Eun Hye Lee, Taeho Jo*, Computational and Structural Biotechnology Journal (2026)

DuAL-Net (Dual Approach Local-global Network) is a hybrid framework that integrates local genomic window analysis with global annotation-based modeling to prioritize disease-associated single-nucleotide polymorphisms in high-dimensional whole-genome sequencing data. As a proof of concept, the authors applied it to 14,094 SNPs within the APOE plus or minus 50-kb region from 1,050 individuals in the Alzheimer's Disease Neuroimaging Initiative and Alzheimer's Disease Sequencing Project (ADSP) cohorts.

Under nested 5-fold cross-validation, the top 100 ranked SNPs achieved an AUC of 0.698 (95% CI 0.659 to 0.737), substantially outperforming bottom-ranked SNPs (AUC = 0.479). Validation on an independent ADSP Alzheimer's Disease Centers cohort (n = 5,570) confirmed generalizability, with top-ranked SNPs reaching AUC = 0.686 versus 0.516 for bottom-ranked SNPs. The framework recovered established APOE risk variants including rs429358 and rs7412, supporting its ability to prioritize biologically relevant SNPs for Alzheimer's disease risk prediction.

Taeho Jo, Eun Hye Lee, for the Alzheimer's Disease Neuroimaging Initiative (ADNI) and the Alzheimer's Disease Sequencing Project (ADSP), Briefings in Bioinformatics (2025)

The authors introduce TrUE-Net, an uncertainty-aware deep learning framework that combines a transformer encoder with a random forest ensemble and uses Monte Carlo Dropout to quantify prediction confidence for Alzheimer's disease classification from whole-genome sequencing data. Using 1,050 ADNI participants (607 AD, 443 cognitively normal) and 14,094 common variants in the APOE 50kb region, SNPs were tokenized into 100-variant segments, embedded to 128 dimensions, and processed through 2 encoder layers with 4 attention heads, then combined with a random forest branch using an ensemble weight of alpha = 0.6.

On 525 test individuals, the model reached 65.1% accuracy (AUC 0.664, F1 0.668) overall. Applying a variance threshold of 0.074 stratified samples into a "certain" subset (24.6%) with accuracy 72.9%, AUC 0.682, and F1 0.821, versus an "uncertain" group with accuracy 62.6% and F1 0.584. TrUE-Net outperformed 5 of 11 baseline models and was comparable to gradient boosting, with bootstrap and McNemar tests confirming gains over simpler classifiers. Limitations include a European-ancestry-only cohort, exclusion of mild cognitive impairment, common-variant focus, and no external validation.

Eun Hye Lee, Yen-Ning Huang, Tamina Park, Taeho Jo, Kwangsik Nho, Andrew J. Saykin, for the Indiana Alzheimer Disease Research Center, Alzheimer's & Dementia(2025)

The study tested whether temporal changes in plasma proteins relate to baseline cognitive stage and progression to Alzheimer's disease dementia (ADD). Longitudinal SomaScan 7K proteomics from 347 Indiana Alzheimer's Disease Research Center participants (mean age 69.3, 61.0% female; 60.7% cognitively normal, 24.3% MCI, 15.0% ADD at baseline; median 24.4 months and 3 visits of follow-up, with 39 incident ADD cases) were analyzed using linear mixed-effects models with time by cognitive stage interactions and Benjamini-Hochberg FDR correction.

Two proteins (ACES, IGFALS) were associated with baseline cognitive status and six (ACES, C7, ZCD1, IL-17C, CC055, SO5A1) with incident ADD, with FDR-adjusted p-values ranging from 0.002 to 0.049. Changes in these proteins correlated with baseline AD imaging and plasma biomarkers, including ACES with amyloid PET uptake, CC055 with tau PET, and links to NFL and GFAP. Using the STREAMLINE machine learning platform across seven algorithms, a proteomics-based model reached 0.848 AUC-ROC for predicting ADD conversion, versus 0.820 for a plasma biomarker model and 0.613 for the base model.

