Independent Validation Dataset Research Articles

Identification of age- and immune-related gene signatures for clinical outcome prediction in lung adenocarcinoma.

The understanding of the factors causing decreased overall survival (OS) in older patients compared to younger patients in lung adenocarcinoma (LUAD) remains. Gene expression profiles of LUAD were obtained from publicly available databases by Kaplan-Meier analysis was performed to determine whether age was associated with patient OS. The immune cell composition in the tumor microenvironment (TME) was evaluated using CIBERSORT. The fraction of stromal and immune cells in tumor samples were also using assessed using multiple tools including ESTIMATE, EPIC, and TIMER. Differentially expressed genes (DEGs) from the RNA-Seq data that were associated with age and immune cell composition were identified using the R package DEGseq. A 22-gene signature composed of DEGs associated with age and immune cell composition that predicted OS were constructed using Least Absolute Shrinkage and Selection Operator (LASSO). In The Cancer Genome Atlas (TCGA)-LUAD dataset, we found that younger patients (≤70) had a significant better OS compared to older patients (>70). In addition, older patients had significantly higher expression of immune checkpoint proteins including inhibitory T cell receptors and their ligands. Moreover, analyses using multiple bioinformatics tools showed increased immune infiltration, including CD4+ T cells, in older patients compared to younger patients. We identified a panel of genes differentially expressed between patients >70 years compared to those ≤70 years, as well as between patients with high or low immune scores and selected 84 common genes to construct a prognostic gene signature. A risk score calculated based on 22 genes selected by LASSO predicted 1, 3, and 5-year OS, with an area under the curve (AUC) of 0.72, 0.72, 0.69, receptively, in TCGA-LUAD dataset and an independent validation dataset available from the European Genome-phenome Archive (EGA). Our results demonstrate that age contributes to OS of LUAD patients atleast in part through its association with immune infiltration in the TME.

A robust immune-related gene pairs signature for predicting the overall survival of esophageal cancer

BackgroundIdentifying reliable biomarkers could effectively predict esophagus carcinoma (EC) patients with poor prognosis. In this work, we constructed an immune-related gene pairs (IRGP) signature to evaluate the prognosis of EC.ResultsThe IRGP signature was trained by the TCGA cohort and validated by three GEO datasets, respectively. Cox regression model together with LASSO was applied to construct the overall survival (OS) associated IRGP. 21 IRGPs consisting of 38 immune-related genes were included in our signature, according to which patients were stratified into high- and low-risk groups. The results of Kaplan-Meier survival analyses indicated that high-risk EC patients had worse OS than low-risk group in the training set, meta-validation set and all independent validation datasets. After adjustment in multivariate Cox analyses, our signature continued to be an independent prognostic factor of EC and the signature-based nomogram could effectively predict the prognosis of EC sufferers. Besides, Gene Ontology analysis revealed this signature is related to immunity. ‘CIBERSORT’ analysis revealed the infiltration levels of plasma cells and activated CD4 memory T cells in two risk groups were significantly different. Ultimately, we validated the expression levels of six selected genes from IRGP index in KYSE-150 and KYSE-450.ConclusionsThis IRGP signature could be applied to select EC patients with high mortality risk, thereby improving prospects for the treatment of EC.

A deep learning-based framework for predicting survival-associated groups in colon cancer by integrating multi-omics and clinical data

