Alzheimers Dement. 2024 Dec;20 Suppl 1:e085804. doi: 10.1002/alz.085804.

ABSTRACT

BACKGROUND: Cerebral small vessel disease (CSVD), which includes cerebral amyloid angiopathy (CAA) and arteriolosclerosis, often co-occurs with Alzheimer's disease (AD) pathology. The medial temporal lobe (MTL) is susceptible to hosting multiple AD pathologies, such as neurofibrillary tangles (NFTs), amyloid-β plaques, phospho-Tar-DNA-Binding-Protein-43 (pTDP-43), as well as CSVD. Whether a causal relationship between these pathologies exists remains largely unknown, but one potential linking mechanism is the dysfunction of perivascular clearance. Our objective was to examine the burden of CSVD in the MTL of a pathological AD cohort and to establish the associations between CSVD and AD-related pathologies, as well as between CSVD and enlarged perivascular spaces (EPVS), a potential indicator of clearance dysfunction.

METHOD: The study included 156 autopsy cases (mean age at death 79.4±10.9 years, 90 females) from the Massachusetts Alzheimer's Disease Research Center (MADRC). One hemisphere was preserved in formalin, and 5 µm-thick sections were cut from predefined regions of the hippocampal body and entorhinal cortex. These sections were subsequently stained using luxol fast blue with hematoxylin&eosin (LHE), and antibodies against amyloid-β, hyperphosphorylated tau (At8), and pTDP-43, following standard histological and immunohistochemical protocols. Utilizing deep-learning models (Aiforia®), we computed the burden of CAA, amyloid-β plaques, NFTs, and pTDP-43 inclusions (Figure 1). Additionally, the severity of arteriolosclerosis and the % area of EPVS were evaluated on the LHE sections.

RESULT: In linear mixed effects models CAA was positively associated with the density of NFTs (Est = 7.21; p = 0.024; R2 = 69%) and amyloid-β plaque burden (Est = 3.01; p<0.001; R2 = 61%) in all regions of interest. Arteriolosclerosis had no direct effect on parenchymal AD-related pathologies but showed a positive interaction with CAA in the association with PVS enlargement. There was no relationship between pTDP-43 inclusions and arteriolosclerosis.

CONCLUSION: These results point towards an association between microvascular pathology and AD-related pathology, possibly mediated by clearance dysfunction.

PMID:39751386 | DOI:10.1002/alz.085804

Alzheimers Dement. 2024 Dec;20 Suppl 1:e092140. doi: 10.1002/alz.092140.

ABSTRACT

BACKGROUND: Assessing tau accumulation in early affected areas like the lateral entorhinal cortex (EC) and inferior temporal gyrus (ITG) enables early prediction of disease progression and cognitive decline. However, positron emission tomography (PET) imaging poses radiation exposure and cost concerns. This research aims to develop a deep learning model predicting tau positivity in these regions using MRI.

METHOD: In this study, we used the Alzheimer's Disease Neuroimaging Initiative (ADNI) cohort, of which dataset was partitioned into train, validation, and test sets (8:1:1 ratio), encompassing a total of 1010 scans, all of whom underwent T1-weighted magnetic resonance imaging (MRI) and [18F] flortaucipir-PET imaging. For the T1-weighted MRI images, FreeSurfer v7.2 was employed to perform pre-processing and extract cortical thickness measurements. Simultaneously, [18F] flortaucipir-PET imaging was processed to compute voxel-wise regions of interest (ROIs) for 66 specific brain regions. Regional tau positivity was established using a cutoff at a z-score of 1.25, with a focus on cognitive normal (CN) subjects within the train set. To predict early tau accumulation regions, we developed an attention mechanism-based encoder-decoder model by adopting a Transformer model into our problem setting, performing sequential predictions for each of the 66 regions. Notably, the model's predictive performance in initial regions significantly influences subsequent predictions. Consequently, we implemented a prioritization strategy, emphasizing predictions from areas where the model demonstrated high accuracy. This approach was designed to enhance the overall predictive accuracy of the model.

RESULT: Predicting five early tau accumulation regions per hemisphere, our model achieved an average AUC of 0.84 and accuracy of 84% for the test dataset (112 participants). Notably, in critical early disease progression regions (fusiform gyrus and ITG), AUC values of 0.84, 0.85, and accuracies of 84.4%, 84% were observed. Furthermore, the proposed prioritization strategy improved performance compared to predictions using vanilla attention-based model.

CONCLUSION: We developed an attention mechanism-based architecture with an encoder-decoder structure. By predicting outcomes not only based on cortical thickness values but also their cross-attention-based contexture information, we could achieve highly accurate tau prediction in early and challenging regions.