Taeho Jo, Paula Bice, Kwangsik Nho, Andrew J. Saykin, the Alzheimer's Disease Sequencing Project, Alzheimer & Dementia TRCI (2025)

The study introduces Deep-Block, a multi-stage deep learning framework built to identify Alzheimer's disease genetic loci from large-scale whole genome sequencing (WGS) data. The pipeline first segments the genome using linkage disequilibrium (LD) patterns, then selects disease-relevant LD blocks through a sparse attention mechanism, and finally quantifies SNP feature importance with TabNet and Random Forest models. The framework was applied to WGS data from 7,416 Alzheimer's Disease Sequencing Project (ADSP) participants, comprising 3,150 cognitively normal individuals and 4,266 with Alzheimer's disease.

Deep-Block identified 30,218 LD blocks ranked by relevance to Alzheimer's disease. Within the top 1,500 blocks, it recovered previously reported variants including APOE rs429358 and rs769449, and highlighted novel SNPs. Expression quantitative trait loci (eQTL) analysis across 13 brain regions provided tissue-specific functional evidence, and results were checked against established AD-associated loci from international biobanks. The authors report that the LD-informed design preserves SNP interactions during dimensionality reduction while scaling to large sequencing datasets.

Taeho Jo, Eun Hye Lee, Paula J Bice, Kwangsik Nho, Andrew J. Saykin, Alzheimer's & Dementia (2025) · Poster PDF

Deep-Block is an attention-based deep learning framework applied to the Alzheimer's Disease Sequencing Project (ADSP) Release 4 whole-genome sequencing (WGS) data (36,361 participant genomes) to identify AD-associated genetic variants while preserving biological context. After quality control, 9,956,115 SNPs and 36,329 participants (99.91%) were retained. The genome was segmented into 48,959 linkage disequilibrium (LD) blocks with un-called SNPs imputed via k-Nearest Neighbors. TabNet performed block-level feature selection, flagging 3,040 blocks whose accuracy exceeded 2.0 standard deviations above the mean, then a dual-model approach combining TabNet and Random Forest scored SNP importance, evaluating 4,869 genetic markers for AD association.

On an independent test set (N = 3,000), Deep-Block achieved AUC 0.70, a 40% improvement over random SNP selection (AUC 0.50). Chromosome 19 harbored the largest number of high-priority variants, including known APOE-related markers, and 95.6% of quality-filtered SNPs were incorporated, indicating thorough genomic coverage. The authors report stable cross-validated performance and conclude that the LD-informed deep learning framework can manage the complexity of large-scale WGS data. Planned extensions include multi-ethnic cohorts and multi-omics integration to deepen insight into AD genetic architecture.

Taeho Jo, Alzheimer's & Dementia (2025) · Poster PDF

This study introduces an uncertainty-aware classification framework that applies Monte Carlo Dropout to quantify predictive uncertainty in Alzheimer's disease (AD) genomic classification. Whole Genome Sequencing data from 1,050 participants (443 cognitively normal, 607 AD dementia) from ADNI and ADSP-FUS1-ADNI-WGS-2 were segmented into non-overlapping windows of 100 SNPs and processed by a transformer with two encoder layers, four attention heads, and feed-forward dimension of 512. Dropout (rate = 0.2) remained active during inference, and a Random Forest classifier trained on flattened genomic data was ensembled with the transformer using a learned weight (alpha = 0.6) tuned via 5-fold cross-validation.

On 525 test samples, the framework achieved accuracy 0.6514, AUC 0.6636, and F1 0.6679. Variance-based stratification showed the certain group (24.6% of samples) reached accuracy 0.7287, AUC 0.6816, and F1 0.8205, improvements of 10.24%, 5.48%, and 23.62% over uncertain samples. Among AD patients, 62.44% were APOE epsilon-4 carriers versus 27.09% in controls, and the transformer alone (alpha = 1.0) yielded the lowest AUC of 0.6283.

