spatial context. While several SRT analysis
frameworks have been developed, the vast majority of them focus on analyzing individual samples, and do not allow for comparisons of spatial gene
expression patterns across experimental conditions, such as healthy vs. diseased states.
Results: Here, we present an approach to identify so-called differential spatial patterns (DSP), i.e., genes that exhibit changes in spatial expression
between groups of samples across conditions. Our framework processes diverse SRT data types and detects DSP by performing differential gene
expression testing across conditions. Notably, this comparison is not currently available in any other spatial omics method. In addition to detecting
DSP across conditions, our framework includes two key features. First, it can identify tissue regions where expression changes across conditions.
Second, it can detect spatial gene expression pattern changes across more than two experimental conditions, with flexible models. With ad hoc
simulations, we demonstrate that our approach has good true positive rates and well-calibrated false discovery rates. Applied to experimental data,
our method identifies biologically relevant DSP genes while maintaining computational efficiency.
Availability: Our framework has been implemented within DESpace Bioconductor R package (from version 2.0.0).

• Lisa Fournier, University of Geneva / Ecole Polytechnique Fédérale de Lausanne / University of Bern, Switzerland
• Garance Haefliger, Idiap Research Institute, Switzerland
• Albin Vernhes, Idiap Research Institute, Switzerland
• Vincent Jung, Ecole Polytechnique Fédérale de Lausanne (EPFL), Switzerland
• Lena Loye, University of Bern, Switzerland
• Valentine Du Bois, University of Geneva, Switzerland
• Intidhar Labidi-Galy, University of Geneva, Switzerland
• Pascal Frossard, Ecole Polytechnique Fédérale de Lausanne (EPFL), Switzerland
• Igor Letovanec, Centre Hospitalier Universitaire Vaudois CHUV / Valais Hospital, Switzerland
• Cédric Vincent-Cuaz, University of Bern, Switzerland
• Raphaëlle Luisier, University of Bern, Switzerland

Intra-tumor heterogeneity (ITH) is a major driver of cancer progression and therapeutic resistance. Although ITH has been extensively profiled at the molecular level, revealing the coexistence of tumor subpopulations with distinct molecular identities, how these programs organize spatially within tumors, forming recurrent regions with unique molecular profiles, tissue architecture, and cellular composition, remains poorly understood. In this study, we push the boundaries of histopathology foundation models (hFMs) to reveal molecularly distinct tumor regions that are invisible to human experts, offering new insights into the molecular organization of invasive breast tumors.
We specialized generalist hFMs through extended pre-training on invasive tumor tissue, enhancing their ability to encode richer, tumor-specific biological concepts. Based solely on these image-derived features, the model identified distinct tumor regions, termed tumor archetypes, that subsequently mapped to unique molecular signatures. Characterizing these tumor archetypes by integrating spatial transcriptomics data, we uncovered distinct patterns of RNA splicing dysregulation alongside other pathways such as TGF-β signaling. Notably, aberrant RNA splicing emerged as the main driver of one archetype across both HER2+ breast cancer and triple-negative breast cancer (TNBC).
Strikingly, these archetypes coexist as spatially distinct regions within the same tumors and recur across patients. Survival analysis revealed that enrichment of the RNA splicing–dysregulated archetype is associated with poorer overall survival, highlighting its clinical relevance.
This work provides a scalable method to uncover tumor archetypes from H&E images, linking tissue morphology to RNA programs and opening new avenues for precision oncology and RNA-focused therapeutic stratification.

• Jonathan Mandl, Bar-Ilan University, Israel
• Yaron Orenstein, Bar-Ilan University, Israel

Motivation: Predicting the RNA-seq coverage of native and exogenous sequences is central to many molecular- and synthetic-biology applications. Substantial progress has been made in developing methods to predict the RNA-seq coverage of native genomic sequences, with the recently developed Shorkie achieving state-of-the-art performance in yeast. However, prediction performance of these methods over exogenous DNA is still unknown. Recent studies measured RNA-seq coverage of large exogenous genomes in yeast, providing a unique opportunity to train machine-learning models on a large exogenous sequence space and to improve both prediction performance and our understanding of regulatory mechanisms.

Results: We introduce ExoShorkie, a method we developed by extending Shorkie through transfer learning across multiple exogenous RNA-seq datasets. We demonstrate that ExoShorkie significantly improves prediction performance on held-out exogenous genomes and outperforms both a native-genome-trained Shorkie baseline and Yorzoi, the only competing method in predicting exogenous RNA-seq coverage in yeast, in cross-validation and in leave-one-genome-out evaluations. Furthermore, through interpretability analyses we reveal biologically meaningful regulatory motifs and distinct regulatory rules in exogenous genomes in yeast, providing new insights into transcriptional regulation.

