SSSumaiya Shomaji

Sumaiya Shomaji

*This is an algorithmically generated author record based on publicly available information.

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article


Sumaiya Shomaji has a total of 14 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • AI-Powered Virtual Cell

    by Fairuz Shadmani Shishir, Rokunuzjahan Rudro, Sumaiya Shomaji

    The convergence of artificial intelligence (AI) and virtual cell technology represents a transformative paradigm in computational biology, fundamentally reshaping how we model, understand, and predict cellular behavior. This comprehensive review examines the revolutionary integration of deep learning, machine learning, and AI-driven approaches in virtual cell platforms, with emphasis on their applications in systems biology, drug discovery, and personalized medicine. We explore how neural networks, particularly deep learning architectures, are being employed to learn cellular dynamics directly from experimental data, bypassing traditional mechanistic modeling limitations. The review covers AI-enhanced virtual cell platforms, including neural ordinary differential equations (NODEs), graph neural networks for biochemical pathways, and reinforcement learning for cellular control systems. We discuss breakthrough applications in AI-driven drug discovery, where virtual cells powered by deep learning accelerate therapeutic compound screening and toxicity prediction. The integration of computer vision techniques for cellular image analysis, natural language processing for biological knowledge extraction, and generative models for synthetic cellular data creation is thoroughly examined. Current challenges, including interpretability, data requirements, and validation, are addressed alongside emerging solutions. We analyze the role of foundation models, transformer architectures, and multi-modal AI systems in advancing virtual cell technology toward autonomous biological discovery. This review demonstrates how AI is not merely augmenting traditional virtual cell approaches but fundamentally transforming them into intelligent, adaptive systems capable of autonomous learning and prediction.

  • Representation Learning for ECG Data

    by Christopher J. Harvey, Amulya Gupta, Sumaiya Shomaji, Zijun Yao, Amit Noheria

    Deep learning (DL) is being applied to electrocardiogram (ECG) analysis, but the common setup of large convolutional networks trained end-to-end on raw waveforms for a single downstream task is data-hungry, computationally expensive, and poorly explainable. Over the last five years, work has shifted toward representation learning. Rather than fitting a separate model for each endpoint, the objective is to learn reusable ECG latent spaces that support compression, predictions, simulation, and some degree of explanation. This chapter focuses on three families of methods that underpin this shift: (i) variational autoencoders (VAEs), (ii) contrastive learning (CL), and (iii) transformer-based ECG foundation models (ECG-FMs). These approaches (i) reduce ECG data dimensionality by much as two to three orders of magnitude while retaining useful information, (ii) often improve performance and robustness esp. when labeled training data are scarce, and (iii) provide opportunities for explainability of model predictions. We summarize the design, variations, strengths, and limitations of different methods and outline how hybrid representation learning can be used to build ECG models that are not only accurate, but also adaptable and computationally tractable for clinical deployment.

  • Parameter Efficient Deep Learning Models for Multi-Target Binding Affinity and hERG Cardiotoxicity Prediction

    by Fairuz Shadmani Shishir, Bishnu Sarker, Cuncong Zhong, Sumaiya Shomaji

    Accurately predicting binding affinity and toxicity is critical to drug discovery, offering the potential to reduce development costs and enhance safety profiles. Traditional computational methods, such as molecular docking and quantitative structure-activity relationship (QSAR) models, rely heavily on handcrafted features and domain expertise, limiting their scalability and generalizability to novel compounds. Recent advances in chemical language models (CLMs), which leverage Simplified Molecular Input Line Entry System (SMILES) representations, offer a scalable alternative by capturing complex molecular properties through learned representations. In this study, we propose a dual-paradigm framework for binding affinity prediction of 3 protein targets and cardiotoxicity for human ether-á-go-go related gene (hERG). The first component utilizes a specialized graph neural network (GNN) that operates directly on molecular graphs, treating atoms as nodes and bonds as edges. The second component employs a CLM fine-tuned via Low-Rank Adaptation (LoRA), enabling efficient adaptation of large pre-trained models to specific toxicological tasks with minimal parameter overhead. Our hybrid framework demonstrates superior performance, achieving an average AUROC of 0.92 across three protein targets, while the LoRA-adapted CLM attains an AUROC of 0.93 for cardiotoxicity prediction, offering a 98% reduction in trainable parameters, outperforming the performance of existing models.

  • Open Access

    A Concept-Driven Disentanglement Framework for Interpretable Graph Neural Networks in Structure-Function Coupling

    by Devin Setiawan, Sumaiya Shomaji, Joaquin Goni, Arian Ashourvan

    Graph Neural Networks (GNNs) achieve state-of-the-art performance in predicting brain functional connectivity (FC) from structural connectivity (SC), yet their "black-box" nature limits interpretability and scientific utility. We present a concept-driven disentanglement framework that builds inherently interpretable GNNs for quantitative hypothesis testing of structure-function relationships. The framework employs an ensemble of GNN branches, each architecturally biased to learn from a predefined structural concept (e.g., strong vs. weak connections) by processing a filtered version of the SC graph. This design enforces verifiably disentangled node embeddings, ensuring each branch captures a distinct structural feature. Using SHAP (Shapley additive explanations), we quantify the predictive contribution of each concept and assess its statistical significance against null distributions. Our framework demonstrates high predictive accuracy for FC, achieving a group-level correlation coefficient of 0.91 on a public human connectome dataset, while simultaneously yielding interpretable neuroscientific insights. This interpretable-by-design methodology bridges the gap between predictive power and scientific transparency, enabling deep learning models to provide mechanistic insights into the brain's network organization.

  • A Deep Learning Framework for Protein-to-Metal Binding Prediction Using Protein Language Models

    by Fairuz Shadmani Shishir, Bishnu Sarker, Farzana Rahman, Sumaiya Shomaji

    This study presents an end-to-end deep learning framework for protein-to-metal-ion binding prediction, a critical task in understanding protein function, structural stability, and metal transport mechanisms. A binding site is a residue location in a protein sequence where a metal binds to a protein. Manual curation of metal binding sites is a tedious process involving mining through research articles, making it expensive, laborious, and time-consuming. Therefore, developing a computational pipeline is essential to predict metal ion binding of unannotated proteins. A significant shortcoming of existing computational methods is the failure to capture the long-term dependency of the residues, the absence of positional information, and a pre-determined set of residues and metal ions. In this paper, we propose a metal-ion binding prediction pipeline using a large language model, emphasizing 1) the comparative performance of five state-of-the-art protein language models (pLMs), 2) the impact of positional encoding of binding sites, and 3) the comparison with classical machine learning techniques. A 10-fold cross-validation evaluation yielded a Matthews Correlation Coefficient (MCC) of 0.89, along with precision, recall, and F1 scores exceeding 95% for the six most extensively studied metal ions reported in the literature.