Drug repurposing offers a faster, lower-cost alternative to de novo drug development, particularly for chronic and neurodegenerative diseases where conventional discovery pipelines are slow and resource-intensive. This paper proposes OTC-DTI (Optimized Transformer-CNN Drug-Target Interaction), a multi-stage framework that combines a Transformer-CNN encoder-decoder for drug-target interaction (DTI) prediction with Bayesian hyperparameter optimization and transfer learning, and applies the resulting model to repurpose existing drugs for Parkinson's disease, a neurodegenerative disorder with substantial unmet therapeutic need. On the Davis and KIBA benchmarks, OTC-DTI achieves a concordance index (CI) of 0.882 and mean squared error (MSE) of 0.255 on Davis, and a CI of 0.885 and MSE of 0.174 on KIBA, improving on the DeepDTA baseline by approximately 2.3% (MSE) and 0.5% (CI) on Davis and 10.3% (MSE) and 2.6% (CI) on KIBA, while remaining competitive with more recent graph- and attention-based architectures. Applying the optimized model to known Parkinson's disease signature genes via the Broad Repurposing Hub, OTC-DTI identifies several candidate compounds, including rifamycin and lycopene, with literature-supported plausibility for further investigation. These results demonstrate that combining hybrid deep learning architectures with Bayesian optimization and transfer learning is a practical and effective strategy for DTI-driven drug repurposing.
Recent advances in Large Language Models (LLMs) have significantly improved the quality of abstractive text summarization. However, the high computational cost and memory requirements of these models limit their practical adoption, particularly in resource-constrained environments. This focused survey reviews recent progress in efficient LLM-based abstractive text summarization, emphasizing methods designed to operate under realistic hardware constraints. The survey first provides a concise background on abstractive and extreme summarization, highlighting challenges introduced by long contexts and aggressive compression. A structured taxonomy is then proposed to categorize existing approaches according to model architecture, training strategy, adaptation technique, and hardware awareness. Building on this taxonomy, the paper analyzes key efficiency-oriented techniques, including parameter-efficient fine-tuning, low-bit quantization, precision optimization, and inference-time acceleration. Furthermore, commonly used evaluation protocols are reviewed, emphasizing the limitations of traditional automated metrics and the growing need for faithfulness- and cost-aware evaluation. A comparative discussion summarizes trade-offs between summarization quality and computational efficiency, offering practical guidelines for model selection and deployment. Finally, open challenges and future research directions are outlined, including faithful extreme summarization, long-context processing, multimodal summarization, and standardized efficiency-aware benchmarks. Overall, this survey provides researchers and practitioners with a clear, structured overview of practical considerations for efficient LLM-based text summarization.
Location-based social networks (LBSNs) generate extensive user check-in data that reflect mobility behaviors and activity preferences. These check-ins reveal frequent sequential short-term mobility patterns that can be exploited to predict users’ upcoming activities and next locations with higher accuracy. Accurately predicting user location trajectories from such data remains a challenging task due to data sparsity, noise, and complex spatiotemporal patterns. This study presents an efficient location trajectory prediction framework that leverages Generalized Sequential Pattern(GSP) mining to model user activity trajectories and predict subsequent location. GSP generates sequential trajectory patterns that model users' temporal transition activities and predict upcoming locations. The framework has been validated using real-world dataset Foursquare and multiple classification techniques, including K-Nearest-Neighbors, Deep-Learning, and Gradient-Boosted-Trees are used. The experimental results demonstrate that all classifiers achieves high accuracy in next location prediction while maintaining fast execution time, reflecting the effectiveness of the GSP-based feature extraction and preprocessing steps. Comparative results show that KNN outperforms other classifiers, achieving an accuracy of 99.6%,while Deep-Learning and Gradient-Boosted-Trees achieve 98.4% and 99.3%,respectively. Additional performance metrics, including precision, recall, and kappa measure, further confirm the robustness of the proposed approach. Compared with our prior study and other related works using the same dataset, the integration of GSP-based model provides improved accuracy, reduces the model’s size, and lowers execution time. Accuracy improves by 2–7% over our prior work and by 11-34% compared with other related approaches. The GSP-based feature extraction effectively captures sequential dependencies in user mobility,enabling efficient and accurate trajectory prediction for LBSN applications.
Breast cancer is one of the most commonly diagnosed cancers in women, and histopathological examination of biopsy tissue is still the diagnostic gold standard. Deep learning has advanced histopathology classification, but most systems share two costs that are easy to overlook: they need task-specific training, and they are often evaluated with random image-level splits that let images of the same patient land in both training and testing. This paper asks a simple question: how far can a classifier get with no training, using only the embeddings of a frozen pathology foundation model? The proposed method, Retrieval-Augmented Classification (RAC), encodes each image once with a frozen encoder, caches the embeddings in a vector datastore, and classifies a new image by retrieving its nearest labelled neighbours and fusing their labels; nothing is trained. RAC is evaluated on BreakHis under a leakage-free, patient-stratified protocol and contrasted with the optimistic random split, showing how much patient leakage inflates the numbers. With frozen pathology encoders, RAC reaches 98.5% accuracy (AUC 99.8%) on a 20% split and 99.6% on a 10% split, matching a hand-crafted topological retrieval baseline and approaching a fully trained Dual-Attention Vision Transformer. The study further identifies which design choices are beneficial and which are detrimental: enlarging the neighbourhood and combining complementary encoders prove beneficial, whereas stain-normalizing the tiles is detrimental to frozen encoders, reducing accuracy by six to seven points. Frozen pathology foundation models with plain retrieval are thus a strong, interpretable baseline for computer-aided breast cancer diagnosis.
The purpose of this research was to evaluate construct validity of AI generated text detection systems using an examination of a genre segmented corpus of pre-existing academic work that was created prior to the advent of such technology. Using 68 scholarly articles published between 1991 to 2003, the results demonstrated inconsistent misclassifications of genuine human-written content with low general false positive rates for the entire data set. The focused misclassifications of certain genres of academic writing (specifically discussion sections) indicate construct contamination due to stylistic or rhetorical patterns which were irrelevant to the AI generated text source and suggest potential unfairness issues within academic environments. The results were assessed using Messick’s validity framework to show the impact of distribution and genre-based effects upon the measurement accuracy of detection systems, reinforcing the value of calibration and transparent threshold determination in higher education environments where the consequences of assessment decision making are high stakes.