• David Cechak, Central European Institute of Technology, Czechia
• Dimosthenis Tzimotoudis, University of Malta, Malta
• Stephanie Sammut, University of Malta, Malta
• Katarina Gresova, Max Delbrück Center, Germany
• Eva Marsalkova, Central European Institute of Technology, Czechia
• David Farrugia, University of Malta, Malta
• Panagiotis Alexiou, University of Malta, Malta

Motivation: MicroRNAs (miRNAs) regulate gene expression by guiding Argonaute proteins to partially complementary sites on target RNAs. While classical prediction methods rely on engineered features such as seed match categories, evolutionary conservation, and site context, recent advances in deep learning offer the potential to learn targeting rules directly from sequence. We developed a sequence-based deep learning model that improves miRNA target site prediction, and further validated the learned target site representations by extending the model to gene-level functional repression prediction.
Results: We introduce miRBind2, a deep learning method for miRNA target site prediction that incorporates a novel pairwise nucleotide representation capturing all possible miRNA-target nucleotide interactions, with a CNN-based architecture. miRBind2 outperforms previous SotA models across four independent datasets from the debiased miRBench benchmark, while using 92% fewer parameters. We show that the convolutional features and weights learned by miRBind2 can be transferred to transcript-level prediction by extending the miRBind2 architecture and fine-tuning it on miRNA perturbation experiments. This miRBind2-3UTR model predicts gene repression from sequence alone. On a dataset of 50,549 miRNA-gene pairs, miRBind2-3UTR significantly outperforms TargetScan. These results show that deep models pretrained on target site data can capture regulatory signals and predict functional repression without requiring conventional engineered biological features.
Availability: Models and source code are freely available via GitHub (https://github.com/BioGeMT/miRBind_2.0). A publicly available web-tool for novel predictions and visualization is available at : (https://huggingface.co/spaces/dimostzim/BioGeMT-miRBind2)

• Sebastian Hoenig, Department of Computer Science, ETH Zürich, Switzerland
• Aayush Grover, Department of Computer Science, ETH Zürich, Switzerland
• Piero Neri, Department of Computer Science, ETH Zürich, Switzerland
• Didier Surdez, Balgrist University Hospital, Faculty of Medicine, University of Zürich, Switzerland
• Valentina Boeva, Department of Computer Science, ETH Zürich, Switzerland

Motivation: Three-dimensional folding of the genome into structures such as chromatin loops is essential for gene regulation. Current experimental methods for mapping these structures, like Hi-C and HiChIP, are labor-intensive and require repeated assays to test hypothesized mutation effects. This motivates the need for predictive approaches that reveal the sequence determinants of chromatin loops.

Results: In this work, we present a novel and interpretable computational pipeline for predicting CTCF-mediated chromatin loops. We propose Chiron3D, a DNA-only model trained in a cell-type-specific manner to predict CTCF HiChIP contact maps. By leveraging pre-trained embeddings from a foundation model, our approach is competitive with baselines that take CTCF ChIP-seq as additional input, while enabling nucleotide-level attribution to the input DNA sequence. Using our framework, we provide likely mechanistic insights into the physical control of loop dynamics. Specifically, we find that the strength of the loop extrusion anchorage site is largely governed by the amount and binding affinity of CTCF sites at the boundaries. Furthermore, we reveal that loop stability is regulated by the amount of intra-loop CTCF binding sites, where fewer sites within the loop lead to a more stable domain. Using targeted, single-nucleotide edit simulations with Chiron3D, we show that both loop strength and stability can be precisely controlled. Together, these results provide novel mechanistic insights into the physical control of genome organization and highlight the potential of decoding the DNA sequence logic in silico.

• Aishwarya Iyer, Maastricht University, Netherlands
• Daan van Beek, MaCSBio, Maastricht University, Netherlands
• Friederike Ehrhart, Maastricht University, Netherlands
• Chris Evelo, Maastricht University, Netherlands
• Theo de Kok, Maastricht University, Netherlands
• Ilja Arts, Maastricht Centre for Systems Biology (MaCSBio), Netherlands
• Michiel Adriaens, Maastricht University, Netherlands
• Martina Summer-Kutmon, Maastricht University, Netherlands

