William Rudman, Michal Golovanevsky, Dana Arad, Yonatan Belinkov, Carsten Eickhoff, Ritambhara Singh, Kyle Mahowald. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

BACKGROUND: Key challenges in leveraging unstructured clinician notes for predictive models include identifying and timing patient outcomes. To address these challenges we applied large language models (LLMs) to identify and temporally localize patient outcomes in clinician notes, and evaluated whether this contextual data enhances predictive modeling for conditions like sepsis. METHODOLOGY: We applied the Medical Concept Annotation Tool (MedCAT) and two LLMs, Meta-Llama-3.1-8B and BioMistral 7B, to clinician notes in the Medical Information Mart for Intensive Care version III to identify and time International Classification of Diseases-based patient outcomes. A physician manually validated the accuracy of the models' ability to identify and time outcomes including sepsis. Finally, downstream time series predictive modeling was used to assess the impact of unstructured clinician notes on sepsis prediction. RESULTS: Meta-Llama-3.1-8B demonstrated the best balance of coverage and precision, reliably identifying outcomes while minimizing false positives. Our manual physician validation evaluated model performance, which varied between outcomes. When comparing derived sepsis timestamps to clinical sepsis onset within a 5-h window, Meta-Llama-3.1-8B showed the smallest median time difference. Finally, incorporating temporally localized outcomes into downstream sepsis prediction improved model performance, area under the receiver operating characteristic (AUC) of 0.84 (95% CI: 0.83-0.86) from 0.77 (95% CI: 0.76-0.79). CONCLUSION: We demonstrate that LLM-driven identification and timing of patient outcomes from unstructured clinician notes is feasible. Contextual outcome identification unlocks the potential of unstructured clinician notes for predictive modeling. We provide TIMED-MIMIC, an automatically generated dataset encompassing 1697 temporally localized patient outcomes, as a publicly available resource.

Multimodal Large Language Models (MLLMs) perform well on tasks such as visual question answering, but it remains unclear whether their reasoning relies more on memorized world knowledge or on the visual information present in the input image.To investigate this, we introduce Visual CounterFact, a new dataset of visually-realistic counterfactuals that put world knowledge priors (e.g, red strawberry) into direct conflict with visual input (e.g, blue strawberry).Using Visual CounterFact, we show that model predictions initially reflect memorized priors, but shift toward visual evidence in mid-to-late layers.This dynamic reveals a competition between the two modalities, with visual input ultimately overriding priors during evaluation.To control this behavior, we propose Pixels Versus Priors (PvP) steering vectors, a mechanism for controlling model outputs toward either world knowledge or visual input through activation-level interventions.On average, PvP successfully shifts 99.3% of color and 80.8% of size predictions from priors to counterfactuals.Together, these findings offer new tools for interpreting and controlling factual behavior in multimodal models.* †

Biomedical knowledge graphs (KGs) encode rich, structured information critical for drug discovery tasks, but extracting meaningful insights from large-scale KGs remains challenging due to their complex structure. Existing biomedical subgraph retrieval methods are tailored for graph neural networks (GNNs), limiting compatibility with other paradigms, including large language models (LLMs). We introduce K-Paths, a model-agnostic retrieval framework that extracts structured, diverse, and biologically meaningful multi-hop paths from dense biomedical KGs. These paths enable prediction of unobserved drug-drug and drug-disease interactions, including those involving entities not seen during training, thus supporting inductive reasoning. K-Paths is training-free and employs a diversity-aware adaptation of Yen's algorithm to extract the K shortest loopless paths between entities in a query, prioritizing biologically relevant and relationally diverse connections. These paths serve as concise, interpretable reasoning chains that can be directly integrated with LLMs or GNNs to improve generalization, accuracy, and enable explainable inference. Experiments on benchmark datasets show that K-Paths improves zero-shot reasoning across state-of-the-art LLMs. For instance, Tx-Gemma 27B improves by 19.8 and 4.0 F1 points on interaction severity prediction and drug repurposing tasks, respectively. Llama 70B achieves gains of 8.5 and 6.2 points on the same tasks. K-Paths also boosts the training efficiency of EmerGNN, a state-of-the-art GNN, by reducing the KG size by 90% while maintaining predictive performance. Beyond efficiency, K-Paths bridges the gap between KGs and LLMs, enabling scalable and explainable LLM-augmented scientific discovery. We release our code and the retrieved paths as a benchmark for inductive reasoning.

