IUT Institutional Repository

IUT Institutional Repository is a digital service that collects, preserves, and distributes digital material. Repositories are important tools for preserving an organization's legacy; they facilitate digital preservation and scholarly communication.

Communities in IUT Institutional Repository

Select a community to browse its collections.

Now showing 1 - 5 of 12

Recent Submissions

GNN andTransformer Fusion Learning for Molecular Classification of BACE1 Inhibitors
(Department of Computer Science and Engineering(CSE), Islamic University of Technology(IUT), Board Bazar, Gazipur-1704, Bangladesh, 2025-10-25) Shadid, Md. Abu Hena; Tabassum, Mahajabin
Alzheimer’s disease (AD) is a progressive and devastating neurodegenerative disor der, primarily manifested through memory loss and cognitive decline [1], [2]. One of the central pathological hallmarks of AD is the accumulation of amyloid-beta (A𝛽) plaques, formed via the sequential cleavage of the amyloid precursor protein (APP) by 𝛽-secretase (BACE1) and 𝛾-secretase [3]. Inhibiting BACE1 is therefore regarded as a compelling therapeutic strategy, as it can impede the formation of neurotoxic A𝛽 aggregates [4], [5]. Nevertheless, the identification of effective BACE1 inhibitors remains arduous and resource-intensive when approached through conventional ex perimental pipelines. In this study, we propose a hybrid deep learning framework that fuses Graph Neural Networks (GNNs) with ChemBERTa, a transformer model pretrained on large chemical corpora. While GNNs capture atom-level and bond level interactions (local structural dependencies), ChemBERTa encodes long-range dependencies and semantic patterns from SMILES representations (global chemical context). By unifying these complementarymodalities, ourmodelovercomesthelim itations of prior GNN+CNN approaches, where CNNs process sequential SMILES in a strictly local fashion and fail to capture non-linear long-range dependencies across molecular structures. Our GNN–ChemBERTa fusion model achieved an accuracy of 92.77% inclassifying active versus inactive BACE1 inhibitors, demonstrating superior predictive power and generalization. Beyond its performance, the model contributes to reducing drug discovery costs, accelerating virtual screening, and minimizing the need for extensive laboratory experimentation. Moreover, a recall value of 93% in dicates that almost all potential active molecules were successfully identified by the model, minimizing the risk of missing true inhibitors. Similarly, a high precision value of 93% demonstrates that the model produces very few false positives, thereby reducing unnecessary laboratory costs associated with testing inactive compounds. Additionally, the ROC–AUC score of 87.88% confirms that the model can effectively distinguish between active and inactive molecules, reflecting strong overall classifica tion performance. By enabling efficient in silico identification of potential inhibitors, this approach not only streamlines the early stages of Alzheimer’s drug development but also holds promise for broader application to other therapeutic targets associated with neurodegenerative diseases.
Medical Anomaly Detection Using Generative Adversarial Network With Self Attention Mechanism
(Department of Computer Science and Engineering(CSE), Islamic University of Technology(IUT), Board Bazar, Gazipur-1704, Bangladesh, 2025-10-25) Fuad, Fahim Abrar; Kaisar, Saqif; Ananta, Tanvir Hassan
Anomaly detection in medical imaging plays a vital role in assisting early diagnosis and treatment. Traditional supervised methods rely heavily on annotated abnormal samples, which are often scarce and diverse, making them impractical for real-world deployment. This work explores unsupervised anomaly detection using a range of GAN-based frameworks, where models learn to capture the distribution of normal data and identify deviations without the need for labeled outliers. Various architec tures, including reconstruction-based and feature-matching approaches, are evalu ated and extended with enhancements such as self-attention mechanisms and posi tional encodings to improve spatial feature learning. Extensive experiments on medi cal imaging datasets demonstrate that the proposed techniques significantly improve detection performance, highlighting the effectiveness of adversarial learning for un supervised medical anomaly detection.
MixSarc: A Bangla-English Code-Mixed Corpus For Implicit Meaning Identification