Motivation: Huntington's disease (HD) exhibits substantial variability in age of onset and disease progression that is not fully explained by CAG repeat length alone. Part of this residual variation is heritable, implicating additional genetic mechanisms. Cis-regulatory variation, genetic variants that alter transcription and splicing of nearby genes, represents one such mechanism that can be quantified through allele-specific expression (ASE) analysis. However, methods for integrating ASE profiles into patient stratification frameworks remain underdeveloped, particularly for rare diseases with small cohorts and sparse data.
Results: We adapt a network-based stratification algorithm, originally developed for somatic tumor mutations, to ASE data. By propagating gene-level ASE imbalance profiles through a protein-protein interaction network, we stratified 20 HD patients into three distinct biological patient subgroups. Differential ASE and gene expression analyses both highlight neuroinflammatory pathways, including IL6 signaling, microglial activation, and cytokine regulation, supporting their role as key sources of inter-patient heterogeneity. Intersection of differentially imbalanced and expressed genes identified four candidate genes with potential eQTL-mediated regulation, MAGI3, HLA-B, CCND3, and PSMB1. These
genes implicate synaptic organization, immune modulation, and proteasomal dysfunction as potential drivers of molecular heterogeneity between patient subgroups. Differences in cortical and striatal neuropathological scores between clusters, even when adjusted for CAG repeat length, provide clinical support for the biological relevance of the identified subgroups.
Availability: All analysis code, Docker containers, and conda environments are available at https://github.com/macsbio/HD-ASE-NBS.
Contact: martina.kutmon@maastrichtuniversity.nl
Supplementary information: Supplementary data are available at Bioinformatics online.

• Florian Hugi, ETH Zurich, Switzerland
• Tanmay Tanna, ETH Zurich, Switzerland
• Randall J Platt, ETH Zurich, Switzerland

Motivation: Record-seq captures transcriptional activity over time in engineered E. coli by integrating cellular RNA-derived spacer sequences into CRISPR arrays, which are subsequently read out by sequencing. Unlike the approximately uniform transcript sampling of RNA-seq, Record-seq records biological signal as spacers sampled by the CRISPR spacer acquisition machinery. Consequently, standard RNA-seq analysis strategies are not directly applicable, limiting sensitivity and interpretability. Our previous pipeline addressed these challenges only partially, retained inherited RNA-seq assumptions, and had limited algorithmic efficiency.
Results: Here, we present an end-to-end computational framework for Record-seq data. To address the primary computational bottleneck – spacer sequence extraction – we implemented a wavefront-alignment approach for efficient quasi-local pattern matching, achieving a 30-fold speedup. We also introduce transcription unit-based feature counting as an alternative to gene-body quantification to better represent prokaryotic transcription and increase statistical power by capturing signal from untranslated regions, which are spacer acquisition hotspots. For downstream analyses, we incorporate multiple normalization strategies and a nonparametric differential expression testing framework specifically designed for sparse datasets. Further, we analyze spacer acquisition patterns and train sequence-based neural models that predict acquisition propensity from genomic sequence and annotations, providing a framework to assess whether acquisition rules generalize as Record-seq is extended to new microbial hosts.
Availability and Implementation: The primary analysis workflow, recoRdseq package, acquisition modelling repository and relevant data are all linked at https://github.com/plattlab/Record-seq-Framework. Acquisition models and training data are on Zenodo at https://doi.org/10.5281/zenodo.18891434.
Supplementary Information: Supplementary data are available online.

• Arttu Jolma, University of Toronto, Canada
• Aldo Hernandez-Corchado, McGill University, Canada
• Ally Yang, University of Toronto, Canada
• Ali Fathi, University of Toronto, Canada
• Kaitlin Laverty, University of Toronto, Canada
• Alexander Brechalov, University of Toronto, Canada
• Rozita Razavi, University of Toronto, Canada
• Mihai Albu, University of Toronto, Canada
• Hong Zheng, University of Toronto, Canada
• Ivan Kulakovskiy, Russian Academy of Sciences, Russia
• Hamed Najafabadi, McGill University, Canada
• Tim Hughes, University of Toronto, Canada

There is ongoing controversy regarding the degree to which transcription factors (TFs) independently specify genomic binding: TF binding motifs are short and degenerate, producing excessive binding site predictions. Here, we present a computational framework, developed in conjunction with large-scale genomic HT-SELEX (GHT-SELEX), to examine intrinsic sequence-specificities of human TFs.

The first component of this framework is MAGIX, a hierarchical Bayesian model that explicitly captures the exponential enrichment dynamics of TF-bound genomic DNA fragments across SELEX cycles, producing quantitative enrichment scores and statistical confidence estimates for genomic binding regions. Applied to in vitro GHT-SELEX data for 179 TFs across 25 families, we find that genomic binding regions discovered by MAGIX often display surprisingly high overlap with ChIP-seq peaks for the same TF, suggesting much higher intrinsic sequence specificity than anticipated.