PMID:39751223 | DOI:10.1002/alz.092140

Alzheimers Dement. 2024 Dec;20 Suppl 1:e088814. doi: 10.1002/alz.088814.

ABSTRACT

BACKGROUND: Alzheimer's disease (AD) has been associated with speech and language impairment. Recent progress in the field has led to the development of automated AD detection using audio-based methods, because it has a great potential for cross-linguistic detection. In this investigation, we utilised a pretrained deep learning model to automatically detect AD, leveraging acoustic data derived from Chinese speech.

METHOD: Speech samples from a picture description task were obtained from 81 native Chinese speakers in Taiwan. This group included 34 normal controls (NC) (14 females; age range: 61-89 years; average age: 73.9 years; SD: 6.7) and 47 patients diagnosed with early AD (24 females; age range: 59-89 years; average age: 76.8 years; SD: 7.6). The audio data were first segmented into 6-second clips, resulting in a total of 1117 clips for the NC group and 1770 for the AD group. Due to the data imbalance, we equalized the groups by randomly selecting 1117 clips from the AD group. The dataset was then divided into a training-to-testing ratio of 8 to 2. The training clips were initially processed using a pre-trained Wav2vec2 model to generate internal acoustic representations. Subsequently, these acoustic representations, now serving as feature data, were input into a two-layer fully connected neural network for additional training and classification.

RESULT: The model achieved a training accuracy of 83%. Notable metrics for test performance were observed as follows: Accuracy - 81.25%, Precision - 80.92%, Recall - 81.79%, and F1 Score - 81.35%. These findings suggest a promising capability of wav2vec2 in Alzheimer's Disease detection, demonstrating a commendable balance between precision and recall.

CONCLUSION: The pre-trained Wav2vec2 model emerges as a promising tool for AD detection through Chinese speech data. This research sets the foundation for additional investigation into the model's potential for cross-linguistic detection using speech data. Additionally, future work can also be conducted to investigate what acoustic features are used and their significances in classification process.

PMID:39751145 | DOI:10.1002/alz.088814

Alzheimers Dement. 2024 Dec;20 Suppl 1:e089241. doi: 10.1002/alz.089241.

ABSTRACT

BACKGROUND: Early detection and accurate forecasting of AD progression are crucial for timely intervention and management. This study leverages multi-modal data, including MRI scans, brain volumetrics, and clinical notes, utilizing Machine Learning (ML), Deep Learning (DL) and a range of ensemble methods to enhance the forecasting accuracy of Alzheimer's disease.

METHOD: We utilize the OASIS-3 longitudinal dataset, tracking 1,098 patients over 30 years. From OASIS-3, we combined three modalities - MRI scans, Freesurfer brain volumetrics, and Clinical Data from the Alzheimer's Disease Research Center (ADRC). We use Convolutional Neural Networks (CNNs), specifically MobileNetV2, ResNet101, ResNet152 and ResNet200 for MRIs and Machine Learning (ML) techniques (Random Forest and K Nearest Neighbors) for Freesurfer featurized brain volumetrics and clinical data. Individual models were tuned for each modality, with the best models combined via ensembles to predict each patient's future Clinical Dementia Rating (CDR). Ensembles evaluated included: aggressive (any modality predicting positive), conservative (all modalities predicting positive), conditional ensembles (majority voting and ADRC or both MRI/Freesurfer), and custom machine learning models built to integrate the modality predictions based on the confidence values returned from the MRI model and the predictions of other models. The figure below shows our experimental pipeline.

RESULT: The study achieved >95% accuracy in predicting future CDR. Ensembles notably reduced harmful False Negatives by 2x-15x, compared to individual modalities, while incurring nominal increases in False Positives. The machine learning trained ensemble demonstrated improved accuracy over the best individual modality predictions The results highlight the potential of multi-modal AI ensemble methods in improving the accuracy of early AD detection and prognosis. The figures below show comparative accuracy and false positives/false negatives rates for each ensemble as compared to the individual modalities.

CONCLUSION: This work demonstrates the potential efficacy of multi-modal data integration via ensemble learning in forecasting Alzheimer's disease, significantly outperforming single-modality methods. It underscores the importance of leveraging diverse data sources and advanced analytical techniques for early diagnosis and intervention in Alzheimer's care, paving the way for future research to explore additional modalities and methods for even greater accuracy and clinical utility.

PMID:39751131 | DOI:10.1002/alz.089241

Alzheimers Dement. 2024 Dec;20 Suppl 1:e088832. doi: 10.1002/alz.088832.