Precise prognostic classification of patients and identifying survival subgroups and their associated genes can be important clinical references when designing treatment strategies for cancer patients. Multi-omics and data integration techniques are powerful tools to achieve this goal. This study aimed to introduce a machine learning method to integrate three types of biological data, and investigate the performance of two other methods, in identifying the survival dependency of patients. The data included TCGA RNA-seq gene expression, DNA methylation, and clinical data from 368 patients with colon cancer also we use an independent external validation data set, containing 232 samples. Three methods including, hyper-parameter optimized autoencoders (HPOAE), normal autoencoder, and penalized principal component analysis (PPCA) were used for simultaneous data integration and estimation under a COX hazards model. The HPOAE was thought to outperform other methods. The HPOAE had the Log Rank Mantel-Cox value of 14.27 ± 2, and a Breslow-Generalized Wilcoxon value of 13.13 ± 1. Ten miRNA, 11 methylated genes, and 28 mRNA all by (importance of marginal cutoff > 0.95) were identified. The study demonstrated that hsa-miR-485-5p targets both ZMYM1 and tp53, the latter of which has been previously associated with cancer in numerous studies. Furthermore, compared to other methods, the HPOAE exhibited a greater capacity for identifying survival subgroups and the genes associated with them in patients with colon cancer. However, all of the results were obtained by computational methods, and clinical and experimental studies are needed to validate these results.

Multimodal recurrence scoring system for prediction of clear cell renal cell carcinoma outcome: a discovery and validation study

Improved markers for predicting recurrence are needed to stratify patients with localised (stage I-III) renal cell carcinoma after surgery for selection of adjuvant therapy. We developed a novel assay integrating three modalities-clinical, genomic, and histopathological-to improve the predictive accuracy for localised renal cell carcinoma recurrence. In this retrospective analysis and validation study, we developed a histopathological whole-slide image (WSI)-based score using deep learning allied to digital scanning of conventional haematoxylin and eosin-stained tumour tissue sections, to predict tumour recurrence in a development dataset of 651 patients with distinctly good or poor disease outcome. The six single nucleotide polymorphism-based score, which was detected in paraffin-embedded tumour tissue samples, and the Leibovich score, which was established using clinicopathological risk factors, were combined with the WSI-based score to construct a multimodal recurrence score in the training dataset of 1125 patients. The multimodal recurrence score was validated in 1625 patients from the independent validation dataset and 418 patients from The Cancer Genome Atlas set. The primary outcome measured was the recurrence-free interval (RFI). The multimodal recurrence score had significantly higher predictive accuracy than the three single-modal scores and clinicopathological risk factors, and it precisely predicted the RFI of patients in the training and two validation datasets (areas under the curve at 5 years: 0·825-0·876 vs 0·608-0·793; p<0·05). The RFI of patients with low stage or grade is usually better than that of patients with high stage or grade; however, the RFI in the multimodal recurrence score-defined high-risk stage I and II group was shorter than in the low-risk stage III group (hazard ratio [HR] 4·57, 95% CI 2·49-8·40; p<0·0001), and the RFI of the high-risk grade 1 and 2 group was shorter than in the low-risk grade 3 and 4 group (HR 4·58, 3·19-6·59; p<0·0001). Our multimodal recurrence score is a practical and reliable predictor that can add value to the current staging system for predicting localised renal cell carcinoma recurrence after surgery, and this combined approach more precisely informs treatment decisions about adjuvant therapy. National Natural Science Foundation of China, and National Key Research and Development Program of China.

A foundational vision transformer improves diagnostic performance for electrocardiograms

The electrocardiogram (ECG) is a ubiquitous diagnostic modality. Convolutional neural networks (CNNs) applied towards ECG analysis require large sample sizes, and transfer learning approaches for biomedical problems may result in suboptimal performance when pre-training is done on natural images. We leveraged masked image modeling to create a vision-based transformer model, HeartBEiT, for electrocardiogram waveform analysis. We pre-trained this model on 8.5 million ECGs and then compared performance vs. standard CNN architectures for diagnosis of hypertrophic cardiomyopathy, low left ventricular ejection fraction and ST elevation myocardial infarction using differing training sample sizes and independent validation datasets. We find that HeartBEiT has significantly higher performance at lower sample sizes compared to other models. We also find that HeartBEiT improves explainability of diagnosis by highlighting biologically relevant regions of the EKG vs. standard CNNs. Domain specific pre-trained transformer models may exceed the classification performance of models trained on natural images especially in very low data regimes. The combination of the architecture and such pre-training allows for more accurate, granular explainability of model predictions.

Molecular evolution in different subtypes ofmultifocal hepatocellular carcinoma.

