Eye (Lond). 2025 Jan 9. doi: 10.1038/s41433-024-03585-1. Online ahead of print.

ABSTRACT

PURPOSE: This study aims to develop a deep-learning-based software capable of detecting and differentiating microaneurysms (MAs) as hyporeflective or hyperreflective on structural optical coherence tomography (OCT) images in patients with non-proliferative diabetic retinopathy (NPDR).

METHODS: A retrospective cohort of 249 patients (498 eyes) diagnosed with NPDR was analysed. Structural OCT scans were obtained using the Heidelberg Spectralis HRA + OCT device. Manual segmentation of MAs was performed by five masked readers, with an expert grader ensuring consistent labeling. Two deep learning models, YOLO (You Only Look Once) and DETR (DEtection TRansformer), were trained using the annotated OCT images. Detection and classification performance were evaluated using the area under the receiver operating characteristic (ROC) curves.

RESULTS: The YOLO model performed poorly with an AUC of 0.35 for overall MA detection, with AUCs of 0.33 and 0.24 for hyperreflective and hyporeflective MAs, respectively. The DETR model had an AUC of 0.86 for overall MA detection, but AUCs of 0.71 and 0.84 for hyperreflective and hyporeflective MAs, respectively. Post-hoc review revealed that discrepancies between automated and manual grading were often due to the automated method's selection of normal retinal vessels.

CONCLUSIONS: The choice of deep learning model is critical to achieving accuracy in detecting and classifying MAs in structural OCT images. An automated approach may assist clinicians in the early detection and monitoring of diabetic retinopathy, potentially improving patient outcomes.

PMID:39789187 | DOI:10.1038/s41433-024-03585-1

Sci Rep. 2025 Jan 10;15(1):1554. doi: 10.1038/s41598-025-85248-z.

ABSTRACT

The increasing prevalence of network connections is driving a continuous surge in the requirement for network security and safeguarding against cyberattacks. This has triggered the need to develop and implement intrusion detection systems (IDS), one of the key components of network perimeter aimed at thwarting and alleviating the issues presented by network invaders. Over time, intrusion detection systems have been instrumental in identifying network breaches and deviations. Several researchers have recommended the implementation of machine learning approaches in IDSs to counteract the menace posed by network intruders. Nevertheless, most previously recommended IDSs exhibit a notable false alarm rate. To mitigate this challenge, exploring deep learning methodologies emerges as a viable solution, leveraging their demonstrated efficacy across various domains. Hence, this article proposes an optimized Long Short-Term Memory (LSTM) for identifying anomalies in network traffic. The presented model uses three optimization methods, i.e., Particle Swarm Optimization (PSO), JAYA, and Salp Swarm Algorithm (SSA), to optimize the hyperparameters of LSTM. In this study, NSL KDD, CICIDS, and BoT-IoT datasets are taken into consideration. To evaluate the efficacy of the proposed model, several indicators of performance like Accuracy, Precision, Recall, F-score, True Positive Rate (TPR), False Positive Rate (FPR), and Receiver Operating Characteristic curve (ROC) have been chosen. A comparative analysis of PSO-LSTMIDS, JAYA-LSTMIDS, and SSA-LSTMIDS is conducted. The simulation results demonstrate that SSA-LSTMIDS surpasses all the models examined in this study across all three datasets.

PMID:39789143 | DOI:10.1038/s41598-025-85248-z

Sci Rep. 2025 Jan 9;15(1):1437. doi: 10.1038/s41598-024-84386-0.

ABSTRACT

The most prevalent form of malignant tumors that originate in the brain are known as gliomas. In order to diagnose, treat, and identify risk factors, it is crucial to have precise and resilient segmentation of the tumors, along with an estimation of the patients' overall survival rate. Therefore, we have introduced a deep learning approach that employs a combination of MRI scans to accurately segment brain tumors and predict survival in patients with gliomas. To ensure strong and reliable tumor segmentation, we employ 2D volumetric convolution neural network architectures that utilize a majority rule. This method helps to significantly decrease model bias and improve performance. Additionally, in order to predict survival rates, we extract radiomic features from the tumor regions that have been segmented, and then use a Deep Learning Inspired 3D replicator neural network to identify the most effective features. The model presented in this study was successful in segmenting brain tumors and predicting the outcome of enhancing tumor and real enhancing tumor. The model was evaluated using the BRATS2020 benchmarks dataset, and the obtained results are quite satisfactory and promising.