ABSTRACT

BACKGROUND: Alzheimer's disease (AD) related pathologies (i.e., neurofibrillary tangles [NFTs], amyloid-β plaques, and phosphorylated-TAR-DNA-binding-protein-43 [pTDP-43]) differ across sexes. However, the interaction between sex and cerebral small vessel disease (CSVD) (i.e., cerebral amyloid angiopathy [CAA] and arteriolosclerosis) on AD-related pathologies has been less well characterized. The medial temporal lobe (MTL) is a crucial region in AD pathophysiology which harbors AD-pathologies at an early disease stage and is vascularized by small vessels prone to CSVD. We therefore aimed to analyze the relationship between sex and CSVD on AD-related pathologies in the MTL of a pathological AD cohort.

METHODS: The study included autopsy cases from the Massachusetts AD Research Center. One hemisphere was formalin-fixed, samples from pre-defined regions (hippocampal body and entorhinal cortex) were stained for luxol fast blue with hematoxylin&eosin, amyloid-b, At8, and pTDP-43. Deep-learning models (Aiforiaâ) were used to obtain quantitative measures for burden of CAA, NFTs, and pTDP-43 inclusions. Arteriolosclerosis (grade 0-3) was determined in vessels >20mm Ø. Mixed effect models explored age and sex (fixed terms) across regions of interest (ROIs: hippocampus, parahippocampal gyrus, entorhinal cortex, amygdala) of the MTL on CSVD subtypes. Furthermore, interactions between sex and CSVD subtypes on AD-related pathologies were tested. Finally, ApoE effect was evaluated in a subgroup with available genotyping.

RESULTS: 157 autopsy cases (80.8±12.6y, 91 females) were included, ApoE status was available in 66/157. Females had higher arteriolosclerosis severity (odds ratio for grade 2/3 = 1.56, 95% confidence interval [CI] 1.11; 2.20, p = 0.01) and lower CAA burden (b = -0.08, 95%CI -0.12; -0.04, p<0.001), when adjusting for age and ROIs. Moreover, females had lower burden of amyloid-b plaques (b = -0.25, 95%CI -0.46; -0.04, p = 0.02) and higher density of NFTs (b = 5.16, 95%CI 1.90; 8.41, p = 0.002). Inclusion of ApoE-status confirmed these findings with higher effect-size, and ApoE4 genotype interacted with female sex predicting higher NFT density when adjusting for arteriolosclerosis and CAA respectively (b = 16.2, 95%CI 4.2; 28.2, p = 0.008 and b = 16.3, 95%CI 3.37; 29.22, p = 0.01). No interaction effect was found between sex and CSVD on AD-pathology.

CONCLUSIONS: In this cohort, sex differentially affected microvascular and AD-related pathologies in the MTL. ApoE4 genotype might act as an effect modifier of sex.

PMID:39751121 | DOI:10.1002/alz.088832

Alzheimers Dement. 2024 Dec;20 Suppl 1:e089093. doi: 10.1002/alz.089093.

ABSTRACT

BACKGROUND: Cell-type specific expression quantitative trait loci (eQTLs) can help dissect cellular heterogeneity in the impact of genetic variation on gene expression for Alzheimer's disease (AD) and AD-related dementia (ADRD). However, due to the high cost and stringent sample collection criteria, it is challenging to obtain large single-nuclei RNA sequencing (snRNA-seq) data with sufficient cohort size to match genotyping data to systematically identify human brain-specific eQTLs for AD/ADRD.

METHOD: In this study, we presented a deep learning-based deconvolution framework on large-scale bulk RNA sequencing (RNA-seq) data to infer cell-type specific eQTLs in the human brains with AD/ADRD. Specifically, we first predicted the brain cell-type specific gene expression for the harmonized bulk RNA dataset (n = 1,092) from Religious Orders Study and Memory and Aging Project (ROS/MAP). We then incorporated the inferred cell-type specific gene expression with matched whole genome sequencing (WGS) data from ROS/MAP to identify the brain cell-type specific eQTLs. These cell-type specific eQTLs were further colocalized with AD genome-wide association study (GWAS) findings to discover potential risk genes and druggable targets for AD.