Taeho Jo, Junpyo Kim, Paula Bice, Kevin Huynh, Tingting Wang, Matthias Arnold, Peter J. Meikle, Corey Giles, Rima Kaddurah-Daouk, Andrew J. Saykin, Kwangsik Nho, eBioMedicine (2023)

This study introduces the Circular Sliding Window Association Test (c-SWAT), a deep learning framework designed to address the challenges of applying neural networks to complex, high-dimensional multi-omics data for Alzheimer's disease (AD) classification. The pipeline combines three techniques: Weighted Gene Co-expression Network Analysis for feature correlation, Convolutional Neural Networks for feature selection, and Random Forest for the final classification step. It was applied to serum-based metabolomics and lipidomic markers from 997 participants in the Alzheimer's Disease Neuroimaging Initiative (ADNI).

Applying c-SWAT achieved a classification accuracy of up to 80.8% and an AUC of 0.808 for distinguishing AD from cognitively normal older adults, corresponding to a 9.4% accuracy improvement and a 0.169 AUC increase over methods without c-SWAT (p = 1.04 x 10^-4). The analysis also highlighted specific lipid biomarkers, including Cer(d16:1/22:0) and PI(37:6), as potential indicators of AD. The authors conclude that c-SWAT enhances classification performance and surfaces candidate lipids that warrant further investigation for understanding the disease.

Taeho Jo, Kwangsik Nho, Shannon L. Risacher, Andrew J. Saykin, AAIC (2023)

This study proposes a deep learning approach that integrates neuroimaging and genetic data to classify Alzheimer's disease (AD) and identify morphological and genetic markers of tau accumulation. On tau-PET data from 300 ADNI participants (66 CN, 66 AD, 97 EMCI, 71 LMCI), the authors trained convolutional neural networks with Layer-wise Relevance Propagation (LRP) to distinguish AD from cognitively normal elderly and to localize AD-related tau deposition patterns.

The tau-PET CNN reached a 5-fold cross-validation classification accuracy of 90.8%, and LRP maps agreed with traditional voxel-wise SPM12 analysis, highlighting significant focal tau deposition in the bilateral temporal lobes including the entorhinal cortex. In parallel, the Sliding Window Association Test (SWAT) was applied to ADNI GWAS data (N = 981; 650 CN, 331 AD), recovering the known APOE region as the top locus and flagging previously unreported SNPs with an AUC of 0.82. A stacking ensemble that combined the imaging and genetic models produced a stronger overall AD classifier.

Taeho Jo, Kwangsik Nho, Andrew J. Saykin, AAIC (2023)

Background and methods. Deep learning applied to whole genome sequencing (WGS) faces the high-dimension low-sample-size (HDLSS) problem, where features vastly outnumber samples. The authors propose SWAT-TAB, a modification of the SWAT-CNN approach that swaps in the TabNet algorithm in the second step. TabNet uses sequential attention to score feature importance at each step and build a trainable mask, selecting the most relevant variants for the label. The method was applied to ADSP WGS data for Chromosome 19, using 7,416 samples (4,266 AD and 3,150 cognitively normal) and 204,486 SNP features.

Results and conclusion. The top-ranked variants on Chromosome 19 were rs429358, rs7256200, rs1969899716, and rs10414043, reflecting known AD genes such as APOE and APOC1 along with several novel features. SWAT-TAB reduced processing time per SNP by 11.58% and improved reproducibility and ease of implementation, requiring only a single layer instead of one layer per feature in the window. The authors conclude the approach is promising for identifying AD-associated variants and building classification models, with larger WGS analyses beyond Chromosome 19 planned.

Taeho Jo, Junpyo Kim, Paula Bice, Kevin Huynh, Tingting Wang, Peter J Meikle, Rima Kaddurah-Daouk, Kwangsik Nho, Andrew J. Saykin,  AAIC (2022)

The study introduces Circling SWAT, an extension of the Sliding Window Association Test, combined with correlation heatmap and dendrogram analysis to derive Phenotype Influence Scores (PIS) for omics data with minimal features. The authors used cross-sectional serum lipidomics from the Alzheimer's Disease Neuroimaging Initiative (ADNI), covering 781 lipids across 216 cognitively normal, 635 MCI, and 382 dementia (AD) participants. A convolutional neural network was trained on the top-scoring lipids to distinguish AD from cognitively normal controls, and the resulting classifier was applied to predict MCI-to-AD conversion within two years using 146 MCI converters and 190 MCI-stable subjects.