The second component, RCADEEM, leverages a protein sequence-derived recognition code to model alternative DNA-binding modes of C2H2 zinc finger (C2H2-ZF) TFs. RCADEEM integrates machine learning-based prediction of C2H2-ZF sequence preferences with GHT-SELEX/MAGIX binding profiles to infer which subsets of ZFs are engaged at individual loci. Applied to 86 C2H2-ZF proteins, RCADEEM reveals that modular, alternative engagement of C2H2-ZF domains is the norm, leading to recognition of multiple distinct motifs by the same TF. These alternative motifs often evolve by internal duplication and divergence within the C2H2-ZF array.

Together, this framework reveals that it is common for TFs to delineate a large fraction of their in vivo genomic binding sites independently of other cellular factors, often through complex, context-dependent DNA-binding behaviours.

• Linan Zhang, Ningbo University, China
• April Sagan, University of Pittsburgh, United States
• Bin Qin, University of Pittsburgh, United States
• Haoyu Wang, University of Pittsburgh, United States
• Elena Kim, University of Pittsburgh, United States
• Baoli Hu, University of Pittsburgh, United States
• Hatice Ulku Osmanbeyoglu, University of Pittsburgh, United States

Transcription factors (TFs) orchestrate cellular responses to environmental signals and intercellular communication. The activity of TFs is influenced by neighboring cells, impacting cellular fate and function. Spatial transcriptomics (ST) allows for the mapping of mRNA expression across tissue samples, providing insights into the local microenvironment. However, the potential of ST data to systematically infer TF activity and its role in cell identity has not been fully exploited. We introduce STAN (Spatially informed Transcription factor Activity Network), a linear mixed-effects computational approach that predicts spatially informed, spot-specific TF activities by integrating curated TF–target gene priors, mRNA expression, spatial coordinates, and histological features. We demonstrate the utility of STAN on lymph node, dorsolateral prefrontal cortex, breast cancer, and glioblastoma ST datasets, identifying TFs associated with specific cell types, spatial regions, pathological zones, and ligand–receptor pairs. STAN enhances the utility of ST data, revealing the intricate interplay between TFs and spatial organization in diverse biological contexts.

• Casper H. Blaauw, VIB.AI, KU Leuven & Hubrecht Institute, Belgium
• Niklas Kempynck, VIB-KU Leuven, Belgium
• Seppe De Winter, VIB-KU Leuven, Belgium
• Vasilieios Konstantakos, VIB-KU Leuven, Belgium
• Eren Can EkÅŸi, VIB-KU Leuven, Belgium
• Sam Dieltiens, VIB-KU Leuven, Belgium
• Darina Abaffyová, VIB-KU Leuven, Belgium
• Valérie Bercier, VIB-KU Leuven, Belgium
• Ibrahim Taskiran, Illumina, United States
• Gert Hulselmans, VIB-KU Leuven, Belgium
• Valerie Christiaens, VIB-KU Leuven, Belgium
• Ludo Van Den Bosch, VIB-KU Leuven, Belgium
• Lukas Mahieu, VIB-KU Leuven, Belgium
• Alexander van Oudenaarden, Hubrecht Institute, Netherlands
• Oliver Hobert, Columbia University, United States
• David R. Kelley, Calico Labs, United States
• Stein Aerts, VIB-KU Leuven, Belgium

The main goal of regulatory genomics is to decipher how spatiotemporal regulation of genes is encoded in the genome. Sequence-to-function models, which learn to link sequence to readouts like chromatin accessibility or gene expression, are the state-of-the-art for this task. These models primarily operate at two levels: predicting the scATAC-seq-based chromatin accessibility for individual enhancers and predicting scRNA-seq-based expression genes based on their surrounding locus.

Here, we present CREsted, a user-friendly package for enhancer modelling covering all steps from preprocessing to interpretation. Starting from scATAC-seq data, CREsted provides tools to preprocess the data, train sequence-to-function models, decipher the learned sequence grammar, and design new cell type-specific enhancers. We showcase CREsted's performance on a variety of tissues and species, ranging from mouse brain cortex and human cancer states to a whole-organism zebrafish developmental atlas.