RESULT: We identified 44,504 genome-wide significant cell-type specific cis-eQTLs (window size = 1Mb, p < 5 × 10-8) from eight brain cell types, including excitatory neurons, inhibitory neurons, microglia, oligodendrocytes, oligodendrocyte precursor cells, astrocytes, and endothelium. Approximate 2,732 eQTLs in astrocytes and 7,628 in excitatory neurons are identical to the results from a large existing snRNA-seq data, associated with the regulation of multiple genes (eGenes, e.g., ARL17B, LRRC37A2, ERAP2, PILRB, ZNF266). We illustrated that the GWAS variant rs199456 (AD GWAS: p = 2.57 × 10-9) co-localized with its regulatory effect on LRRC37A2, LRRC37A and ARL17B gene expression in excitatory neurons.

CONCLUSION: In summary, this study presented comprehensive brain cell type-specific eQTL analysis and identified potential eQTL-regulated likely causal genes from AD GWAS findings using a deep learning-based deconvolution framework. It offers an opportunity to efficiently discover the effect of GWAS loci on gene expression at cell type-specific manners using genotyping data and matched bulk RNA-seq data instead of costly snRNA-seq data. Functional validation of candidate eQTLs and associated genes are warranted in the future.

PMID:39751088 | DOI:10.1002/alz.089093

Alzheimers Dement. 2024 Dec;20 Suppl 1:e089010. doi: 10.1002/alz.089010.

ABSTRACT

BACKGROUND: Alzheimer's Disease (AD) is a widespread neurodegenerative disease with Mild Cognitive Impairment (MCI) acting as an interim phase between normal cognitive state and AD. The irreversible nature of AD and the difficulty in early prediction present significant challenges for patients, caregivers, and the healthcare sector. Deep learning (DL) methods such as Recurrent Neural Networks (RNN) have been utilized to analyze Electronic Health Records (EHR) to model disease progression and predict diagnosis. However, these models do not address some inherent irregularities in EHR data such as irregular time intervals between clinical visits. Furthermore, most DL models are not interpretable. To address these issues, we developed a novel DL architecture called Time-Aware RNN (TA-RNN) to predict MCI to AD conversion at the next clinical visit.

METHOD: TA-RNN comprises of a time embedding layer, attention-based RNN, and prediction layer based on multi-layer perceptron (MLP) (Figure 1). For interpretability, a dual-level attention mechanism within the RNN identifies significant visits and features impacting predictions. TA-RNN addresses irregular time intervals by incorporating time embedding into longitudinal cognitive and neuroimaging data based on attention weights to create a patient embedding. The MLP, trained on demographic data and the patient embedding predicts AD conversion. TA-RNN was evaluated on Alzheimer's Disease Neuroimaging Initiative (ADNI) and National Alzheimer's Coordinating Center (NACC) datasets based on F2 score and sensitivity.

RESULT: Multiple TA-RNN models were trained with two, three, five, or six visits to predict the diagnosis at the next visit. In one setup, the models were trained and tested on ADNI. In another setup, the models were trained on the entire ADNI dataset and evaluated on the entire NACC dataset. The results indicated superior performance of TA-RNN compared to state-of-the-art (SOTA) and baseline approaches for both setups (Figure 2A and 2B). Based on attention weights, we also highlighted significant visits (Figure 3A) and features (Figure 3B) and observed that CDRSB and FAQ features and the most recent visit had highest influence in predictions.

CONCLUSION: We propose TA-RNN, an interpretable model to predict MCI to AD conversion while handling irregular time intervals. TA-RNN outperformed SOTA and baseline methods in multiple experiments.

PMID:39751068 | DOI:10.1002/alz.089010

Brief Bioinform. 2024 Nov 22;26(1):bbae686. doi: 10.1093/bib/bbae686.

ABSTRACT

Recent studies suggest cGAS-STING pathway may play a crucial role in the genesis and development of hepatocellular carcinoma (HCC), closely associated with classical pathways and tumor immunity. We aimed to develop models predicting survival and anti-PD-1/PD-L1 outcomes centered on the cGAS-STING pathway in HCC. We identified classical pathways highly correlated with cGAS-STING pathway and constructed transformer survival model preserving raw structure of pathways. We also developed explainable XGBoost model for predicting anti-PD-1/PD-L1 outcomes using SHAP algorithm. We trained and validated transformer survival model on pan-cancer cohort and tested it on three independent HCC cohorts. Using 0.5 as threshold across cohorts, we divided each HCC cohort into two groups and calculated P values with log-rank test. TCGA-LIHC: C-index = 0.750, P = 1.52e-11; ICGC-LIRI-JP: C-index = 0.741, P = .00138; GSE144269: C-index = 0.647, P = .0233. We trained and validated [area under the receiver operating characteristic curve (AUC) = 0.777] XGBoost model on immunotherapy datasets and tested it on GSE78220 (AUC = 0.789); we also tested XGBoost model on HCC anti-PD-L1 cohort (AUC = 0.719). Our deep learning model and XGBoost model demonstrate potential in predicting survival risks and anti-PD-1/PD-L1 outcomes in HCC. We deployed these two prediction models to the GitHub repository and provided detailed instructions for their usage: deep learning survival model, https://github.com/mlwalker123/CSP_survival_model; XGBoost immunotherapy model, https://github.com/mlwalker123/CSP_immunotherapy_model.