Multifocal hepatocellular carcinoma (MF-HCC) accounts for > 40% of HCCs, exhibiting a poor prognosis than single primary HCCs. Characterizing molecular features including dynamic changes of mutational signature along with clonal evolution, intrahepatic metastatic timing, and genetic footprint in the preneoplastic stage underlying different subtypes of MF-HCC are important for understanding their molecular evolution and developing a precision management strategy. We conducted whole-exome sequencing in 74 tumor samples from spatially distinct regions in 35 resected lesions and adjacent noncancerous tissues from 11 patients, 15 histologically confirmed preneoplastic lesions, and six samples from peripheral blood mononuclear cells. A previously published MF-HCC cohort (n = 9) was included as an independent validation dataset. We combined well-established approaches to investigate tumor heterogeneity, intrahepatic metastatic timing, and molecular footprints in different subtypes of MF-HCCs. We classified MF-HCCs patients into three subtypes, including intrahepatic metastasis, multicentric occurrence, and mixed intrahepatic metastasis and multicentric occurrence. The dynamic changes in mutational signatures between tumor subclonal expansions demonstrated varied etiologies (e.g., aristolochic acid exposure) underlying the clonal progression in different MF-HCC subtypes. Furthermore, the clonal evolution in intrahepatic metastasis exhibited an early metastatic seeding at 10-4-0.01cm3 in primary tumor volume (below the limits of clinical detection), further validated in an independent cohort. In addition, mutational footprints in the preneoplastic lesions for multicentric occurrence patients revealed common preneoplastic arising clones, evidently being ancestors of different tumor lesions. Our study comprehensively characterized the varied tumor clonal evolutionary history underlying different subtypes of MF-HCC and provided important implications for optimizing personalized clinical management for MF-HCC.

Graph-ETMB: A graph neural network-based model for tumour mutation burden estimation

As a critical indicator of how easily the human immune system recognizes tumour cells, tumour mutational burden (TMB) is widely used to identify the potential effectiveness of immune checkpoint inhibitor therapy. However, the difficulties associated with the whole exome sequencing (WES) process, such as high tissue sampling requirements, high costs, and long turnaround times, have hindered the widespread clinical use of WES. Furthermore, the mutation landscape varies across cancer types, and the distribution of TMBs varies across cancer subtypes. Therefore, there is an urgent clinical need to develop a small cancer-specific panel to estimate TMB accurately, predict immunotherapy response cost-effectively and assist physicians in precise decision-making. This paper uses a graph neural network framework (Graph-ETMB) to address the cancer specificity problem in TMB. The correlation and tractability between mutated genes are described through message-passing and aggregation algorithms between graph networks. Then the graph neural network is trained in the lung adenocarcinoma data through a semi-supervised approach, resulting in a mutation panel containing 20 genes with a length of only 0.16 Mb. The number of genes to be detected is smaller than most commercial panels currently in clinical use. In addition, the efficacy of the designed panel in predicting immunotherapy response was further determined in an independent validation dataset, exploring the association between TMB and immunotherapy efficacy.

A Multi-Variable Sentinel-2 Random Forest Machine Learning Model Approach to Predicting Perennial Ryegrass Biomass in Commercial Dairy Farms in Southeast Australia