Michal Golovanevsky, William Rudman, Vedant Palit, Carsten Eickhoff, Ritambhara Singh. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2025.

Epigenetic mechanisms play a crucial role in driving transcript expression and shaping the phenotypic plasticity of glioblastoma stem cells (GSCs), contributing to tumor heterogeneity and therapeutic resistance. These mechanisms dynamically regulate the expression of key oncogenic and stemness-associated genes, enabling GSCs to adapt to environmental cues and evade targeted therapies. Importantly, epigenetic reprogramming allows GSCs to transition between cellular states, including therapy-resistant mesenchymal-like phenotypes, underscoring the need for epigenetic-targeting strategies to disrupt these adaptive processes. Understanding these epigenetic drivers of gene expression provides a foundation for novel therapeutic interventions aimed at eradicating GSCs and improving glioblastoma outcomes. Using machine learning (ML), we employ cross-patient prediction of transcript expression in GSCs by combining epigenetic features from various sources, including ATAC-seq, CTCF ChIP-seq, RNAPII ChIP-seq, H3K27Ac ChIP-seq, and RNA-seq. We investigate different ML and deep learning (DL) models for this task and ultimately build our final pipeline using XGBoost. The model trained on one patient generalizes to other 11 patients with high performance. Notably, H3K27Ac alone from a single patient is sufficient to predict gene expression in all 11 patients. Furthermore, the distribution of H3K27Ac peaks across the genomes of all patients is remarkably similar. These findings suggest that GSCs share a common distributional pattern of enhancer activity characterized by H3K27Ac, which can be utilized to predict gene expression in GSCs across patients. In summary, while GSCs are known for their transcriptomic and phenotypic heterogeneity, we propose that they share a common epigenetic pattern of enhancer activation that defines their underlying transcriptomic expression pattern. This pattern can predict gene expression across patient samples, providing valuable insights into the biology of GSCs.

Inferring gene regulatory networks from gene expression data is an important and challenging problem in the biology community. We propose OTVelo, a methodology that takes time-stamped single-cell gene expression data as input and predicts gene regulation across two time points. It is known that the rate of change of gene expression, which we will refer to as gene velocity, provides crucial information that enhances such inference; however, this information is not always available due to the limitations in sequencing depth. Our algorithm overcomes this limitation by estimating gene velocities using optimal transport. We then infer gene regulation using time-lagged correlation and Granger causality via regularized linear regression. Instead of providing an aggregated network across all time points, our method uncovers the underlying dynamical mechanism across time points. We validate our algorithm on 13 simulated datasets with both synthetic and curated networks and demonstrate its efficacy on 9 experimental data sets.

Hi-C experiments allow researchers to study and understand the 3D genome organization and its regulatory function. Unfortunately, sequencing costs and technical constraints severely restrict access to high-quality Hi-C data for many cell types. Existing frameworks rely on a sparse Hi-C dataset or cheaper-to-acquire ChIP-seq data to predict Hi-C contact maps with high read coverage. However, these methods fail to generalize to sparse or cross-cell-type inputs because they do not account for the contributions of epigenomic features or the impact of the structural neighborhood in predicting Hi-C reads. We propose GrapHiC, which combines Hi-C and ChIP-seq in a graph representation, allowing more accurate embedding of structural and epigenomic features. Each node represents a binned genomic region, and we assign edge weights using the observed Hi-C reads. Additionally, we embed ChIP-seq and relative positional information as node attributes, allowing our representation to capture structural neighborhoods and the contributions of proteins and their modifications for predicting Hi-C reads. We show that GrapHiC generalizes better than the current state-of-the-art on cross-cell-type settings and sparse Hi-C inputs. Moreover, we can utilize our framework to impute Hi-C reads even when no Hi-C contact map is available, thus making high-quality Hi-C data accessible for many cell types.

