Sanghamitra Bandyopadhyay

Summary: Epigenetic enzymes, writers, readers and erasers regulate chromatin landscapes and participate in tumor heterogeneity. While therapeutic targeting of these enzymes has shown clinical promise, the comparative efficacy of mono-versus dual-inhibitor strategies remain unclear. Here, we introduce a multi-modal platform that uses NicE-viewSeq and integrates automated deep learning based spatially resolved chromatin accessibility profiling with high-throughput sequencing following epigenetic inhibitor application. Accessible chromatin landscapes were altered along with nucleosome positioning following inhibition of either LSD1 or HDACs alone, or both together. Coordinated modulation of histone marks and the CoREST complex on chromatin was observed across inhibitory conditions. Transcription factor binding analysis identified three predominant families, ETS, RUNT, and bZIP with enhanced chromatin association upon treatments. Mechanistically, a CoREST-RUNX regulatory axis was uncovered wherein JunB, a member of bZIP family displaces CoREST-RUNX at differentially accessible regions, triggering apoptotic pathways. Therefore, JunB-mediated mechanism reveals a convergent therapeutic vulnerability, offering new avenues for optimizing different combinatorial epigenetic therapy in cancer.

Background: Cancer heterogeneity results in patients with the same diagnosis responding differently to drugs, making treatments extremely challenging. Advances in computational power enable personalized treatments that suppress tumors and extend patient survival. Therefore, accurate prediction of cancer cell response to a particular medication is of utmost importance. Current deep learning-based models have achieved impressive accuracy, but they often function as a “black box” and cannot explain the reason for the prediction. To address this limitation, we develop a deep learning-based model, BKDRP, which incorporates prior biological information into the architecture, along with molecular fingerprints of drugs, while embedding biological priors into its architecture. Specifically, it incorporates the fact that genes encode proteins that combine to form protein complexes, which in turn regulate biological pathways, ultimately targeted by drugs. Results: We evaluate BKDRP on the GDSC (Genomics of Drug Sensitivity and Cancer) cell line dataset using multi-omics gene expression, protein expression, mutation, and copy number variation. Four rigorous experiments have been conducted to test the model’s generalizability: prediction of unknown drug–cell line responses, responses to unseen drugs (LODO: Leave-One-Drug-Out), responses to unseen cell lines (LOCLO: Leave-On-Cell-Line-Out), and responses across unseen cancer types (LOCO: Leave-One-Cancer-Out). The performance of the proposed method and baseline algorithms is assessed using two metrics: Area Under the ROC Curve (AUC) and Area Under the Precision-Recall Curve (AUPR). The experimental results demonstrate that BKDRP performs well in different evaluation techniques. Notably, BKDRP has achieved an AUC of 0.8845, surpassing traditional machine learning and deep learning approaches and demonstrating robustness in handling biological variability across cancer types. A case study of lung adenocarcinoma (LUAD) highlights known biomarkers ( KRAS , EGFR , STK11 ), key proteins (SOCS1, HSPA8, SMC3), and drugs (Erlotinib, Palbociclib) that are consistent with the literature. Conclusions: In conclusion, BKDRP presents a novel biological knowledge-driven deep neural network model for cancer drug response prediction that shows strong predictive accuracy and interpretability. By integrating multi-omics data and incorporating domain knowledge, BKDRP has the strong potential for applications in biomarker discovery and the advancement of personalized oncology.

One of the key challenges in spatial transcriptomics data analysis is the lack of sufficient data to train models. To address this shortcoming, multiple generative models have been developed to generate synthetic spatial transcriptomics samples in a controlled environment. However, these models often fail in out-of-the-box generation in the presence of noise (such as outliers). To tackle this challenge, we propose RSTG (Robust Spatial Transcriptomic Generator), an autoencoder that incorporates the β -ELBO loss, to generate realistic and high-quality spatial transcriptomic sequences. Our model uncovers data’ intrinsic structure by approximating its underlying distribution through variational inference, resulting in more interpretable and robust density estimation. We validate the effectiveness of RSTG across multiple tasks, including the recovery of cellular positions in both the 2D spatial and location domains. Our method shows improved performance, both qualitatively and quantitatively, across multiple datasets, including brain and liver samples generated using MERFISH, MERSCOPE, and Visium technologies. We further illustrate the robustness of RSTG to outliers by contaminating a portion of the data with anomalies (such as white noise, batch effects, and dropouts) as well as on a reallife degraded sample. The results show that our proposal maintains high quality and stability even when the training data are contaminated, across a variety of experimental settings and in comparison with existing approaches.

