M. A. Hakim Newton

Protein-ligand pose prediction is a core task in structure-based drug discovery because it determines how a ligand fits within a protein pocket and directly affects downstream virtual screening and lead-optimization workflows. Recent graph neural network (GNN) methods have shown promise for protein-ligand pose prediction, while improving the accuracy of an initial docked pose remains an important task. In this work, we present KTransPose, a GNN-based protein-ligand pose-refinement framework that combines a dual-branch architecture, a docking-oriented multi-component loss, and iterative refinement. The dual-branch architecture combines a global pose transformation branch that predicts rotation and translation with a local coordinate correction branch that predicts atom-wise coordinate adjustments. The multi-component loss combines root mean squared distance (RMSD) when computed without rotational or translational alignment, mean squared error (MSE) after alignment by the Kabsch algorithm, intraligand distance regularization, and a protein-ligand clash penalty. We instantiate the framework with several GNN architectures and find that a TransformerConv-based model provides the strongest overall performance. Using the same preprocessing, initial docked poses, and evaluation procedure, KTransPose improves upon MedusaGraph on both PDBbind-2020 and CASF-2016 benchmark datasets, reducing the mean RMSD from 5.08 Å to 4.47 Å and from 4.81 Å to 4.24 Å respectively, corresponding to improvements of approximately 12% in both datasets.

Anti-money laundering (AML) machine learning research is limited by the restricted availability of real banking transaction data. Synthetic benchmarks can support reproducible evaluation. However, existing AML datasets and generators, both statistical and agent-based, do not always jointly represent reporting thresholds, staged laundering behavior, temporal variation, routing structure, and overlap between ordinary and suspicious activity. This paper presents AMLNet, a knowledge-guided framework for synthetic AML benchmark generation and machine learning evaluation. AMLNet generates suspicious activity as staged transaction sequences shaped by regulatory thresholds, laundering typologies, timing rules, routing constraints, and ordinary customer behavior. The benchmark instance contains 1.09 million transactions with a suspicious rate of 0.13%. Domain-informed checks show class imbalance, heavy-tailed transaction amounts, ordinary–suspicious amount overlap, staged laundering patterns, multi-hop routing, and variation across payment channels and locations. We evaluate AMLNet using transaction-level and node-level machine learning branches with representative tabular and graph-based detectors. Detector comparisons show different strengths across models: Random Forest gives the highest node-level F1-score, XGBoost gives the highest ROC-AUC, and GraphSAGE gives the highest PR-AUC under the AMLNet node-level protocol. Feature ablations show that temporal and structural features improve detection beyond basic transaction attributes. Cross-domain experiments on Elliptic++ show that direct transfer is difficult, but fine-tuning an AMLNet-trained graph model gives ROC-AUC and PR-AUC close to target-only training. AMLNet provides a reusable benchmark and evaluation protocol for studying machine learning methods under rare-event AML conditions.

Online education is common in the current digital world. There has been some progress in adopting synchronous online education in the higher education sector, using general purpose online meeting software. The current global trend is to adopt human-technology integration to achieve better classroom engagement with the help of blending learning and to achieve equal or better performance than physical learning. In this paper, we investigate the research question “What are the key factors that affect the perceived quality ratings of online university classrooms?” To find the answers, using expert and convenience sampling methods, we perform web-based online quantitative cross-sectional surveys on university teachers and students. We use correlation analysis and decision trees as educational data mining methods to analyse and explain the collected data. Our findings from this investigation are that having appropriate online tools, having interactions in online classes, getting math lessons and labs well, and lively and engaging teaching methods are important for students. For teachers, getting participation, interaction, and feedback from students are key to improve online classes. Our study also seeks a hierarchical explanation behind the relations between the perceived quality ratings and the identified factors. Our findings would help higher education researchers in redesigning future online classrooms.

The Internet of Things (IoT) has transformed various sectors, from home automation to healthcare, leveraging a multitude of sensors and actuators communicating through cloud, fog, and edge networks. However, the diversity in device manufacturing and communication protocols necessitates interoperable communication interfaces. Most existing IoT interoperability solutions often rely on cloud-based centralised architectures and suffer from latency and scalability issues. This work specifically focuses on scenarios where decisions need to be made with IoT edge devices in real-time, even in situations where there might be internet disruptions, low bandwidth, or no internet connection. While typical IoT interoperability solutions support edge devices, their reliance on cloud-based architectures makes them unsuitable for mission-critical applications, environmental monitoring, or water quality monitoring, where internet connectivity cannot be guaranteed. To tackle these challenges, the project InterEdge proposed a theoretical interoperability model supporting hierarchical decentralised communication between edge devices. The aforementioned framework has four levels to handle network, syntactic, semantic, and organisational aspects of interoperability. As part of the same project, this work focuses on the implementation of the syntactic and semantic levels of the aforementioned framework. This work involves tackling the implementation challenges, particularly considering key issues related to transmission latency and memory requirements. We have created profiles for edge devices and data formats to store their essential and extra information. Using the profiles, communications can be established and maintained seamlessly among edge devices. We have conducted a comparative analysis between InterEdge implementation and three other implementations of established open standards. The experimental results demonstrate that the syntactic and semantic levels of the implemented interoperability solution, InterEdge, significantly outperforms the existing open standards in terms of standard benchmarking metrics such as code size, memory usage, and response latency. The contribution of this paper lies in these implementation results, which provide concrete evidence of the superior performance of our proposed solution, InterEdge, thereby validating its efficacy in real-world IoT scenarios.

Forecasting accurate water quality is very important in aquaculture, environment monitoring, and many other applications. Many internal and external factors influence water quality. Therefore, water quality parameters exhibit complex time series characteristics. Consequently, long-range accurate prediction of water quality parameters suffers from poor propagation of information from past timepoints to further future timepoints. Moreover, to synchronise the prediction model with the changes in the time series characteristics, periodic retraining of the prediction model is required and such retraining is to be done on resource-restricted computation devices. In this work, we present a low-cost training approach to improve long-range multi-step water quality prediction. We train a short-range predictor to save training effort. Then, we strive to achieve and/or improve long-range prediction using multi-step iterative ensembling during inference. Experimental results on 9 water quality datasets demonstrate that the proposed method achieves significantly lower error than the existing state-of-the-art approaches. Our approach significantly outperforms the existing approaches in several standard metrics, even in the case of future timepoints at long distances.