One of the most valuable and nutritionally essential agricultural commodities worldwide is milk. The European Union and New Zealand are the second- and third-largest exporting regions of milk products and rely heavily on pasture-based production systems. They are comparable to the Australian systems investigated in this study. With projections of herd decline, increased milk yield must be obtained from a combination of animal genetics and feed efficiencies. Accurate pasture biomass estimation across all seasons will improve feed efficiency and increase the productivity of dairy farms; however, the existing time-consuming and manual methods of pasture measurement limit improvements to utilisation. In this study, Sentinel-2 (S2) band and spectral index (SI) information were coupled with the broad season and management-derived datasets using a Random Forest (RF) machine learning (ML) framework to develop a perennial ryegrass (PRG) biomass prediction model accurate to +/−500 kg DM/ha, and that could predict pasture yield above 3000 kg DM/ha. Measurements of PRG biomass were taken from 11 working dairy farms across southeastern Australia over 2019–2021. Of the 68 possible variables investigated, multiple simulations identified 12 S2 bands and 9 SI, management and season as the most important variables, where Short-Wave Infrared (SWIR) bands were the most influential in predicting pasture biomass above 4000 kg DM/ha. Conditional Latin Hypercube Sampling (cLHS) was used to split the dataset into 80% and 20% for model calibration and internal validation in addition to an entirely independent validation dataset. The combined internal model validation showed R2 = 0.90, LCCC = 0.72, RMSE = 439.49 kg DM/ha, NRMSE = 15.08, and the combined independent validation had R2 = 0.88, LCCC = 0.68, RMSE = 457.05 kg DM/ha, NRMSE = 19.83. The key findings of this study indicated that the data obtained from the S2 bands and SI were appropriate for making accurate estimations of PRG biomass. Furthermore, including SWIR bands significantly improved the model. Finally, by utilising an RF ML model, a single ‘global’ model can automate PRG biomass prediction with high accuracy across extensive regions of all seasons and types of farm management.

Prenatal Diagnosis of Placenta Accreta Spectrum Disorders: Deep Learning Radiomics of Pelvic MRI.

Diagnostic performance of placenta accreta spectrum (PAS) by prenatal MRI is unsatisfactory. Deep learning radiomics (DLR) has the potential to quantify the MRI features of PAS. To explore whether DLR from MRI can be used to identify pregnancies with PAS. Retrospective. 324 pregnant women (mean age, 33.3 years) suspected PAS (170 training and 72 validation from institution 1, 82 external validation from institution 2) with clinicopathologically proved PAS (206 PAS, 118 non-PAS). 3-T, turbo spin-echo T2-weighted images. The DLR features were extracted using the MedicalNet. An MRI-based DLR model incorporating DLR signature, clinical model (different clinical characteristics between PAS and non-PAS groups), and MRI morphologic model (radiologists' binary assessment for the PAS diagnosis) was developed. These models were constructed in the training dataset and then validated in the validation datasets. The Student t-test or Mann-Whitney U, χ2 or Fisher exact test, Kappa, dice similarity coefficient, intraclass correlation coefficients, least absolute shrinkage and selection operator logistic regression, multivariate logistic regression, receiver operating characteristic (ROC) curve, DeLong test, net reclassification improvement (NRI) and integrated discrimination improvement (IDI), calibration curve with Hosmer-Lemeshow test, decision curve analysis (DCA). P < 0.05 indicated a significant difference. The MRI-based DLR model had a higher area under the curve than the clinical model in three datasets (0.880 vs. 0.741, 0.861 vs. 0.772, 0.852 vs. 0.675, respectively) or MRI morphologic model in training and independent validation datasets (0.880 vs. 0.760, 0.861, vs. 0.781, respectively). The NRI and IDI were 0.123 and 0.104, respectively. The Hosmer-Lemeshow test had nonsignificant statistics (P = 0.296 to 0.590). The DCA offered a net benefit at any threshold probability. An MRI-based DLR model may show better performance in diagnosing PAS than a clinical or MRI morphologic model. 3 TECHNICAL EFFICACY STAGE: 2.

Identifying the neurodevelopmental and psychiatric signatures of genomic disorders associated with intellectual disability: a machine learning approach