PMID:39749665 | DOI:10.1093/bib/bbae686

J Chem Inf Model. 2025 Jan 3. doi: 10.1021/acs.jcim.4c01737. Online ahead of print.

ABSTRACT

As combination therapy becomes more common in clinical applications, predicting adverse effects of combination medications is a challenging task. However, there are three limitations of the existing prediction models. First, they rely on a single view of the drug and cannot fully utilize multiview information, resulting in limited performance when capturing complex structures. Second, they ignore subgraph information at different scales, which limits the ability to model interactions between subgraphs. Third, there has been limited research on effectively integrating multiview features of molecules. Therefore, we propose ComNet, a deep learning model that improves the accuracy of side effect prediction by integrating multiview features of drugs. First, to capture diverse features of drugs, a multiview feature extraction module is proposed, which not only uses molecular fingerprints but also extracts semantic information on SMILES and spatial information on 3D conformations. Second, to enhance the modeling ability of complex structures, a multiscale subgraph fusion mechanism is proposed, which can fuse local and global graph structures of drugs. Finally, a multiview feature fusion mechanism is proposed, which uses an attention mechanism to adaptively adjust the weights of different views to achieve multiview data fusion. Experiments on several publicly available data sets show that ComNet performs better than existing methods in various complex scenarios, especially in cold-start scenarios. Ablation experiments show that each core structure in ComNet contributes to the overall performance. Further analysis shows that ComNet not only converges rapidly and has good generalization ability but also identifies different substructures in the molecule. Finally, a case study on a self-collected data set validates the superior performance of ComNet in practical applications.

PMID:39749659 | DOI:10.1021/acs.jcim.4c01737

Front Comput Neurosci. 2024 Dec 19;18:1523973. doi: 10.3389/fncom.2024.1523973. eCollection 2024.

NO ABSTRACT

PMID:39749286 | PMC:PMC11693727 | DOI:10.3389/fncom.2024.1523973

Heliyon. 2024 Nov 14;10(22):e40345. doi: 10.1016/j.heliyon.2024.e40345. eCollection 2024 Nov 30.

ABSTRACT

Cyclin Dependent Kinase 1 (CDK1) plays a crucial role in cell cycle regulation, and dysregulation of its activity has been implicated in various cancers. Although several CDK1 inhibitors are currently in clinical trials, none have yet been approved for therapeutic use. This research utilized deep learning techniques, specifically Recurrent Neural Networks with Long Short-Term Memory (LSTM), to generate potential CDK1 inhibitors. Molecular docking, evaluation of molecular properties, and molecular dynamics simulations were conducted to identify the most promising candidates. The results showed that the generated ligands exhibited substantial improvements in target affinity and drug-likeness. Molecular docking results showed that the generated ligands had an average binding affinity of -10.65 ± 0.877 kcal/mol towards CDK1. The Quantitative Estimate of Drug-likeness (QED) values for the generated ligands averaged 0.733 ± 0.10, significantly higher than the 0.547 ± 0.15 observed for known CDK1 inhibitors (p < 0.001). Molecular dynamics simulations further confirmed the stability and favorable interactions of the selected ligands with the CDK1 complex. The identification of novel CDK1 inhibitors with improved binding affinities and drug-likeness properties could potentially fill the gap in the ongoing development of CDK inhibitors. However, it is imperative to note that extensive experimental validation is required prior to advancing these generated ligands to subsequent stages of drug development.

PMID:39748968 | PMC:PMC11693894 | DOI:10.1016/j.heliyon.2024.e40345

IEEE Access. 2024;12:191219-191230. doi: 10.1109/access.2024.3516572. Epub 2024 Dec 12.

ABSTRACT

Given telemetry datasets (e.g., GPS location, speed, direction, distance.), the Older Driver Classification (ODC) problem identifies two groups of drivers: normal and abnormal. The ODC problem is essential in many societal applications, including road safety, insurance risk assessment, and targeted interventions for elderly drivers with dementia or Mild Cognitive Impairment (MCI). The problem is challenging because of the volume and heterogeneity of temporally-detailed vehicle datasets. This paper proposes a novel spatial deep-learning approach that leverages Grid-Index based data augmentation to enhance the detection of abnormal driving behaviors. Through extensive experiments and a real-world case study, the proposed approach consistently identifies abnormal drivers with high accuracy. The findings demonstrate the potential of grid-based methods to improve telematics-based driving behavior analysis significantly. This approach offers valuable implications for enhancing road safety measures, optimizing insurance risk assessments, and developing targeted interventions for at-risk drivers.