Top-ranked lipids by PIS included LPE(22:6), PC(39:5), and PC(15-MHDA_22:6), while deDE(18:2) and deDE(20:4) were excluded because their AD associations were driven by AD-related medication. Using the top 50 lipids, the CNN reached 73.6% accuracy for AD vs CN classification, 5.1% higher than a model using only age, sex, and APOE epsilon 4. Applied to MCI conversion, the AD/CN classifier achieved 69.9% accuracy, 8.8% higher than the demographic and APOE baseline. The authors conclude that Circling SWAT is promising for lipidome analysis, identifying AD-related lipids and predicting disease progression.

Taeho Jo, Kwangsik Nho, Paula Bice, Andrew J Saykin, For The Alzheimer’s Disease Neuroimaging Initiative, Briefings in Bioinformatics (2022)

This work introduces SWAT-CNN, a three-step deep learning approach for identifying phenotype-associated single nucleotide polymorphisms (SNPs) and classifying Alzheimer's disease (AD). The genome is first divided into 134,955 non-overlapping fragments of 40 SNPs, then a convolutional neural network selects phenotype-associated fragments while an overlapping Sliding Window Association Testing procedure computes phenotype influence scores, and finally top-ranked SNPs are fed into a CNN classifier that distinguishes AD from cognitively normal subjects.

Using the ADNI cohort of 981 non-Hispanic European participants (650 cognitively normal, 331 AD) and 5,398,183 SNPs with MAF above 5%, the top 4,000 SNPs achieved 75.02% accuracy and an AUC of 0.8157, outperforming XGBoost by 3.8% and Random Forest by 9.6%, and clearly exceeding the 66.7% accuracy from APOE epsilon4 alleles alone. The method recovered the well-known APOE/APOC1/TOMM40 locus and highlighted novel candidate loci including SNX14, SNX16, BICD1, WDR72 and GLT1D1. Training used 10-fold cross-validation with early stopping.

Taeho Jo, Kwangsik Nho, Andrew J. Saykin, AAIC (2021)

Genome-wide association study (GWAS) designs are widely used to identify genetic loci associated with Alzheimer's disease (AD), but the per-SNP statistical testing creates a heavy multiple testing burden and limits detection of non-linear effects. This work reports preliminary results of a deep learning framework for genome-wide AD-associated variant identification. Genome-wide genotyping data (12,448,786 SNPs after imputation) were drawn from 916 ADNI participants (458 cognitively normal controls, 458 AD patients). A convolutional neural network (CNN) with convolutional, pooling, and fully connected softmax layers was applied in a two-stage design, with data split 60:20:20 into training, testing, and validation, and AUC used to score performance.

The first stage identified 2,335 candidate genetic regions comprising 93,400 SNPs associated with AD. The second stage computed per-SNP p-values from AD influence z-scores derived from the deep learning model, with the APOE region emerging as the most highly associated locus (p < 5×10−8). Case/control classification using the identified SNPs reached mean AUCs of 0.74, 0.79, 0.82, and 0.90 at p-value thresholds of 1×10−5 (114 SNPs), 1×10−4 (243 SNPs), 1×10−3 (724 SNPs), and 1×10−2 (2,846 SNPs), with mean accuracies of 0.66, 0.69, 0.73, and 0.81 respectively. The authors conclude that the framework can identify AD-associated loci and reduce the computational cost of detecting non-linear SNP-SNP interactions, with planned extensions integrating additional omics layers and independent replication.

Taeho Jo, Kwangsik Nho, Shannon L. Risacher & Andrew J. Saykin for the Alzheimer’s Neuroimaging Initiative, BMC Bioinformatics (2020)

This study develops a deep learning framework to identify informative features in tau positron emission tomography (PET) for Alzheimer's disease (AD) classification. Using data from 300 ADNI participants, the authors trained a 3D convolutional neural network (CNN) on tau PET scans normalized to Montreal Neurologic Institute space, then applied layer-wise relevance propagation (LRP) to visualize the imaging regions that drove each prediction. Model performance was evaluated with five-fold cross-validation.