The increasing adoption of graph neural networks (GNNs) in critical real-world applications brings with it the need for interpretability and explainability of these models. The existing paradigms of GNN explainability typically operate at the extremes. Instance-level methods offer a microscopic view focused on individual predictions, whereas model-level approaches provide a macroscopic summary of the classifier behavior on a target class. We introduce the group-level paradigm, which views explanations at the right resolution, grounded in the intuition that groups of instances embedded close together in the latent space share a common rationale for belonging to the same target class. We show that the instance-level and model-level paradigms are the special cases of this broader framework. To realize this paradigm, we propose GroupEx, a novel method for interpreting GNNs at the group level. GroupEx has the capability to identify the subgroups of graphs within a target class using a group identity (GI) score. It employs a group extractor that can extract a subgroup of graphs that share a common explanation for their class identity by clustering graph embeddings in the latent space of the classifier. A specially devised explanation generator can then discover a group-level explanation by optimizing a novel isomorphism objective, which ensures that the discovered explanation commonly explains the class identity of all the instances in the group. Empirical results on real-world and synthetic datasets demonstrate that GroupEx provides deeper insights into GNN decision-making than the existing state-of-the-art explainability methods, enabling a more structured and interpretable understanding of model predictions.

One of the key challenges in spatial transcriptomics data analysis is the lack of sufficient data to train models. To address this shortcoming, multiple generative models have been developed to generate synthetic spatial transcriptomics samples in a controlled environment. However, these models often fail in out-of-the-box generation in the presence of noise (such as outliers). To tackle this challenge, we propose RSTG (Robust Spatial Transcriptomic Generator), an autoencoder that incorporates the $\beta$-ELBO loss, to generate realistic and high-quality spatial transcriptomic sequences. Our model uncovers data' intrinsic structure by approximating its underlying distribution through variational inference, resulting in more interpretable and robust density estimation. We validate the effectiveness of RSTG across multiple tasks, including the recovery of cellular positions in both the 2D spatial and location domains. Our method shows improved performance, both qualitatively and quantitatively, across multiple datasets, including brain and liver samples generated using MERFISH, MERSCOPE, and Visium technologies. We further illustrate the robustness of RSTG to outliers by contaminating a portion of the data with anomalies (such as white noise, batch effects, and dropouts) as well as on a real-life degraded sample. The results show that our proposal maintains high quality and stability even when the training data are contaminated, across a variety of experimental settings and in comparison with existing approaches.

Classifying drugs into symptomatic (SYM) and disease-modifying (DM) categories is essential for understanding their therapeutic effect and plays a key role in drug repurposing. While many computational approaches focus on drug-target prediction, they often ignore the nature of the drug's action on disease progression. This study proposes a graph-based strategy to classify drugs as SYM or DM based on their effect on disease treatment. We construct a heterogeneous network comprising genes, diseases, and drugs, and apply a guided shortest path traversal framework for drug effect classification. During this traversal, certain genes appear frequently in the shortest metapaths linking diseases and drugs. These recurrent genes are identified based on their frequency of occurrence in known drug-disease paths. For a new drug-disease pair, if the traversal path contains recurrent genes marked for a specific treatment type, we classify the drug accordingly. Over and above classifying the drugs, the proposed method incorporates the metapath-based framework to improve interpretability. Experimental results show that our model achieves significantly better classification accuracy compared to advanced machine learning and deep learning methods. A case study on multiple sclerosis further supports the biological relevance of our approach. All data and scripts supporting the findings of this study are publicly available at https://github.com/aishik11/drug_effect_classification