Gromov-Wasserstein distance has many applications in machine learning due to its ability to compare measures across metric spaces and its invariance to isometric transformations. However, in certain applications, this invariant property can be too flexible, thus undesirable. Moreover, the Gromov-Wasserstein distance solely considers pairwise sample similarities in input datasets, disregarding the raw feature representations. We propose a new optimal transport formulation, called Augmented Gromov-Wasserstein (AGW), that allows for some control over the level of rigidity to transformations. It also incorporates feature alignments, enabling us to better leverage prior knowledge on the input data for improved performance. We first present theoretical insights into the proposed method. We then demonstrate its usefulness for single-cell multi-omic alignment tasks and heterogeneous domain adaptation in machine learning.

SUMMARY: Single-cell Hi-C (scHi-C) protocol helps identify cell-type-specific chromatin interactions and sheds light on cell differentiation and disease progression. Despite providing crucial insights, scHi-C data is often underutilized due to the high cost and the complexity of the experimental protocol. We present a deep learning framework, scGrapHiC, that predicts pseudo-bulk scHi-C contact maps using pseudo-bulk scRNA-seq data. Specifically, scGrapHiC performs graph deconvolution to extract genome-wide single-cell interactions from a bulk Hi-C contact map using scRNA-seq as a guiding signal. Our evaluations show that scGrapHiC, trained on seven cell-type co-assay datasets, outperforms typical sequence encoder approaches. For example, scGrapHiC achieves a substantial improvement of 23.2% in recovering cell-type-specific Topologically Associating Domains over the baselines. It also generalizes to unseen embryo and brain tissue samples. scGrapHiC is a novel method to generate cell-type-specific scHi-C contact maps using widely available genomic signals that enables the study of cell-type-specific chromatin interactions. AVAILABILITY AND IMPLEMENTATION: The GitHub link: https://github.com/rsinghlab/scGrapHiC contains the source code of scGrapHiC and associated scripts to preprocess publicly available datasets to produce the results and visualizations we have discuss in this manuscript.

SUMMARY: Measurement of single-cell gene expression at different timepoints enables the study of cell development. However, due to the resource constraints and technical challenges associated with the single-cell experiments, researchers can only profile gene expression at discrete and sparsely sampled timepoints. This missing timepoint information impedes downstream cell developmental analyses. We propose scNODE, an end-to-end deep learning model that can predict in silico single-cell gene expression at unobserved timepoints. scNODE integrates a variational autoencoder with neural ordinary differential equations to predict gene expression using a continuous and nonlinear latent space. Importantly, we incorporate a dynamic regularization term to learn a latent space that is robust against distribution shifts when predicting single-cell gene expression at unobserved timepoints. Our evaluations on three real-world scRNA-seq datasets show that scNODE achieves higher predictive performance than state-of-the-art methods. We further demonstrate that scNODE's predictions help cell trajectory inference under the missing timepoint paradigm and the learned latent space is useful for in silico perturbation analysis of relevant genes along a developmental cell path. AVAILABILITY AND IMPLEMENTATION: The data and code are publicly available at https://github.com/rsinghlab/scNODE.

Optimal transport (OT) compares probability distributions by computing a meaningful alignment between their samples. CO-optimal transport (COOT) takes this comparison further by inferring an alignment between features as well. While this approach leads to better alignments and generalizes both OT and Gromov-Wasserstein distances, we provide a theoretical result showing that it is sensitive to outliers that are omnipresent in real-world data. This prompts us to propose unbalanced COOT for which we provably show its robustness to noise in the compared datasets. To the best of our knowledge, this is the first such result for OT methods in incomparable spaces. With this result in hand, we provide empirical evidence of this robustness for the challenging tasks of heterogeneous domain adaptation with and without varying proportions of classes and simultaneous alignment of samples and features across two single-cell measurements.