PMID:39748855 | PMC:PMC11694628 | DOI:10.1109/access.2024.3516572

Front Plant Sci. 2024 Dec 19;15:1512632. doi: 10.3389/fpls.2024.1512632. eCollection 2024.

ABSTRACT

Molecular-based detection of pathogens from potato tubers hold promise, but the initial sample extraction process is labor-intensive. Developing a robotic tuber sampling system, equipped with a fast and precise machine vision technique to identify optimal sampling locations on a potato tuber, offers a viable solution. However, detecting sampling locations such as eyes and stolon scar is challenging due to variability in their appearance, size, and shape, along with soil adhering to the tubers. In this study, we addressed these challenges by evaluating various deep-learning-based object detectors, encompassing You Look Only Once (YOLO) variants of YOLOv5, YOLOv6, YOLOv7, YOLOv8, YOLOv9, YOLOv10, and YOLO11, for detecting eyes and stolon scars across a range of diverse potato cultivars. A robust image dataset obtained from tubers of five potato cultivars (three russet skinned, a red skinned, and a purple skinned) was developed as a benchmark for detection of these sampling locations. The mean average precision at an intersection over union threshold of 0.5 (mAP@0.5) ranged from 0.832 and 0.854 with YOLOv5n to 0.903 and 0.914 with YOLOv10l. Among all the tested models, YOLOv10m showed the optimal trade-off between detection accuracy (mAP@0.5 of 0.911) and inference time (92 ms), along with satisfactory generalization performance when cross-validated among the cultivars used in this study. The model benchmarking and inferences of this study provide insights for advancing the development of a robotic potato tuber sampling device.

PMID:39748820 | PMC:PMC11693691 | DOI:10.3389/fpls.2024.1512632

Mayo Clin Proc Digit Health. 2024 Dec;2(4):548-558. doi: 10.1016/j.mcpdig.2024.08.005. Epub 2024 Aug 26.

ABSTRACT

OBJECTIVE: To report the development and performance of 2 distinct deep learning models trained exclusively on retinal color fundus photographs to classify Alzheimer disease (AD).

PATIENTS AND METHODS: Two independent datasets (UK Biobank and our tertiary academic institution) of good-quality retinal photographs derived from patients with AD and controls were used to build 2 deep learning models, between April 1, 2021, and January 30, 2024. ADVAS is a U-Net-based architecture that uses retinal vessel segmentation. ADRET is a bidirectional encoder representations from transformers style self-supervised learning convolutional neural network pretrained on a large data set of retinal color photographs from UK Biobank. The models' performance to distinguish AD from non-AD was determined using mean accuracy, sensitivity, specificity, and receiving operating curves. The generated attention heatmaps were analyzed for distinctive features.

RESULTS: The self-supervised ADRET model had superior accuracy when compared with ADVAS, in both UK Biobank (98.27% vs 77.20%; P<.001) and our institutional testing data sets (98.90% vs 94.17%; P=.04). No major differences were noted between the original and binary vessel segmentation and between both eyes vs single-eye models. Attention heatmaps obtained from patients with AD highlighted regions surrounding small vascular branches as areas of highest relevance to the model decision making.

CONCLUSION: A bidirectional encoder representations from transformers style self-supervised convolutional neural network pretrained on a large data set of retinal color photographs alone can screen symptomatic AD with high accuracy, better than U-Net-pretrained models. To be translated in clinical practice, this methodology requires further validation in larger and diverse populations and integrated techniques to harmonize fundus photographs and attenuate the imaging-associated noise.

PMID:39748801 | PMC:PMC11695061 | DOI:10.1016/j.mcpdig.2024.08.005

Pathologica. 2024 Dec;116(6):390-403. doi: 10.32074/1591-951X-1069.

ABSTRACT

The search for reliable prognostic markers in oral squamous cell carcinoma (OSCC) remains a critical need. Tumor-infiltrating lymphocytes (TILs), particularly T lymphocytes, play a pivotal role in the immune response against tumors and are strongly correlated with favorable prognoses. Computational pathology has proven highly effective for histopathological image analysis, automating tasks such as cell detection, classification, and segmentation.

