De Lin Institute of Technology

Abstract Gene coexpression networks (GCNs) provide a useful representation of coordinated gene activity, but their reconstruction from single-cell RNA sequencing (scRNA-seq) data is challenged by high dimensionality, sparsity and heterogeneous noise. Most existing approaches rely on marginal correlation or machine-learning regressors, which tend to capture indirect associations and typically lack rigorous statistical error control.We propose dGGAPM (Direct Gene--Gene Associations via Precision Matrix), a precision-matrix-based framework for inferring direct gene--gene associations from scRNA-seq data. dGGAPM employs node-wise sparse regression to estimate partial correlations between genes, and then applies an adaptive thresholding rule derived from high-dimensional asymptotic theory to obtain a sparse network under a pre-specified false discovery rate (FDR). This yields a signed gene coexpression network in which edges correspond to statistically supported conditional dependencies, providing a principled alternative to ad hoc correlation cutoffs.We evaluate dGGAPM on three time-resolved scRNA-seq datasets covering human and mouse embryonic stem cells and mouse hematopoietic stem cells, using ChIP-seq–derived regulatory networks as proxies for ground truth. Across these benchmarks, dGGAPM achieves competitive or improved area under the ROC curve compared with several widely used methods (scLink, GENIE3, SCODE and GRNBoost2), and recovers biologically coherent modules enriched for processes such as lineage commitment, DNA replication and hematopoiesis. Case studies on key regulators including NANOG, SOX2 and GATA2 further illustrate that dGGAPM can reveal interpretable modules and hub genes. These results indicate that precision-matrix-based inference with explicit FDR control is a useful and robust strategy for reconstructing gene coexpression networks from single-cell transcriptomic data.

Purpose This paper aims to evaluate how effectively large language models (LLMs) represent and generate minority cultural knowledge, specifically Taiwanese Hakka culture. To address this, the study proposes a structured and replicable evaluation framework integrating Bloom’s taxonomy and retrieval-augmented generation (RAG). The research is guided by the following questions: (1) How do LLMs perform across different cognitive domains when processing Hakka cultural content? (2) To what extent does the integration of RAG enhance the accuracy and contextual appropriateness of LLM outputs? And (3) How do different LLM architectures compare in their ability to recall, analyse and creatively synthesize culturally grounded information? Design/methodology/approach This study proposes a cognitive benchmarking framework to evaluate how LLMs process and apply culturally specific knowledge. The framework integrates Bloom’s taxonomy with RAG to assess model performance across six hierarchical cognitive domains: remembering, understanding, applying, analysing, evaluating and creating. Using a curated Taiwanese Hakka digital cultural archive as the primary testbed, the evaluation measures LLM-generated responses’ semantic accuracy and cultural relevance. Findings The evaluation results indicate that LLMs augmented with RAG exhibit marked improvements over baseline models in the cognitive domains of remembering, understanding and analysing. These enhancements are particularly evident in tasks requiring factual accuracy, contextual relevance and semantic precision, underscoring RAG’s effectiveness in addressing the knowledge sparsity typically observed in underrepresented cultural data sets. However, a notable limitation persists across all models including those equipped with RAG in the domain of creating. This suggests that while retrieval mechanisms bolster the reproduction and comprehension of cultural knowledge, they do not yet sufficiently support culturally nuanced generative synthesis. Originality/value This study introduces a novel evaluation framework integrating cognitive domain benchmarks with RAG-enhanced LLMs to assess cultural knowledge processing. The research advances culturally grounded artificial intelligence (AI) systems and digital archival quality by empirically demonstrating RAG’s impact on improving factual accuracy in lower and mid-level tasks. The findings affirm the strategic value of retrieval integration for enhancing representational fidelity in cultural AI applications, while also highlighting the need for future research into hybrid architectures that combine external grounding with culturally adaptive generation strategies.

Most studies in bodily self-consciousness were limited to the case of a single body. We performed VR experiments to test the hypothesis that it is possible for healthy subjects to experience the Double Body Effect-the experiential combination of double body ownership and double body-location. Under proprioceptive-vestibular and visual manipulations, participants wobbled involuntarily while watching two identical avatars doing exactly the same. The results showed that, in both the 1PP and 3PP conditions, it was indeed possible for healthy subjects to experience the Double Body Effect. This reveals that body ownership and body location are more flexible than most studies have considered so far. It also suggests that self-location and body-location are not the same experiences, and that the relation between self and body is more complicated than both traditional dualism and contemporary reductionism. Finally, our findings can serve as a preliminary model for understanding the perplexing phenomenology of heautoscopy.