OBJECTIVE: Alzheimer's disease (AD) is the most common neurodegenerative disorder with one of the most complex pathogeneses, making effective and clinically actionable decision support difficult. The objective of this study was to develop a novel multimodal deep learning framework to aid medical professionals in AD diagnosis. MATERIALS AND METHODS: We present a Multimodal Alzheimer's Disease Diagnosis framework (MADDi) to accurately detect the presence of AD and mild cognitive impairment (MCI) from imaging, genetic, and clinical data. MADDi is novel in that we use cross-modal attention, which captures interactions between modalities-a method not previously explored in this domain. We perform multi-class classification, a challenging task considering the strong similarities between MCI and AD. We compare with previous state-of-the-art models, evaluate the importance of attention, and examine the contribution of each modality to the model's performance. RESULTS: MADDi classifies MCI, AD, and controls with 96.88% accuracy on a held-out test set. When examining the contribution of different attention schemes, we found that the combination of cross-modal attention with self-attention performed the best, and no attention layers in the model performed the worst, with a 7.9% difference in F1-scores. DISCUSSION: Our experiments underlined the importance of structured clinical data to help machine learning models contextualize and interpret the remaining modalities. Extensive ablation studies showed that any multimodal mixture of input features without access to structured clinical information suffered marked performance losses. CONCLUSION: This study demonstrates the merit of combining multiple input modalities via cross-modal attention to deliver highly accurate AD diagnostic decision support.

MOTIVATION: Gapped k-mer kernels with support vector machines (gkm-SVMs) have achieved strong predictive performance on regulatory DNA sequences on modestly sized training sets. However, existing gkm-SVM algorithms suffer from slow kernel computation time, as they depend exponentially on the sub-sequence feature length, number of mismatch positions, and the task's alphabet size. RESULTS: In this work, we introduce a fast and scalable algorithm for calculating gapped k-mer string kernels. Our method, named FastSK, uses a simplified kernel formulation that decomposes the kernel calculation into a set of independent counting operations over the possible mismatch positions. This simplified decomposition allows us to devise a fast Monte Carlo approximation that rapidly converges. FastSK can scale to much greater feature lengths, allows us to consider more mismatches, and is performant on a variety of sequence analysis tasks. On multiple DNA transcription factor binding site prediction datasets, FastSK consistently matches or outperforms the state-of-the-art gkmSVM-2.0 algorithms in area under the ROC curve, while achieving average speedups in kernel computation of ∼100× and speedups of ∼800× for large feature lengths. We further show that FastSK outperforms character-level recurrent and convolutional neural networks while achieving low variance. We then extend FastSK to 7 English-language medical named entity recognition datasets and 10 protein remote homology detection datasets. FastSK consistently matches or outperforms these baselines. AVAILABILITY AND IMPLEMENTATION: Our algorithm is available as a Python package and as C++ source code at https://github.com/QData/FastSK. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

International audience

Through sequence-based classification, this paper tries to accurately predict the DNA binding sites of transcription factors (TFs) in an unannotated cellular context. Related methods in the literature fail to perform such predictions accurately, since they do not consider sample distribution shift of sequence segments from an annotated (source) context to an unannotated (target) context. We, therefore, propose a method called "Transfer String Kernel" (TSK) that achieves improved prediction of transcription factor binding site (TFBS) using knowledge transfer via cross-context sample adaptation. TSK maps sequence segments to a high-dimensional feature space using a discriminative mismatch string kernel framework. In this high-dimensional space, labeled examples of the source context are re-weighted so that the revised sample distribution matches the target context more closely. We have experimentally verified TSK for TFBS identifications on 14 different TFs under a cross-organism setting. We find that TSK consistently outperforms the state-of-the-art TFBS tools, especially when working with TFs whose binding sequences are not conserved across contexts. We also demonstrate the generalizability of TSK by showing its cutting-edge performance on a different set of cross-context tasks for the MHC peptide binding predictions.

Motivation: Computational methods that predict differential gene expression from histone modification signals are highly desirable for understanding how histone modifications control the functional heterogeneity of cells through influencing differential gene regulation. Recent studies either failed to capture combinatorial effects on differential prediction or primarily only focused on cell type-specific analysis. In this paper we develop a novel attention-based deep learning architecture, DeepDiff, that provides a unified and end-to-end solution to model and to interpret how dependencies among histone modifications control the differential patterns of gene regulation. DeepDiff uses a hierarchy of multiple Long Short-Term Memory (LSTM) modules to encode the spatial structure of input signals and to model how various histone modifications cooperate automatically. We introduce and train two levels of attention jointly with the target prediction, enabling DeepDiff to attend differentially to relevant modifications and to locate important genome positions for each modification. Additionally, DeepDiff introduces a novel deep-learning based multi-task formulation to use the cell-type-specific gene expression predictions as auxiliary tasks, encouraging richer feature embeddings in our primary task of differential expression prediction. Results: Using data from Roadmap Epigenomics Project (REMC) for ten different pairs of cell types, we show that DeepDiff significantly outperforms the state-of-the-art baselines for differential gene expression prediction. The learned attention weights are validated by observations from previous studies about how epigenetic mechanisms connect to differential gene expression. Availability and implementation: Codes and results are available at deepchrome.org. Supplementary information: Supplementary data are available at Bioinformatics online.