Background: Cancer heterogeneity results in patients with the same diagnosis responding differently to drugs, making treatments extremely challenging. Advances in computational power enable personalized treatments that suppress tumors and extend patient survival. Therefore, accurate prediction of cancer cell response to a particular medication is of utmost importance. Current deep learning-based models have achieved impressive accuracy, but they often function as a "black box'' and cannot explain the reason for the prediction. To address this limitation, we develop a deep learning-based model, BKDRP, which incorporates prior biological information into the architecture, along with molecular fingerprints of drugs, while embedding biological priors into its architecture. Specifically, it incorporates the fact that genes encode proteins that combine to form protein complexes, which in turn regulate biological pathways, ultimately targeted by drugs. Results: We evaluate BKDRP on the GDSC (Genomics of Drug Sensitivity and Cancer) cell line dataset using multi-omics gene expression, protein expression, mutation, and copy number variation. Four rigorous experiments have been conducted to test the model's generalizability: prediction of unknown drug–cell line responses, responses to unseen drugs (LODO: Leave-One-Drug-Out), responses to unseen cell lines (LOCLO: Leave-On-Cell-Line-Out), and responses across unseen cancer types (LOCO: Leave-One-Cancer-Type-Out). The performance of the proposed method and baseline algorithms is assessed using two metrics: Area Under the ROC Curve (AUC) and Area Under the Precision-Recall Curve (AUPR). The experimental results demonstrate that BKDRP performs well in different evaluation techniques. Notably, BKDRP has achieved an AUC of 0.8776, surpassing traditional machine learning and deep learning approaches and demonstrating robustness in handling biological variability across cancer types. A case study of lung adenocarcinoma (LUAD) highlights known biomarkers (KRAS, EGFR, STK11), key proteins (SOCS1, HSPA8, SMC3), and drugs (Erlotinib, Palbociclib) consistent with the literature. Conclusions: In conclusion, BKDRP presents a novel biological knowledge-driven deep neural network model for cancer drug response prediction that shows strong predictive accuracy and interpretability. By integrating multi-omics data and incorporating domain knowledge, BKDRP has the strong potential for applications in biomarker discovery and the advancement of personalized oncology.

Graph neural networks (GNNs) have become the prevailing methodology for addressing graph data-related tasks, permeating critical domains like recommendation systems and drug development. The necessity for trustworthiness and interpretability of GNNs has risen to the forefront, especially given their direct impact on end users’ lives. To address this need, we present Gen-GraphEx, a model-agnostic, model-level explanation method that prioritizes user centricness by eliminating the need for having access to the hidden layers of the GNN model it seeks to explain. Given a particular class label, Gen-GraphEx learns a graph generative model (GGM) that produces explanation graphs that not only contain discriminative patterns that the GNN has learned for that class but also lie in distribution with real graphs that belong to that class according to the GNN. Unlike existing state-of-the-art models, Gen-GraphEx also has the unique ability to interpolate the GGMs of two target classes to generate instances that lie near the decision boundary of the two classes giving a deeper insight into the model’s decision-making. Its advantages over existing methods in the literature also include nonreliance on another subsequent deep learning module for explanation generation, ability to generate graphs with various node and edge features, and being more computationally efficient. Extensive validation and thorough comparative analysis of the proposed approach is carried out across an array of real and synthetic datasets that consistently demonstrate its exceptional performance and competitiveness ranking alongside state-of-the-art model-level explainers. Our code is available at https://github.com/amisayan/Gen-GraphEx

The development of new drugs is an expensive and time-consuming process, often hindered by the lack of reliable models to predict drug-target interactions (DTIs) and their mechanisms of action (MoA). Existing deep learning-based methods for DTI prediction typically focus only on binary classification of interactions, overlooking the complex mechanisms underlying these interactions. Moreover, the absence of comprehensive datasets for modeling MoA further complicates this task. To address these limitations, we introduce DrugMAP, a novel multimodal deep learning model that integrates graph neural networks and transformer-based architectures to predict both DTIs and their MoA. We construct a large-scale dataset from multiple public sources, adding a new level of complexity by including detailed MoA annotations for thousands of drug-target pairs. DrugMAP simultaneously leverages the molecular and atomic-level structures of drugs and target proteins, utilizing multi-representational encoders for enhanced feature extraction. Experimental results show that DrugMAP outperforms state-of-the-art models for both DTI and MoA prediction across multiple benchmark datasets. Our model achieves a 3.5% improvement in AUC for MoA prediction, demonstrating its potential for guiding drug discovery and understanding adverse drug events. The model and dataset are available at https://github.com/JUCompBio/DTI-MOA