PMID:39789043 | DOI:10.1038/s41598-024-84386-0

J Craniofac Surg. 2024 Oct 29. doi: 10.1097/SCS.0000000000010663. Online ahead of print.

ABSTRACT

BACKGROUND: With the use of machine learning algorithms, artificial intelligence (AI) has become a viable diagnostic and treatment tool for oral cancer. AI can assess a variety of information, including histopathology slides and intraoral pictures.

AIM: The purpose of this systematic review is to evaluate the efficacy and accuracy of AI technology in the detection and diagnosis of oral cancer between 2020 and 2024.

METHODOLOGY: With an emphasis on AI applications in oral cancer diagnostics, a thorough search approach was used to find pertinent publications published between 2020 and 2024. Using particular keywords associated with AI, oral cancer, and diagnostic imaging, databases such as PubMed, Scopus, and Web of Science were searched. Among the selection criteria were actual English-language research papers that assessed the effectiveness of AI models in diagnosing oral cancer. Three impartial reviewers extracted data, evaluated quality, and compiled the findings using a narrative synthesis technique.

RESULTS: Twelve papers that demonstrated a range of AI applications in the diagnosis of oral cancer satisfied the inclusion criteria. This study showed encouraging results in lesion identification and prognostic prediction using machine learning and deep learning algorithms to evaluate oral pictures and histopathology slides. The results demonstrated how AI-driven technologies might enhance diagnostic precision and enable early intervention in cases of oral cancer.

CONCLUSION: Unprecedented prospects to transform oral cancer diagnosis and detection are provided by artificial intelligence. More resilient AI systems in oral oncology can be achieved by joint research and innovation efforts, even in the face of constraints like data set variability and regulatory concerns.

PMID:39787481 | DOI:10.1097/SCS.0000000000010663

Neuroradiol J. 2025 Jan 9:19714009251313511. doi: 10.1177/19714009251313511. Online ahead of print.

ABSTRACT

INTRODUCTION: The prevalence of neurodegenerative diseases has significantly increased, necessitating a deeper understanding of their symptoms, diagnostic processes, and prevention strategies. Frontotemporal dementia (FTD) and Alzheimer's disease (AD) are two prominent neurodegenerative conditions that present diagnostic challenges due to overlapping symptoms. To address these challenges, experts utilize a range of imaging techniques, including magnetic resonance imaging (MRI), diffusion tensor imaging (DTI), functional MRI (fMRI), positron emission tomography (PET), and single-photon emission computed tomography (SPECT). These techniques facilitate a detailed examination of the manifestations of these diseases. Recent research has demonstrated the potential of artificial intelligence (AI) in automating the diagnostic process, generating significant interest in this field.

MATERIALS AND METHODS: This narrative review aims to compile and analyze articles related to the AI-assisted diagnosis of FTD and AD. We reviewed 31 articles published between 2012 and 2024, with 23 focusing on machine learning techniques and 8 on deep learning techniques. The studies utilized features extracted from both single imaging modalities and multi-modal approaches, and evaluated the performance of various classification models.

RESULTS: Among the machine learning studies, Support Vector Machines (SVM) exhibited the most favorable performance in classifying FTD and AD. In deep learning studies, the ResNet convolutional neural network outperformed other networks.

CONCLUSION: This review highlights the utility of different imaging modalities as diagnostic aids in distinguishing between FTD and AD. However, it emphasizes the importance of incorporating clinical examinations and patient symptom evaluations to ensure comprehensive and accurate diagnoses.

PMID:39787363 | DOI:10.1177/19714009251313511

J Med Chem. 2025 Jan 9. doi: 10.1021/acs.jmedchem.4c02740. Online ahead of print.

ABSTRACT

Applying artificial intelligence techniques to flexibly model the binding between the ligand and protein has attracted extensive interest in recent years, but their applicability remains improved. In this study, we have developed CarsiDock-Flex, a novel two-step flexible docking paradigm that generates binding poses directly from predicted structures. CarsiDock-Flex consists of an equivariant deep learning-based model termed CarsiInduce to refine ESMFold-predicted protein pockets with the induction of specific ligands and our existing CarsiDock algorithm to redock the ligand into the induced binding pockets. Extensive evaluations demonstrate the effectiveness of CarsiInduce, which can successfully guide the transition of ESMFold-predicted pockets into their holo-like conformations for numerous cases, thus leading to the superior docking accuracy of CarsiDock-Flex even on unseen sequences. Overall, our approach offers a novel design for flexible modeling of protein-ligand binding poses, paving the way for a deeper understanding of protein-ligand interactions that account for protein flexibility.