The past decade has seen a revolution in genomic technologies that enabled a flood of genome-wide profiling of chromatin marks. Recent literature tried to understand gene regulation by predicting gene expression from large-scale chromatin measurements. Two fundamental challenges exist for such learning tasks: (1) genome-wide chromatin signals are spatially structured, high-dimensional and highly modular; and (2) the core aim is to understand what are the relevant factors and how they work together. Previous studies either failed to model complex dependencies among input signals or relied on separate feature analysis to explain the decisions. This paper presents an attention-based deep learning approach; AttentiveChrome, that uses a unified architecture to model and to interpret dependencies among chromatin factors for controlling gene regulation. AttentiveChrome uses a hierarchy of multiple Long Short-Term Memory (LSTM) modules to encode the input signals and to model how various chromatin marks cooperate automatically. AttentiveChrome trains two levels of attention jointly with the target prediction, enabling it to attend differentially to relevant marks and to locate important positions per mark. We evaluate the model across 56 different cell types (tasks) in human. Not only is the proposed architecture more accurate, but its attention scores also provide a better interpretation than state-of-the-art feature visualization methods such as saliency map.

Identifying context-specific entity networks from aggregated data is an important task, arising often in bioinformatics and neuroimaging applications. Computationally, this task can be formulated as jointly estimating multiple different, but related, sparse undirected graphical models (UGM) from aggregated samples across several contexts. Previous joint-UGM studies have mostly focused on sparse Gaussian graphical models (sGGMs) and can’t identify context-specific edge patterns directly. We, therefore, propose a novel approach, SIMULE (detecting Shared and Individual parts of MULtiple graphs Explicitly) to learn multi-UGM via a constrained $$\ell $$ 1 minimization. SIMULE automatically infers both specific edge patterns that are unique to each context and shared interactions preserved among all the contexts. Through the $$\ell $$ 1 constrained formulation, this problem is cast as multiple independent subtasks of linear programming that can be solved efficiently in parallel. In addition to Gaussian data, SIMULE can also handle multivariate Nonparanormal data that greatly relaxes the normality assumption that many real-world applications do not follow. We provide a novel theoretical proof showing that SIMULE achieves a consistent result at the rate $$O(\log (Kp)/n_{tot})$$ . On multiple synthetic datasets and two biomedical datasets, SIMULE shows significant improvement over state-of-the-art multi-sGGM and single-UGM baselines (SIMULE implementation and the used datasets @ https://github.com/QData/SIMULE ).

MOTIVATION: Histone modifications are among the most important factors that control gene regulation. Computational methods that predict gene expression from histone modification signals are highly desirable for understanding their combinatorial effects in gene regulation. This knowledge can help in developing 'epigenetic drugs' for diseases like cancer. Previous studies for quantifying the relationship between histone modifications and gene expression levels either failed to capture combinatorial effects or relied on multiple methods that separate predictions and combinatorial analysis. This paper develops a unified discriminative framework using a deep convolutional neural network to classify gene expression using histone modification data as input. Our system, called DeepChrome, allows automatic extraction of complex interactions among important features. To simultaneously visualize the combinatorial interactions among histone modifications, we propose a novel optimization-based technique that generates feature pattern maps from the learnt deep model. This provides an intuitive description of underlying epigenetic mechanisms that regulate genes. RESULTS: We show that DeepChrome outperforms state-of-the-art models like Support Vector Machines and Random Forests for gene expression classification task on 56 different cell-types from REMC database. The output of our visualization technique not only validates the previous observations but also allows novel insights about combinatorial interactions among histone modification marks, some of which have recently been observed by experimental studies. AVAILABILITY AND IMPLEMENTATION: Codes and results are available at www.deepchrome.org CONTACT: [email protected] SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.