Furthermore, we discuss the evolution of the field towards modelling cell type-specific gene expression-predicting models. Using the toolkit provided in CREsted, we trained a model to predict gene expression across every cell type of an entire animal, the nematode C. elegans. Using this model, we extract and cluster high-importance sequence motifs from the genome, which we use to build an organism-wide motif grammar across tissues. Furthermore, we use in-silico perturbations to link regulatory elements to genes, and fine-map eQTLs by scoring naturally occurring variants. Finally, we use atlases from two related nematode species to for cross-species comparison and augmentation. Altogether, this provides a view into genome regulation at a unique scale.

• Adrien Osakwe, McGill University, Canada
• Eric Haoran Huang, McGill University, Canada
• Yue Li, McGill University, Canada

Temporal modelling of single-cell gene expression is essential for capturing dynamic cellular processes, yet a systematic framework for evaluating time-aware trajectory inference methods has not yet been established. Here, we present a modular and scalable benchmark designed to assess methods across three critical tasks: forecast accuracy (temporal cell alignment) for projecting cells to unseen time points, embedding coherence between original and projected data, and cell-type lineage fidelity. We evaluated ten state-of-the-art methods, which are broadly categorized into 8 forecasting-based and 2 optimal transport (OT)-based methods across eight diverse datasets spanning four species. Our results show that while several methods achieve high forecast accuracy, they often fail to preserve biological signals, both in their latent spaces and in cell lineage reconstruction. Notably, most methods confer low lineage fidelity and often underperform compared to a correlation baseline. We further demonstrate that integrating pseudotime can effectively denoise trajectories by aligning the data snapshots with the intrinsic biological clock in each cell. Finally, to streamline benchmarking for temporal single-cell analysis, we built one of the first self-contained Python packages for the research community: https://github.com/li-lab-mcgill/scTimeBench.

• Rufus Daw, The University of Manchester, United Kingdom
• Harry Deijnen, The University of Manchester, United Kingdom
• Magnus Rattray, The University of Manchester, United Kingdom
• John Grainger, The University of Manchester, United Kingdom

Single-cell RNA sequencing (scRNA-seq) cell annotation techniques rely on the matching of known defining marker genes to a given cell population. However, these methods may lack robustness to dynamic fluctuations in cell marker expression between patients, samples and pathologies. The advent of easy-to-implement predictive technologies, like generative-AI (gen-AI), has facilitated the introduction of computational workflows that improve otherwise inaccurate context-dependent cell type annotation. Here, we introduce CellTypeAI, a streamlined, scalable program developed for tissue context-dependent cell annotation of scRNA-seq datasets using modern gen-AI models, enhanced by retrieval augmented generation methods. Our implementation builds upon local gen-AI hosting technologies and directly integrates into scRNA-seq analysis pipelines. We show that CellTypeAI provides improved annotation accuracy compared to current conventional annotation methods and nascent cloud-based gen-AI approaches. As CellTypeAI leverages locally-run AI models, it can be applied to sensitive datasets, unlike approaches utilising online gen-AI tools such as ChatGPT, DeepSeek, or Claude. CellTypeAI presents a novel solution for tissue-specific cell type identification, overcoming traditional marker-based limitations via locally-deployed gen-AI models.

• Abdullah Al Noman, Virginia Tech, United States
• Palash Sashittal, Virginia Tech, United States

Cell differentiation is a dynamic process in which cells traverse through high-dimensional gene expression space under the influence of gene regulatory networks and environmental cues. Recent advances in single-cell RNA sequencing (scRNA-seq) have enabled us to measure high-resolution snapshots of this dynamic process. Several computational methods have been developed to reconstruct cellular flow maps from these snapshots, revealing the trajectories taken by cells in gene expression space. While existing methods provide increasingly detailed descriptions of cellular trajectories, the stability of these trajectories to perturbations is largely unexplored. As such, it remains unclear how robust inferred trajectories are to perturbations, which genes most strongly influence long-term fate outcomes, and where instability arises between competing fate commitments. While sensitivity and stability analysis tools from dynamical systems theory provide a principled way to study the stability of differentiation trajectories, existing approaches are not designed for the high-dimensionality and sparsity of scRNA-seq data. Here, we introduce FateSens, a sensitivity-based computational framework for analyzing gene regulatory dynamics using flow maps derived from scRNA-seq data. FateSens performs sensitivity analysis of differentiation trajectories derived from scRNA-seq data to identify regulatory genes and fate boundaries. To demonstrate its utility, we applied FateSens to study neutrophil-monocyte differentiation using scRNA-seq data of mouse hematopoiesis. While FateSens relies only on transcriptomic measurements, this dataset also contains lineage tracing barcodes that provide ground-truth fate relationships. Our results show that FateSens accurately recovers regulators consistent with known biology and identifies fate boundaries that are supported by lineage tracing data.