PMID:39787296 | DOI:10.1021/acs.jmedchem.4c02740

PLoS Comput Biol. 2025 Jan 9;21(1):e1012738. doi: 10.1371/journal.pcbi.1012738. eCollection 2025 Jan.

ABSTRACT

Human mobility between different regions is a major factor in large-scale outbreaks of infectious diseases. Deep learning models incorporating infectious disease transmission dynamics for predicting the spread of multi-regional outbreaks due to human mobility have become a hot research topic. In this study, we incorporate the Graph Transformer Neural Network and graph learning mechanisms into a metapopulation SIR model to build a hybrid framework, Metapopulation Graph Transformer Neural Network (M-Graphormer), for high-dimensional parameter estimation and multi-regional epidemic prediction. The framework effectively solves the problem that existing models may lose some hidden spatial dependencies in the data when dealing with the dynamic graph structure of the network due to human mobility. We performed multi-wave infectious disease prediction in multiple regions based on real epidemic data. The results show that the framework is capable of performing high-dimensional parameter estimation and accurately predicting epidemic transmission dynamics in multiple regions even with low data quality. In addition, we retrospectively extrapolate the temporal evolution patterns of contact rate under different interventions implemented in different regions, reflecting the dynamics of intervention intensity and the need for flexibility in adjusting interventions in different regions. To provide early warning of infectious disease transmission, we retrospectively predicted the arrival time of infectious diseases using data from the early stages of outbreaks.

PMID:39787070 | DOI:10.1371/journal.pcbi.1012738

PLoS One. 2025 Jan 9;20(1):e0313776. doi: 10.1371/journal.pone.0313776. eCollection 2025.

ABSTRACT

BACKGROUND: Chat Generative Pre-trained Transformer (ChatGPT) is a 175-billion-parameter natural language processing model that uses deep learning algorithms trained on vast amounts of data to generate human-like texts such as essays. Consequently, it has introduced new challenges and threats to medical education. We assessed the use of ChatGPT and other AI tools among medical students in Uganda.

METHODS: We conducted a descriptive cross-sectional study among medical students at four public universities in Uganda from 1st November 2023 to 20th December 2023. Participants were recruited by stratified random sampling. We used a semi-structured questionnaire to collect data on participants' socio-demographics and use of AI tools such as ChatGPT. Our outcome variable was use of AI tools. Data were analyzed descriptively in Stata version 17.0. We conducted a modified Poisson regression to explore the association between use of AI tools and various exposures.

RESULTS: A total of 564 students participated. Almost all (93%) had heard about AI tools and more than two-thirds (75.7%) had ever used AI tools. Regarding the AI tools used, majority (72.2%) had ever used ChatGPT, followed by SnapChat AI (14.9%), Bing AI (11.5%), and Bard AI (6.9%). Most students use AI tools to complete assignments (55.5%), preparing for tutorials (39.9%), preparing for exams (34.8%) and research writing (24.8%). Students also reported the use of AI tools for nonacademic purposes including emotional support, recreation, and spiritual growth. Older students were 31% less likely to use AI tools compared to younger ones (Adjusted Prevalence Ratio (aPR):0.69; 95% CI: [0.62, 0.76]). Students at Makerere University were 66% more likely to use AI tools compared to students in Gulu University (aPR:1.66; 95% CI:[1.64, 1.69]).

CONCLUSION: The use of ChatGPT and other AI tools was widespread among medical students in Uganda. AI tools were used for both academic and non-academic purposes. Younger students were more likely to use AI tools compared to older students. There is a need to promote AI literacy in institutions to empower older students with essential skills for the digital age. Further, educators should assume students are using AI and adjust their way of teaching and setting exams to suit this new reality. Our research adds further evidence to existing voices calling for regulatory frameworks for AI in medical education.

PMID:39787055 | DOI:10.1371/journal.pone.0313776

Eur Heart J Acute Cardiovasc Care. 2025 Jan 9:zuaf001. doi: 10.1093/ehjacc/zuaf001. Online ahead of print.

ABSTRACT

BACKGROUND: Acute heart failure (AHF) poses significant diagnostic challenges in the emergency room (ER) because of its varied clinical presentation and limitations of traditional diagnostic methods. This study aimed to develop and evaluate a deep-learning model using electrocardiogram (ECG) data to enhance AHF identification in the ER.