(Department of Computer Science and Engineering(CSE), Islamic University of Technology(IUT), Board Bazar, Gazipur-1704, Bangladesh, 2025-10-25) Ahmed, Tamim; Alam, Kazi Samin Yasar; Chowdhury, Md Tanbir
Thisthesisfocusesondetectinghumor,sarcasm,offensiveness,andvulgarityinBangla English code-mixed text, an area largely overlooked in existing natural language pro cessing (NLP) research. A novel dataset has been proposed, which will be created by scraping and filtering social media content, followed by manual annotation across fourattributes. Twotransformer-basedapproacheswereexploredinsmallscale: multi class and multi-label text classification. The study also proposes future directions, in cluding dataset balancing, comparative evaluation of transformer models and large language models (LLMs), and the introduction of a SarOff Score to better capture sarcasm-offense overlap. By addressing the complexities of code-mixed tone detec tion, this work advances NLP in low-resource, multilingual settings
LLM-BasedAuto-Labeling of Developer Discussions AComparative Study of Zero-Shot, Sampling Methods, Ensembles and Judge-Guided Strategies
(Department of Computer Science and Engineering(CSE), Islamic University of Technology(IUT), Board Bazar, Gazipur-1704, Bangladesh, 2025-10-25) Shakhawat, Chowdhury Ashfaq; Soyeb, Md; Haque, Iftekharul
Software bugs have long posed challenges to the delivery of reliable digital services, promptingextensiveresearch intoautomatedbuglabeling. Whilesignificantadvance ments have been made, existing approaches often struggle with high false positive rates and face difficulties in practical deployment due to reliance on structured bug reports. Most contemporary studies utilize structured datasets containing developer generated bug reports, typically written in natural language. These reports require manual or semi-automated extraction of relevant inputs, a process that is both time consuming and error-prone. With the emergence of Large Language Models (LLMs), a new research opportu nity arises: can LLMs effectively extract failure-inducing inputs from unstructured, community-driven sources such as GitHub, Stack Overflow, and other developer fo rums? In this study, we propose a novel end-to-end pipeline that leverages LLMs for bug labeling directly from raw, unstructured text. Our methodology focuses on au tomated labeling, utilizing prompt-based approaches to optimize the performance of generative models. Wecuratedandannotateda datasetcomprising1885StackOverflow questions posted between 2023 and2025,andfurthervalidatedourapproachusingadatasetofGitHub issue reports. Through extensive experimentation, we assess the accuracy and ro bustness of our pipeline across diverse input formats. Unlike existing solutions, our proposed framework emphasizes simplicity, scalability, and cost-effectiveness, mak ing it well-suited for integration into real-world software development workflows.
From Lightweight CNNs to SpikeNets: Benchmarking Accuracy–Energy Tradeoffs with Pruned Spiking SqueezeNet
(Department of Computer Science and Engineering(CSE), Islamic University of Technology(IUT), Board Bazar, Gazipur-1704, Bangladesh, 2025-10-25) Dipto, Tawsif Tashwar; Kabir, Radib Bin; Ahamed, Mehedi
SpikingNeuralNetworks(SNNs)areemergingasenergy-efficientalternativestoCon volutional Neural Networks (CNNs) for edge-device deployment. However, SNNs typically exhibit a performance gap compared to CNNs, and training them remains challenging due to the non-differentiability of spiking neurons. In this work, we sys tematically analyze lightweight CNN-to-SNNconversionpipelinesenhancedwithpa rameters optimization and pruning. We benchmarked models on CIFAR-10, CIFAR 100, and TinyImageNet, evaluating accuracy, F1-score, parameter count, computa tional complexity, and energy consumption. CNNs are profiled using MAC opera tions, while SNNs are measured with Accumulate (AC) and MAC operations to cap ture event-driven sparsity. Our results show that SNNs achieve up to 15.7× higher energy efficiency compared to CNN baselines while maintaining comparable accu racy. Notably, pruning the SNN SqueezeNet improved CIFAR-10 accuracy by 6% and reduced parameters by19%comparedtotheoriginalSNNSqueezeNet. Moreover,the pruned SNN SqueezeNet-P exhibits only a 1% lower accuracy than CNN SqueezeNet while achieving an88.1%improvementinenergyefficiency. Thesefindingshighlight the potential of SNNs as low-power alternatives for edge intelligence and provide a benchmarked evaluation framework for future studies on energy-efficient neural net works.