The 3D CNN classified AD versus cognitively normal (CN) participants with an average accuracy of 90.8%. LRP consistently highlighted the hippocampus, parahippocampus, thalamus, and fusiform as the most informative regions. In mild cognitive impairment (MCI) participants, the classifier's AD probability scores correlated with medial temporal lobe tau deposition (r = 0.43 in early MCI and r = 0.49 in late MCI). The authors conclude that combining 3D CNN with LRP on tau PET can surface informative features for AD classification and may support early detection during prodromal stages of the disease.

Taeho Jo, Kwangsik Nho, Shannon L. Risacher, Andrew J. Saykin, AAIC (2020)

The authors present a deep learning framework that reads [18F] flortaucipir tau PET images to distinguish Alzheimer disease from cognitively normal participants and then explains which brain regions drove each decision. A 3D convolutional neural network is trained on ADNI tau PET data, with layer-wise relevance propagation used to highlight the voxels most informative for classification.

Reported results include an average classification accuracy of 90.8 percent under five-fold cross-validation. The relevance maps concentrate on the hippocampus, parahippocampus, thalamus, and fusiform, consistent with conventional voxel-wise analyses showing AD associated tau deposition in the bilateral temporal lobes. Applied to early and late mild cognitive impairment participants, the classifier scores track tau accumulation, which the authors argue supports use of the approach for early detection during the prodromal stages of AD.

Taeho Jo, Kwangsik Nho, Andrew J. Saykin, Frontiers in Aging Neuroscience (2019)

This systematic review surveys deep learning methods applied to neuroimaging for Alzheimer's disease diagnostic classification and prognostic prediction. The authors searched PubMed and Google Scholar for work published between January 2013 and July 2018, screening 389 initial hits down to 16 studies that met the full inclusion criteria. All 16 reviewed studies used data from the Alzheimer's Disease Neuroimaging Initiative (ADNI), with MRI used in 13 studies, FDG-PET in 10, combined MRI and FDG-PET in 12, and one study combining amyloid PET with FDG-PET.

Across the reviewed literature, hybrid methods that combined traditional machine learning for classification with a stacked auto-encoder for feature selection produced accuracies of up to 98.8% for AD classification and 83.7% for predicting conversion from mild cognitive impairment (MCI) to AD. Pure deep learning approaches, such as CNN-based models, reached up to 96.0% for AD classification and 84.2% for MCI conversion prediction. The review reports that multimodal neuroimaging combined with biomarkers achieved the best performance, and that PET data yielded better results than MRI alone.

Taeho Jo, Kwangsik Nho, Shannon L. Risacher, Andrew J. Saykin, AAIC (2019)

Early accurate diagnosis of Alzheimer's disease (AD) is important for initiating therapeutic development and effective patient care. The authors investigate a deep learning approach that fuses multimodal neuroimaging, genetic, and demographic data to distinguish AD from cognitively normal adults (CN). Demographic information, 3D MRI, PET ([18F]FDG and [18F]Florbetapir), and APOE data from ADNI (N=329; 185 CN and 144 AD) were processed with FreeSurfer and SPM whole-brain voxel-based analysis. The Multimodal-3DCNN pipeline first passes MRI and PET volumes through a 3D convolutional neural network (3D-CNN), then fuses the CNN output with APOE ε4 genotype and demographic information (age, sex, education, handedness, etc.) via a gram matrix, and finally uses a deep neural network (DNN) to classify AD versus CN. Performance was evaluated with 5-fold cross-validation.

Using only MRI plus APOE ε4 and demographic information, the model reached 78% accuracy (AUC = 0.87). Adding [18F]Florbetapir PET to MRI, APOE ε4, and demographics improved performance to 87% accuracy (AUC = 0.93). The best performance, 94% accuracy (AUC = 0.95), was achieved by combining MRI, [18F]FDG PET, and [18F]Florbetapir PET together with APOE ε4 and demographic information. The authors conclude that Multimodal-3DCNN yields high diagnostic classification accuracy and has potential to identify AD individuals to support trial enrichment and therapeutic development, with future work planned on disease progression prediction and additional genetic markers.