METHODS: In this retrospective cohort study, we analyzed the ECG data of 19,285 patients who visited ERs of three hospitals between 2016 and 2020; 9,119 with available left ventricular ejection fraction and N-terminal prohormone of brain natriuretic peptide level data and who were diagnosed with AHF were included in the study. We extracted morphological and clinical parameters from ECG data to train and validate four machine learning models: baseline linear regression and more advanced models including XGBoost, Light GBM, and CatBoost.

RESULTS: The CatBoost algorithm outperformed other models, showing superior area under the receiver operating characteristic and area under the precision-recall curve diagnostic accuracy across both internal (0.89 ± 0.01 and 0.89 ± 0.01) and external (0.90 and 0.89) validation datasets, respectively. The model demonstrated high accuracy, precision, recall, and f1 score, indicating robust performance in AHF identification.

CONCLUSION: The developed machine learning model significantly enhanced AHF detection in the ER using conventional 12-lead ECGs combined with clinical data. These findings suggest that ECGs, a common tool in the ER, can effectively help screen for AHF.

PMID:39787045 | DOI:10.1093/ehjacc/zuaf001

Bioinformatics. 2025 Jan 9:btae762. doi: 10.1093/bioinformatics/btae762. Online ahead of print.

ABSTRACT

MOTIVATION: Transcription factors (TFs) are DNA-binding proteins that regulate gene expression. Traditional methods predict a protein as a TF if the protein contains any DNA-binding domains (DBDs) of known TFs. However, this approach fails to identify a novel TF that does not contain any known DBDs. Recently proposed TF prediction methods do not rely on DBDs. Such methods use features of protein sequences to train a machine learning model, and then use the trained model to predict whether a protein is a TF or not. Because the 3-dimensional (3D) structure of a protein captures more information than its sequence, using 3D protein structures will likely allow for more accurate prediction of novel TFs.

RESULTS: We propose a deep learning-based TF prediction method (StrucTFactor), which is the first method to utilize 3D secondary structural information of proteins. We compare StrucTFactor with recent state-of-the-art TF prediction methods based on ∼525 000 proteins across 12 datasets, capturing different aspects of data bias (including sequence redundancy) possibly influencing a method's performance. We find that StrucTFactor significantly (p-value<0.001) outperforms the existing TF prediction methods, improving the performance over its closest competitor by up to 17% based on Matthews correlation coefficient.

AVAILABILITY: Data and source code are available at https://github.com/lieboldj/StrucTFactor and on our website at https://apps.cosy.bio/StrucTFactor/.

SUPPLEMENTARY INFORMATION: Included.

PMID:39786868 | DOI:10.1093/bioinformatics/btae762

Can J Exp Psychol. 2025 Jan 9. doi: 10.1037/cep0000360. Online ahead of print.

ABSTRACT

Working memory is associated with general intelligence and is crucial for performing complex cognitive tasks. Neuroimaging investigations have recognized that working memory is supported by a distribution of activity in regions across the entire brain. Identification of these regions has come primarily from general linear model analyses of statistical parametric maps to reveal brain regions whose activation is linearly related to working memory task conditions. This approach can fail to detect nonlinear task differences or differences reflected in distributed patterns of activity. In this study, we take advantage of the increased sensitivity of multivariate pattern analysis in a multiple-constraint deep learning classifier to analyze patterns of whole-brain blood oxygen level dependent (BOLD) activity in children performing two different conditions of the emotional n-back task. Regional (supervoxel) whole-brain activation patterns from functional imaging runs of 20 children were used to train a set of neural network classifiers to identify task category (0-back vs. 2-back) and activation co-occurrence probability, which encoded functional connectivity. These simultaneous constraints promote the discovery of coherent networks that contribute towards task performance in each memory load condition. Permutation analyses discovered the global activation patterns and interregional coactivations that distinguish memory load. Examination of model weights identified the brain regions most predictive of memory load and the functional networks integrating these regions. Community detection analyses identified functional networks integrating task-predictive regions and found distinct patterns of network activation for each task type. Comparisons to functional network literature suggest more focused attentional network activation during the 2-back task. (PsycInfo Database Record (c) 2025 APA, all rights reserved).

PMID:39786863 | DOI:10.1037/cep0000360

Bioinformatics. 2025 Jan 9:btaf010. doi: 10.1093/bioinformatics/btaf010. Online ahead of print.