In the present study, we developed a StarDist-based model to automatically detect T lymphocytes in hematoxylin and eosin (H&E)-stained whole-slide images (WSIs) of OSCC, bypassing the need for traditional immunohistochemistry (IHC). Using QuPath, we generated training datasets from annotated slides, employing IHC as the ground truth. Our model was validated on Cancer Genome Atlas-derived OSCC images, and survival analyses demonstrated that higher TIL densities correlated with improved patient outcomes.

This work introduces an efficient, AI-powered workflow for automated immune profiling in OSCC, offering a reproducible and scalable approach for diagnostic and prognostic applications.

PMID:39748724 | DOI:10.32074/1591-951X-1069

Microsc Res Tech. 2025 Jan 2. doi: 10.1002/jemt.24788. Online ahead of print.

ABSTRACT

Microscopic imaging aids disease diagnosis by describing quantitative cell morphology and tissue size. However, the high spatial resolution of these images poses significant challenges for manual quantitative evaluation. This project proposes using computer-aided analysis methods to address these challenges, enabling rapid and precise clinical diagnosis, course analysis, and prognostic prediction. This research introduces advanced deep learning frameworks such as squeeze-and-excitation and dilated dense convolution blocks to tackle the complexities of quantifying small and intricate breast cancer tissues and meeting the real-time requirements of pathological image analysis. Our proposed framework integrates a dense convolutional network (DenseNet) with an attention mechanism, enhancing the capability for rapid and accurate clinical assessments. These multi-classification models facilitate the precise prediction and segmentation of breast lesions in microscopic images by leveraging lightweight multi-scale feature extraction, dynamic region attention, sub-region classification, and regional regularization loss functions. This research will employ transfer learning paradigms and data enhancement methods to enhance the models' learning further and prevent overfitting. We propose the fine-tuning employing pre-trained architectures such as VGGNet-19, ResNet152V2, EfficientNetV2-B1, and DenseNet-121, modifying the final pooling layer in each model's last block with an SPP layer and associated BN layer. The study uses labeled and unlabeled data for tissue microscopic image analysis, enhancing models' robust features and classification abilities. This method reduces the costs and time associated with traditional methods, alleviating the burden of data labeling in computational pathology. The goal is to provide a sophisticated, efficient quantitative pathological image analysis solution, improving clinical outcomes and advancing the computational field. The model, trained, validated, and tested on a microscope breast image dataset, achieved recognition accuracy of 99.6% for benign and malignant secondary classification and 99.4% for eight breast subtypes classification. Our proposed approach demonstrates substantial improvement compared to existing methods, which generally report lower accuracies for breast subtype classification ranging between 85% and 94%. This high level of accuracy underscores the potential of our approach to provide reliable diagnostic support, enhancing precision in clinical decision-making.

PMID:39748498 | DOI:10.1002/jemt.24788

Cardiovasc Diabetol. 2025 Jan 2;24(1):3. doi: 10.1186/s12933-024-02564-w.

ABSTRACT

BACKGROUND: Prior studies have demonstrated an association between retinal vascular features and cardiovascular disease (CVD), however most studies have only evaluated a few simple parameters at a time. Our aim was to determine whether a deep-learning artificial intelligence (AI) model could be used to predict CVD outcomes from routinely obtained diabetic retinal screening photographs and to compare its performance to a traditional clinical CVD risk score.

METHODS: We included 6127 individuals with type 2 diabetes without myocardial infarction or stroke prior to study entry. The cohort was divided into training (70%), validation (10%) and testing (20%) cohorts. Clinical 10-year CVD risk was calculated using the pooled cohort equation (PCE) risk score. A polygenic risk score (PRS) for coronary heart disease was also obtained. Retinal images were analysed using an EfficientNet-B2 network to predict 10-year CVD risk. The primary outcome was time to first major adverse CV event (MACE) including CV death, myocardial infarction or stroke.

RESULTS: 1241 individuals were included in the test cohort (mean PCE 10-year CVD risk 35%). There was a strong correlation between retinal predicted CVD risk and the PCE risk score (r = 0.66) but not the polygenic risk score (r = 0.05). There were 288 MACE events. Higher retina-predicted risk was significantly associated with increased 10-year risk of MACE (HR 1.05 per 1% increase; 95% CI 1.04-1.06, p < 0.001) and remained so after adjustment for the PCE and polygenic risk score (HR 1.03; 95% CI 1.02-1.04, p < 0.001). The retinal risk score had similar performance to the PCE (both AUC 0.697) and when combined with the PCE and polygenic risk score had significantly improved performance compared to the PCE alone (AUC 0.728). An increase in retinal-predicted risk within 3 years was associated with subsequent increased MACE likelihood.

