Artificial Intelligence (AI) models for gene expression changes have emerged as vital tools in biomedical research, mapping transcriptomic responses to sequence variations and molecular perturbations to accelerate biomarker discovery, drug discovery and precision medicine. Gene expression change prediction algorithms have become a central focus in biomedical research, leveraging machine learning and statistical modeling to forecast transcriptional responses under various conditions. However, existing literature lacks a unified structural framework, often conflating core execution models with downstream phenotypic applications. To address this gap, in this paper we present a systematic, reproducible survey and critical synthesis of state-of-the-art AI methodologies in gene expression change prediction from 2013 to 2025. Using a transparent literature screening protocol across four major repositories, we introduce a novel, mutually exclusive Bidirectional Hierarchical Taxonomy that cleanly segregates architectures by their upstream driver and their modeling landscape. Moving beyond purely descriptive summaries, our core contribution provides a rigorous evaluation of model performance across standard statistical metrics, noting that sequence-to-expression models achieve peak Pearson (r) and Spearman ( ρ ) correlations between 0.45 and 0.82 depending on Genomic context, while chemical and genetic perturbation engines are evaluated via Mean Squared Error (MSE), F1-scores, and Area Under the Receiver Operating Characteristic (ROC-AUC) curves exceeding 0.85 in in-distribution settings. Finally, we distill critical open challenges specifically data sparsity, cell-line generalization bias, and out-of-distribution transfer limits to establish an actionable Roadmap for future computational framework design.
Lung cancer remains the leading cause of cancer-related mortality worldwide, and accurate automated classification of pulmonary lesions from computed tomography images represents a clinically significant challenge. This study presents a comprehensive multi-architecture evaluation framework for automated lung cancer detection, systematically assessing the contributions of structured preprocessing, adversarial augmentation, and architectural design to classification performance and generalization stability. Seven deep learning architectures were evaluated across two geographically distinct CT imaging datasets comprising 262 subjects from Iranian and Iraqi institutional cohorts under subject-level stratified 10-fold cross-validation. Evaluated configurations encompassed a purpose-built convolutional neural network, three hybrid pipelines coupling CNN-extracted features with classical classifiers, and three fine-tuned pretrained architectures. A multi-stage preprocessing pipeline and Fast Gradient Sign Method adversarial augmentation were systematically applied to isolate their individual and combined contributions. Gradient-weighted Class Activation Mapping was employed to assess spatial interpretability, and non-parametric statistical testing was conducted to establish the significance of observed performance differences. The standalone CNN under the combined preprocessing and adversarial augmentation configuration achieved the highest accuracy of 95.80
Sepsis is a life-threatening condition with a case-fatality rate exceeding 25 0.851 ± 0.017 , an AUPRC of 0.563 ± 0.045 , and a specificity of 0.964 ± 0.013 . The model outperforms five established baselines, including AdaCare, ConCare and Clinical-BERT. Calibration analysis indicates that predicted probabilities are well aligned with observed outcomes. Ablation studies confirm that each architectural component contributes to performance improvements, while attention analysis shows weight concentrated on features consistent with established sepsis biomarkers, including lactate, heart rate, and respiratory rate. The proposed ETSEW framework demonstrates that modelling ICU data as asynchronous event streams can significantly improve early sepsis prediction. By addressing irregular sampling, class imbalance, and uncertainty estimation, the model provides both accurate and reliable predictions. These characteristics make ETSEW a promising approach for real-world clinical decision support systems.
Human longevity is a complex phenotype influenced by numerous genetic variants whose reported associations vary across studies and populations. Integrating curated human genetic data with experimental evidence and post-transcriptional regulatory information may help to identify longevity-associated genes and the biological processes they share. We performed a source-reviewed historical reanalysis of LongevityMap Build 3. Candidate genes were required to have at least two significant study units, evidence from at least two reported population labels and a significant-result proportion of at least 0.60, and the robustness of selection was assessed across 100 threshold combinations and alternative definitions of the selection criteria. Gene Ontology Biological Process enrichment was evaluated against the LongevityMap gene background, with leave-region-out sensitivity analyses. Candidates were integrated with six Open Genes datasets after removal of human-association records duplicating LongevityMap publications, and additional exploratory candidates were identified. Experimentally validated miRNA–target interactions were analysed for the primary and integrated gene sets, with assay-restricted and coverage-matched null comparisons. Of 550 LongevityMap records, 547 were analysable and represented 270 publications. After source review, 213 genes had interpretable single-gene evidence and 21 met the selection criteria. These 21 genes correspond to 17 genomic regions, because APOE, TOMM40 and APOC1 lie in the 19q13.32 region and HSPA1A, HSPA1L and HLA-DRB1 in the MHC region; 13 of the 21 genes, including PON1, KL, TOMM40 and APOE, were selected under all alternative criteria. No Gene Ontology term remained significant after Benjamini–Hochberg correction. The strongest nominal signals concerned cholesterol and lipid transport, but they depended on a small group of apolipoprotein-related genes. Of the 1,460 Open Genes human-longevity records, 438 duplicated LongevityMap publications and were excluded. After adjustment for research coverage, the prioritized genes were over-represented among genes with non-overlapping human longevity evidence (adjusted odds ratio 6.07; FDR = 0.011), but not among genes with functional or experimental evidence. Fifty-seven additional exploratory candidates were identified, giving an integrated set of 78 genes. hsa-miR-34a-5p ranked first in both miRNA analyses, targeting 18 of the 21 primary genes and 69 of the 78 integrated genes. However, its target coverage did not exceed that of random gene sets with comparable research coverage (FDR = 0.58 and 0.95), and no miRNA showed significant excess coverage. When only reporter, western blot, qPCR or ELISA evidence was considered, hsa-miR-17-5p ranked first. This study provides a conservative, source-reviewed framework for prioritizing human longevity-associated genes and integrating them with complementary ageing evidence and validated miRNA–target relationships. The prioritized genes are consistently supported by external human evidence, whereas the miRNA rankings mainly reflect the extent to which genes and miRNAs have been studied. The identified genes and miRNAs should be regarded as exploratory candidates rather than confirmed causal determinants, biomarkers or therapeutic targets of human longevity.
The exponential growth of epidemiological data from distributed healthcare systems presents unprecedented opportunities and challenges for communicable disease surveillance. This paper presents a comprehensive cloud-based analytics framework leveraging Apache Spark for distributed processing of 1.8 million epidemiological records to enable real-time communicable disease prediction using sparse clinical inputs. We propose a privacy-preserving architecture that integrates local language models via Ollama at the edge layer for initial data structuring, ensuring patient privacy and data sovereignty before cloud transmission. The core predictive component employs a Random Forest Classifier trained on 86 clinical features spanning demographic, physiological, hematological, immunological, and symptomatic domains. Our framework achieves 89.0