ABSTRACT

MOTIVATION: Accurate prediction of protein side-chain conformations is necessary to understand protein folding, protein-protein interactions and facilitate de novo protein design.

RESULTS: Here we apply torsional flow matching and equivariant graph attention to develop FlowPacker, a fast and performant model to predict protein side-chain conformations conditioned on the protein sequence and backbone. We show that FlowPacker outperforms previous state-of-the-art baselines across most metrics with improved runtime. We further show that FlowPacker can be used to inpaint missing side-chain coordinates and also for multimeric targets, and exhibits strong performance on a test set of antibody-antigen complexes.

AVAILABILITY: Code is available at https://gitlab.com/mjslee0921/flowpacker.

SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

PMID:39786861 | DOI:10.1093/bioinformatics/btaf010

Int J Cardiovasc Imaging. 2025 Jan 9. doi: 10.1007/s10554-025-03322-z. Online ahead of print.

ABSTRACT

Amid an aging global population, heart failure has become a leading cause of hospitalization among older people. Its high prevalence and mortality rates underscore the importance of accurate mortality prediction for swift disease progression assessment and better patient outcomes. The evolution of artificial intelligence (AI) presents new avenues for predicting heart failure mortality. Yet current research has predominantly leveraged structured data and unstructured clinical notes from electronic health records (EHR), underutilizing the prognostic value of chest X-rays (CXRs). This study aims to harness deep learning methodologies to explore the feasibility of enhancing the precision of predicting in-hospital all-cause mortality in heart failure patients using CXRs data. We propose a novel multimodal deep learning network based on the spatially and temporally decoupled Transformer (MN-STDT) for in-hospital mortality prediction in heart failure by integrating longitudinal CXRs and structured EHR data. The MN-STDT captures spatial and temporal information from CXRs through a Hybrid Spatial Encoder and a Distance-Aware Temporal Encoder, ultimately fusing features from both modalities for predictive modeling. Initial pre-training of the spatial encoder was conducted on CheXpert, followed by full model training and evaluation on the MIMIC-IV and MIMIC-CXR datasets for mortality prediction tasks. In a comprehensive view, the MN-STDT demonstrated the best performance, with an AUC-ROC of 0.8620, surpassing all baseline models. Comparative analysis revealed that the AUC-ROC of the multimodal model (0.8620) was significantly higher than that of models using only structured data (0.8166) or chest X-ray data alone (0.7479). This study demonstrates the value of CXRs in the prognosis of heart failure, showing that the combination of longitudinal CXRs with structured EHR data can significantly improve the accuracy of mortality prediction in heart failure. Feature importance analysis based on SHAP provides interpretable decision support, paving the way for potential clinical applications.

PMID:39786626 | DOI:10.1007/s10554-025-03322-z

Breast Cancer Res Treat. 2025 Jan 9. doi: 10.1007/s10549-024-07589-z. Online ahead of print.

ABSTRACT

PURPOSE: Traditional computer-assisted detection (CADe) algorithms were developed for 2D mammography, while modern artificial intelligence (AI) algorithms can be applied to 2D mammography and/or digital breast tomosynthesis (DBT). The objective is to compare the performance of a traditional machine learning CADe algorithm for synthetic 2D mammography to a deep learning-based AI algorithm for DBT on the same mammograms.

METHODS: Mammographic examinations from 764 patients (mean age 58 years ± 11) with 106 biopsy-proven cancers and 658 cancer-negative cases were analyzed by a CADe algorithm (ImageChecker v10.0, Hologic, Inc.) and an AI algorithm (Genius AI Detection v2.0, Hologic, Inc.). Synthetic 2D images were used for CADe analysis, and DBT images were used for AI analysis. For each algorithm, an overall case score was defined as the highest score of all lesion marks, which was used to determine the area under the receiver operating characteristic curve (AUC).

RESULTS: The overall AUC was higher for 3D AI than 2D CADe (0.873 versus 0.693, P < 0.001). Lesion-specific sensitivity of 3D AI was higher than 2D CADe (94.3 versus 72.6%, P = 0.002). Specificity of 3D AI was higher than 2D CADe (54.3 versus 16.7%, P < 0.001), and the rate of false marks on non-cancer cases was lower for 3D AI than 2D CADe (0.91 versus 3.24 per exam, P < 0.001).

CONCLUSION: A deep learning-based AI algorithm applied to DBT images significantly outperformed a traditional machine learning CADe algorithm applied to synthetic 2D mammographic images, with regard to AUC, sensitivity, and specificity.