CONCLUSIONS: A deep-learning AI model could accurately predict MACE from routine retinal screening photographs with a comparable performance to traditional clinical risk assessment in a diabetic cohort. Combining the AI-derived retinal risk prediction with a coronary heart disease polygenic risk score improved risk prediction. AI retinal assessment might allow a one-stop CVD risk assessment at routine retinal screening.

PMID:39748380 | DOI:10.1186/s12933-024-02564-w

Sci Rep. 2025 Jan 2;15(1):230. doi: 10.1038/s41598-024-81862-5.

ABSTRACT

MiRNAs and lncRNAs are two essential noncoding RNAs. Predicting associations between noncoding RNAs and diseases can significantly improve the accuracy of early diagnosis.With the continuous breakthroughs in artificial intelligence, researchers increasingly use deep learning methods to predict associations. Nevertheless, most existing methods face two major issues: low prediction accuracy and the limitation of only being able to predict a single type of noncoding RNA-disease association. To address these challenges, this paper proposes a method called K-Means and multigraph Contrastive Learning for predicting associations among miRNAs, lncRNAs, and diseases (K-MGCMLD). The K-MGCMLD model is divided into four main steps. The first step is the construction of a heterogeneous graph. The second step involves down sampling using the K-means clustering algorithm to balance the positive and negative samples. The third step is to use an encoder with a Graph Convolutional Network (GCN) architecture to extract embedding vectors. Multigraph contrastive learning, including both local and global graph contrastive learning, is used to help the embedding vectors better capture the latent topological features of the graph. The fourth step involves feature reconstruction using the balanced positive and negative samples and the embedding vectors fed into an XGBoost classifier for multi-association classification prediction. Experimental results have shown that AUC value for miRNA-disease association is 0.9542, lncRNA-disease association is 0.9603, and lncRNA-miRNA association is 0.9687. Additionally, this study has conducted case analyses using K-MGCMLD, which has validated the associations of all the top 30 miRNAs predicted to be associated with lung cancer and Alzheimer's diseases.

PMID:39747154 | DOI:10.1038/s41598-024-81862-5

Nat Commun. 2025 Jan 2;16(1):136. doi: 10.1038/s41467-024-54970-z.

ABSTRACT

Microsatellite instability (MSI) is a critical phenotype of cancer genomes and an FDA-recognized biomarker that can guide treatment with immune checkpoint inhibitors. Previous work has demonstrated that next-generation sequencing data can be used to identify samples with MSI-high phenotype. However, low tumor purity, as frequently observed in routine clinical samples, poses a challenge to the sensitivity of existing algorithms. To overcome this critical issue, we developed MiMSI, an MSI classifier based on deep neural networks and trained using a dataset that included low tumor purity MSI cases in a multiple instance learning framework. On a challenging yet representative set of cases, MiMSI showed higher sensitivity (0.895) and auROC (0.971) than MSISensor (sensitivity: 0.67; auROC: 0.907), an open-source software previously validated for clinical use at our institution using MSK-IMPACT large panel targeted NGS data. In a separate, prospective cohort, MiMSI confirmed that it outperforms MSISensor in low purity cases (P = 8.244e-07).

PMID:39746944 | DOI:10.1038/s41467-024-54970-z

Mol Cell Proteomics. 2024 Dec 31:100901. doi: 10.1016/j.mcpro.2024.100901. Online ahead of print.

ABSTRACT

Animal venoms, distinguished by their unique structural features and potent bioactivities, represent a vast and relatively untapped reservoir of therapeutic molecules. However, limitations associated with comprehensively constructing and expressing highly complex venom and venom-like molecule libraries have precluded their therapeutic evaluation via high throughput screening. Here, we developed an innovative computational approach to design a highly diverse library of animal venoms and "metavenoms". We employed programmable M13 hyperphage display to preserve critical disulfide-bonded structures for highly parallelized single-round biopanning with quantitation via high-throughput DNA sequencing. Our approach led to the discovery of Kunitz type domain containing proteins that target the human itch receptor Mas-related G protein-coupled receptor X4 (MRGPRX4), which plays a crucial role in itch perception. Deep learning-based structural homology mining identified two endogenous human homologs, tissue factor pathway inhibitor (TFPI) and serine peptidase inhibitor, Kunitz type 2 (SPINT2), which exhibit agonist-dependent potentiation of MRGPRX4. Highly multiplexed screening of animal venoms and metavenoms is therefore a promising approach to uncover new drug candidates.

PMID:39746545 | DOI:10.1016/j.mcpro.2024.100901