We present an end-to-end workflow for privacyenhancing synthetic dataset generation and privacy-risk auditing by fine-tuning and prompting the open source Llama-3.1-8B with differential-privacy (DP) techniques. Two corpora are produced according to our workflow: a Raw-Generated Dataset without protection and a DP-Semantic Dataset created by adding DP noise controlled by a privacy budget. Using Anonymeter, aligned with EU Article 29 WP guidelines, we quantify singlingout, linkability, and inference risk. Experiments show high reidentification risk in the raw corpus, whereas the DP corpus, under a moderate budget, drastically lowers all three metrics while preserving data utility. Our reproducible, open framework offers a regulation-compliant recipe for safe data synthesis, augmentation, and city-scale applications of large language models.

A 450 nm blue-light laser diode underwater wireless optical communication (UWOC) system with 155 Mbps, 622 Mbps, and 1.25 Gbps is tested in a various water density. In a 1.03 g/cm<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">3</sup> density, the UWOC communication can transmit information up to 3 meters. The optical power attenuation is increased in higher seawater density and longer distances. The seawater density contributes to the UWOC system absorption and scattering, which degrades the signal. Different modulation speeds can also increase the BER value in different water densities. Moreover, the UWOC system was tested with chlorophyll-containing water at different turbidity levels. Adding more chlorophyll to the water increased light absorption, which caused the deterioration of the UOWC results.

Existing drug recommendation systems lack integration with up-to-date clinical guidelines (the latest diabetes association standards of care and clinical guidelines that align with local government health care regulations) and lack high-precision drug interaction processing, explainability, and dynamic dosage adjustment. As a result, the recommendations generated by these systems are often inaccurate and do not align with local standards, greatly limiting their practicality.To develop a personalized drug recommendation and dosage optimization system named Diabetes Drug Recommendation System (DDRs), integrating Fast Healthcare Interoperability Resources-standardized electronic health record (EHR) data and up-to-date clinical guidelines for accurate and practical recommendations.We analyzed patients' EHR and International Classification of Diseases-tenth edition codes and integrated them with a drug interaction database to reduce adverse reactions. ADA guidelines and Taiwan's National Health Insurance (NHI) chronic disease guidelines served as data sources. Bio-GPT and Retrieval-Augmented Generation (RAG) were used to build the clinical guideline database and ensure recommendations align with the latest standards, with references provided for interpretability. Finally, optimal dosage was dynamically calculated by integrating patient disease progression trends from the EHR.DDRs achieved superior drug recommendation accuracy (Precision-Recall Area Under the Curve = 0.7951, Jaccard = 0.5632, F1-score = 0.7158), with a low drug-drug interaction rate (4.73%) and dosage error (±6.21%). Faithfulness of recommendations reached 0.850. Field validation with three physicians showed that the system reduced literature review time by 30 to 40% and delivered clinically actionable recommendations.DDRs is the first system to integrate EHR data, LLMs, RAG, ADA guidelines, and Taiwan NHI policies for diabetes treatment. The system demonstrates high accuracy, safety, and interpretability, offering practical decision support in routine clinical settings.

Abstract This study introduces an innovative recurrent neural network called long short‐term memory (LSTM) as a prediction model, which is used to predict angle‐dependent reflectance distributions of colour samples with glossy and matte surfaces. A two‐dimensional (2D) reflectance measurement system was developed to measure the angle‐dependent reflectance in this study. Its structure mainly included a semicircular rotating mechanism, a high‐resolution digital camera and a high‐quality white light‐emitting diode. A semicircular rotating mechanism was designed to rotate from 10° to 170° in the vertical direction. Two ColorGauge miniaturised colour charts with glossy and matte surfaces were selected as test chips. The test chips on ColorGauge miniaturised colour charts included fives colours of glossy white, glossy black, matte red, matte green and matte blue. The reflectance distributions of the test chips were measured by the 2D reflectance measurement system, and the measured reflectance data were used as training data in the LSTM model. In comparison with second‐ and third‐order regressions, the mean CIE lightness difference (0.09) using the LSTM model was lower. Therefore, it was verified that the LSTM model performed well in predicting reflectance distributions. In addition, the LSTM model was also validated on the additional test samples (10 matte chromatic samples and five glossy achromatic samples). The maximum and minimum mean CIE lightness differences were 3.77 and 0.64 for matte chromatic samples, and 2.34 and 0.42 for glossy achromatic samples, respectively. The results of small prediction errors indicated that the LSTM model presents excellent prediction performance.