PMID:39786500 | DOI:10.1007/s10549-024-07589-z

Drug Saf. 2025 Jan 9. doi: 10.1007/s40264-024-01505-6. Online ahead of print.

ABSTRACT

BACKGROUND: Natural language processing (NLP) and machine learning (ML) techniques may help harness unstructured free-text electronic health record (EHR) data to detect adverse drug events (ADEs) and thus improve pharmacovigilance. However, evidence of their real-world effectiveness remains unclear.

OBJECTIVE: To summarise the evidence on the effectiveness of NLP/ML in detecting ADEs from unstructured EHR data and ultimately improve pharmacovigilance in comparison to other data sources.

METHODS: A scoping review was conducted by searching six databases in July 2023. Studies leveraging NLP/ML to identify ADEs from EHR were included. Titles/abstracts were screened by two independent researchers as were full-text articles. Data extraction was conducted by one researcher and checked by another. A narrative synthesis summarises the research techniques, ADEs analysed, model performance and pharmacovigilance impacts.

RESULTS: Seven studies met the inclusion criteria covering a wide range of ADEs and medications. The utilisation of rule-based NLP, statistical models, and deep learning approaches was observed. Natural language processing/ML techniques with unstructured data improved the detection of under-reported adverse events and safety signals. However, substantial variability was noted in the techniques and evaluation methods employed across the different studies and limitations exist in integrating the findings into practice.

CONCLUSIONS: Natural language processing (NLP) and machine learning (ML) have promising possibilities in extracting valuable insights with regard to pharmacovigilance from unstructured EHR data. These approaches have demonstrated proficiency in identifying specific adverse events and uncovering previously unknown safety signals that would not have been apparent through structured data alone. Nevertheless, challenges such as the absence of standardised methodologies and validation criteria obstruct the widespread adoption of NLP/ML for pharmacovigilance leveraging of unstructured EHR data.

PMID:39786481 | DOI:10.1007/s40264-024-01505-6

J Chem Inf Model. 2025 Jan 9. doi: 10.1021/acs.jcim.4c01732. Online ahead of print.

ABSTRACT

Proteolysis-targeting chimeras (PROTACs) are heterobifunctional molecules that target undruggable proteins, enhance selectivity and prevent target accumulation through catalytic activity. The unique structure of PROTACs presents challenges in structural identification and drug design. Liquid chromatography (LC), combined with mass spectrometry (MS), enhances compound annotation by providing essential retention time (RT) data, especially when MS alone is insufficient. However, predicting RT for PROTACs remains challenging. To address this, we compiled the PROTAC-RT data set from literature and evaluated the performance of four machine learning algorithms─extreme gradient boosting (XGBoost), random forest (RF), K-nearest neighbor (KNN) and support vector machines (SVM)─and a deep learning model, fully connected neural network (FCNN), using 24 molecular fingerprints and descriptors. Through screening combinations of molecular fingerprints, descriptors and chromatographic condition descriptors (CCs), we developed an optimized XGBoost model (XGBoost + moe206+Path + Charge + CCs) that achieved an R2 of 0.958 ± 0.027 and an RMSE of 0.934 ± 0.412. After hyperparameter tuning, the model's R2 improved to 0.963 ± 0.023, with an RMSE of 0.896 ± 0.374. The model showed strong predictive accuracy under new chromatographic separation conditions and was validated using six experimentally determined compounds. SHapley Additive exPlanations (SHAP) not only highlights the advantages of XGBoost but also emphasizes the importance of CCs and molecular features, such as bond variability, van der Waals surface area, and atomic charge states. The optimized XGBoost model combines moe206, path, charge descriptors, and CCs, providing a fast and precise method for predicting the RT of PROTACs compounds, thus facilitating their annotation.

PMID:39786356 | DOI:10.1021/acs.jcim.4c01732

Alzheimers Dement. 2024 Dec;20 Suppl 2:e087666. doi: 10.1002/alz.087666.

ABSTRACT

BACKGROUND: Drugs targeting Alzheimer's disease (AD) pathology are likely to be most effective in the presymptomatic stage, where individuals harbor AD pathology but have not manifested symptoms. Neuroimaging approaches can help to identify such individuals, but are costly for population-wide screening. Cost-effective screening is needed to identify those who may benefit from neuroimaging, such as those at risk of developing clinical disease. We present a deep learning algorithm that uses accelerometry recordings to predict clinically diagnosed AD in dementia-free patients.