Taeho Jo, Kwangsik Nho, Shannon L. Risacher, Jingwen Yan, Andrew J. Saykin, AAIC (2018)

This AAIC 2018 abstract by Jo, Nho, Risacher, Yan, and Saykin introduces Multimodal-CNN, a deep learning approach that improves MRI-based classification of Alzheimer's disease by combining neuroimaging with clinical information. Two-dimensional image patches were extracted from the training set using hippocampal segmentations, while clinical variables and genetic data (APOE epsilon-4 allele count) were encoded as 2D matrices and merged with intermediate CNN layers, allowing the network to learn from both imaging and non-imaging modalities.

The analysis used baseline neuroimaging from 103 participants with Alzheimer's disease and 144 cognitively normal controls. By incorporating the encoded clinical and genetic features into the CNN pipeline, the authors report improved diagnostic accuracy over MRI-only classification, supporting the value of multimodal fusion for AD detection.

Taeho Jo, Jie Hou, Jesse Eickholt & Jianlin Cheng, Scientific Reports (2015)

The paper introduces DN-Fold, a deep learning network method that predicts whether a given query-template protein pair belongs to the same structural fold. Inputs are drawn from protein sequence and structural features extracted from each pair, and the approach is benchmarked against 18 other methods on Lindahl's dataset and on a large SCOP 1.75 benchmark containing about one million protein pairs, evaluated at three levels of fold recognition (family, superfamily, and fold) based on evolutionary distance.

The ensembled DN-Fold achieved a correct recognition rate for Top 1 predictions of 84.5%, 61.5%, and 33.6% at the family, superfamily, and fold levels respectively, and for Top 5 predictions 91.2%, 76.5%, and 60.7%. These results place DN-Fold competitively among state-of-the-art fold recognition methods.

Taeho Jo & Jianlin Cheng, BMC Bioinformatics (2014)

This paper casts protein fold recognition as a binary classification problem and introduces RF-Fold, a random forest method that identifies structurally similar template proteins for a given target. The classifier uses 500 decision trees trained on 84 pairwise similarity features spanning five categories: sequence/family information, sequence-sequence alignment, sequence-profile alignment, profile-profile alignment, and structural information. The model is trained and evaluated with 10-fold cross-validation on 976 by 975 target-template pairs from the FOLDpro dataset, a highly imbalanced setting with 7,438 positive and 944,162 negative examples.

On the Lindahl benchmark, RF-Fold reaches 84.5%, 63.4%, and 40.8% accuracy at the family, superfamily, and fold levels for the top-one ranked template, and 91.5%, 79.3%, and 58.3% for the top-five case. Compared with 17 existing methods, RF-Fold delivers comparable or better performance, and against FOLDpro it improves top-one accuracy by roughly 9% at the superfamily level and 14.3% at the fold level.

Taeho Jo, Mariko Shono, Masato Wada, Sayaka Ito, Junko Nomoto, Yukichi Hara, Membrane (2010)

The authors predict the three-dimensional structure of Heterosigma akashiwo Na+-ATPase (HANA), an algal membrane protein, by homology modeling based on the crystal structure of the K+-bound form of shark Na+/K+-ATPase (PDB ID: 2ZXE). The overall HANA fold resembles the shark enzyme, including the three cytoplasmic domains (A, N, and P) that are characteristic of P-type ATPases. A distinguishing feature is a long TM7-8 junction that forms a large extracellular domain in HANA, occupying the structural role played by the beta-subunit in shark Na+/K+-ATPase.

Using valence mapping on the constructed model, the study identifies two putative K+-binding sites within the transmembrane domain of HANA. Combined with previously reported ion requirements for ATPase activity and phosphoenzyme (EP) formation, the presence of these K+-binding sites suggests that HANA may transport K+ ions in a manner analogous to animal Na+/K+-ATPases.