METHOD: Participants were from The Memory and Aging Project (MAP), a longitudinal cohort study of older adults focused on aging and dementia. As part of this study, participants were asked to wear a wrist accelerometer for ten days. We designed a feedforward neural network that synthesizes clinical and accelerometric features to predict clinical AD. Clinical features were selected based on ease of collection, and included age, sex, education, social isolation and purpose in life. The dataset included participants without dementia at the time of recording, who survived for and had a known clinical AD outcome within five years of the recording.

RESULT: The training dataset consisted of 875 unique patients. The test dataset consisted of 395 unique patients (mean age 81.5, SD= 6.8). 14.4% of the test set developed clinical AD within five years. On the test set, the model achieved 89% sensitivity (SN), 70% specificity (SP), and an F1 score of 0.87 (Figure 1). This is superior to accelerometric-only (SN 67%, SP 70%, F1 0.68) and clinical-only (SN 76%, SP 80%, F1 0.78) models. Model accuracy was similar for patients both with and without mild cognitive impairment at baseline. When the binarized model output was used as a predictor for five-year AD-free survival via Cox proportional hazards (Figure 2), it achieved a C-index of 0.729 [95% CI 0.69 b-0.77], a hazard ratio of 6.90 [95% CI 4.15-11.47], and a log-likelihood ratio of 71.05. Further tuning and validation of the model are underway.

CONCLUSION: A deep learning model using wrist accelerometry and easily obtained clinical features shows promise in predicting 5-year AD outcomes in adults without dementia at baseline.

PMID:39786306 | DOI:10.1002/alz.087666

Alzheimers Dement. 2024 Dec;20 Suppl 2:e086725. doi: 10.1002/alz.086725.

ABSTRACT

BACKGROUND: The location of proposed brain MRI markers of small vessel disease (SVD) might reflect their pathogenesis and may translate into differential associations with cognition. We derived regional MRI markers of SVD and studied: (i) associations with cognitive performance, (ii) patterns most likely to reflect underlying SVD, (iii) mediating effects on the relationships of age and cardiovascular disease (CVD) risk with cognition.

METHOD: In 891 participants from The Multi-Ethnic Study of Atherosclerosis, we segmented enlarged perivascular spaces (ePVS), white matter hyperintensities (WMH) and microbleeds (MBs) using deep learning-based algorithms, and calculated white matter (WM) microstructural integrity measures of fractional anisotropy (FA), trace (TR) and free water (FW) using automated DTI-processing pipelines. Measures of global and domain-specific cognitive performance were derived from a comprehensive cognitive evaluation based on the UDS v3 neuropsychological battery.

RESULT: Mean (SD) age was 73.6 (7.9) years; 474 (53%) participants were women. In generalized linear models adjusted for demographics, vascular risk factors, and APOE ε4 carriership, higher basal ganglia ePVS count was associated with worse global, language, and attention cognitive performance (Table 1). Higher periventricular WMH volume was associated with worse global, delayed memory, language, phonemic, and attention performance. Higher WM FA was associated with better global, delayed memory, language, and attention performance. Higher WM TR was associated with worse global, delayed memory, language, phonemic, and attention performance. Exploratory factor analysis revealed that basal ganglia ePVS (standardized loading=0.51), thalamus ePVS (0.43), periventricular WMH (0.85), subcortical WMH (0.65), and WM FA (-0.73) and TR (0.84) loaded onto the same factor, likely reflecting underlying SVD. Structural equation models demonstrated that SVD mediated the effect of age on cognition (β[95%CI]= -0.071[-0.088,-0.053]) through the pathways: Age→SVD→Cognition (-0.044[-0.063,-0.026]) and Age→SVD→Brain Atrophy→Cognition (-0.006[-0.012,-0.002]) - Figure 1, and the effect of CVD risk on cognition (-0.028[-0.044,-0.012]) through the pathways: CVD Risk→SVD→Cognition (-0.021[-0.031,-0.013]) and CVD Risk→SVD→Brain Atrophy→Cognition (-0.007[-0.012,-0.003]) - Figure 2.

CONCLUSION: The location of the proposed MRI markers of SVD likely reflects distinct etiopathogenic substrates and should be considered when examining associations with cognitive or other health-related outcomes. SVD mediates the relationships of age and CVD risk with cognition via both atrophy-related and unrelated pathways.

PMID:39786253 | DOI:10.1002/alz.086725

Alzheimers Dement. 2024 Dec;20 Suppl 2:e090583. doi: 10.1002/alz.090583.

ABSTRACT

BACKGROUND: Neurodegenerative diseases are a heterogeneous group of illnesses. Differences across patients exist in the underlying biological drivers of disease. Furthermore, cross-diagnostic disease mechanisms exist, and different pathologies often co-occur in the brain. Clinical symptoms fail to capture this heterogeneity. Molecular biomarker-driven approaches are needed to improve patient identification & stratification with the aim of moving towards more targeted patient treatment strategies.

METHOD: The UK Biobank Pharma Proteomics Project has generated a proteomics dataset of unprecedented size. It consists of plasma proteomic profiles measured using the Olink 3k protein panel from over 54,000 individuals, complemented with broad phenotypic & genetic information. It is a unique resource to explore the biological correlation between neurodegenerative diseases as well as with other indications. We applied a multi-task deep learning (MTL) approach to generate models that predict disease status based on plasma proteomics profiles for a broad range of indications, including neurodegenerative diseases such as Alzheimer's and Parkinson's disease. The MTL approach first learns fundamental biological patterns by knowledge sharing between diseases in the initial segment of the network. Subsequently, understanding of a specific indication is refined with dedicated training. This improves model generalizability and statistical power. The neural network also produces low-dimensional embeddings of proteomic profiles that can be used for sample clustering and to derive insights about disease-associated processes.

RESULT: The average precision-recall area-under-the-curve (PR-AUC) of the MTL models across all diseases is 0.72 vs. 0.67 for the baseline single task logistic regression models. For Alzheimer's disease, the MTL classifier has a PR-AUC of 0.76. Feature importance scores were calculated using the SHAP method. Top features for Alzheimer's disease included several known biomarkers (e.g., GFAP, NPTXR). A UMAP projection of all diseases using the feature importance scores clusters diseases by disease category. Sample clustering revealed biologically interpretable patient subgroups, such as a Parkinson's cluster linked to lysosomal biology.

CONCLUSION: High performance of the MTL approach signifies good characterization of cross-disease biology. This is corroborated by the model's capability to produce meaningful low-dimensional representations of plasma proteomics profiles that can be used for identification of cross-diagnostic protein signatures and subtypes of neurodegenerative diseases. UKB application number 65851.

PMID:39786064 | DOI:10.1002/alz.090583

Alzheimers Dement. 2024 Dec;20 Suppl 6:e086486. doi: 10.1002/alz.086486.

ABSTRACT

BACKGROUND: Pivotal Alzheimer's Disease (AD) trials typically require thousands of participants, resulting in long enrollment timelines and substantial costs. We leverage deep learning predictive models to create prognostic scores (forecasted control outcome) of trial participants and in combination with a linear statistical model to increase statistical power in randomized clinical trials (RCT). This is a straightforward extension of the traditional RCT analysis, allowing for ease of use in any clinical program. We demonstrate the application of these methods retrospectively on 3 pivotal Phase III clinical trials in mild-to-moderate AD (NCT00236431, NCT00574132, and NCT00575055).

METHOD: A probabilistic deep learning model was trained on the trajectories of nearly 7000 participants who had varying degrees of cognitive impairment, ranging from mild cognitive impairment (MCI) to moderate AD. These trajectories were collected observational studies and the control arms of RCTs. This trained model was used to forecast the control outcomes of participants in the three trials retrospectively, by entering their individual trial baseline data. The resultant forecasts are known as prognostic scores and represent comprehensive predictions across a broad range of AD outcomes. We evaluated the potential reduction in estimated variance and how this could translate to required sample size by incorporating the prognostic score as a covariate in the primary linear statistical model of each study, analyzing the 11-component Alzheimer's Disease Assessment Scale-Cognitive Subscale (ADAS-Cog11) and the Clinical Dementia Rating Sum-of-Boxes (CDR-SB) endpoints as applicable.

RESULT: Prognostic scores have the potential to decrease estimated variance between 5% to 10% and placebo arm sample size between 7% and 17% in the 3 studies when comparing standard + prognostic score vs. standard adjustment.

CONCLUSION: Prognostic scores have the potential to increase the statistical power in clinical trials; this would enable a reduced number of subjects required to detect a significant treatment effect. Potential sample size reduction during trial planning must be carefully estimated using independent validation studies to reduce the risk of under-powering the trial.

PMID:39782540 | DOI:10.1002/alz.086486