BackgroundGenomic conditions can be associated with developmental delay, intellectual disability, autism spectrum disorder, and physical and mental health symptoms. They are individually rare and highly variable in presentation, which limits the use of standard clinical guidelines for diagnosis and treatment. A simple screening tool to identify young people with genomic conditions associated with neurodevelopmental disorders (ND-GCs) who could benefit from further support would be of considerable value. We used machine learning approaches to address this question.MethodA total of 493 individuals were included: 389 with a ND-GC, mean age = 9.01, 66% male) and 104 siblings without known genomic conditions (controls, mean age = 10.23, 53% male). Primary carers completed assessments of behavioural, neurodevelopmental and psychiatric symptoms and physical health and development. Machine learning techniques (penalised logistic regression, random forests, support vector machines and artificial neural networks) were used to develop classifiers of ND-GC status and identified limited sets of variables that gave the best classification performance. Exploratory graph analysis was used to understand associations within the final variable set.ResultsAll machine learning methods identified variable sets giving high classification accuracy (AUROC between 0.883 and 0.915). We identified a subset of 30 variables best discriminating between individuals with ND-GCs and controls which formed 5 dimensions: conduct, separation anxiety, situational anxiety, communication and motor development.LimitationsThis study used cross-sectional data from a cohort study which was imbalanced with respect to ND-GC status. Our model requires validation in independent datasets and with longitudinal follow-up data for validation before clinical application.ConclusionsIn this study, we developed models that identified a compact set of psychiatric and physical health measures that differentiate individuals with a ND-GC from controls and highlight higher-order structure within these measures. This work is a step towards developing a screening instrument to identify young people with ND-GCs who might benefit from further specialist assessment.

Multi-institutional Development and Validation of Contouring Guidelines for Para-aortic Elective Nodal Irradiation in Prostate Cancer Based on Patterns of Involvement on Targeted Molecular Imaging Positron Emission Tomography/Computed Tomography

Molecular imaging better identifies anatomic regions of metastatic spread of prostate cancer compared with conventional imaging, resulting in para-aortic (PA) nodal metastases being increasingly identified. Consequently, some radiation oncologists electively treat the PA lymph node region in patients with gross or high risk of PA nodal involvement. The anatomic locations of at-risk PA lymph nodes for prostate cancer are unknown. Our objective was to use molecular imaging to develop guidelines for the optimal delineation of the PA clinical target volume (CTV) in patients with prostate cancer. We conducted a multi-institutional retrospective cohort study of patients with prostate cancer undergoing 18F-fluciclovine or 18F-DCFPyL prostate-specific membrane antigen positron emission tomography (PET)/computed tomography (CT). Images of patients with PET-positive PA nodes were imported into the treatment planning system, avid nodes were contoured, and measurements were taken in relation to anatomic landmarks. A contouring guideline that encompassed the location of ≥95% of PET-positive PA nodes was created using descriptive statistics and then validated in an independent data set. Five hundred fifty-nine patients had molecular PET/CT imaging in the development data set (78% 18F-fluciclovine, 22% prostate-specific membrane antigen). Seventy-six patients (14%) had evidence of PA nodal metastasis. We determined that expanding the CTV to 1.8 cm left of the aorta, 1.4 cm right of the inferior vena cava (IVC), 7 mm posterior to the aorta/IVC or to the vertebral body, and superiorly to the T11/T12 vertebral interface, with the anterior border 4 mm anterior to the aorta/IVC and inferior border at the bifurcation of the aorta/IVC, resulted in coverage of ≥95% of PET-positive PA nodes. When the guideline was used in the independent validation data set (246 patients with molecular PET/CT imaging, of whom 31 had PA nodal metastasis), 97% of nodes were encompassed, thereby validating our guideline. We used molecular PET/CT imaging to determine the anatomic locations of PA metastases to develop contouring guidelines for creating a prostate cancer PA CTV. Although the optimal patient selection and clinical benefits of PA radiation therapy remain uncertain, our results will aid in delineating the optimal target when PA radiation therapy is pursued.

Decoding the lncRNAome Across Diverse Cellular Stresses Reveals Core p53-effector Pan-cancer Suppressive lncRNAs.

Identification of pan-cancer suppressive lncRNAs transcriptionally regulated by p53 across different cellular stresses by integrating multilayered high-throughput molecular profiles. This study provides critical new insights into the p53 tumor suppressor by revealing the lncRNAs in the p53 cell-cycle regulatory network and their impact on cancer cell growth and patient survival.

Mendelian randomization indicates a causal contribution of type 2 diabetes to retinal vein occlusion.

Retinal vein occlusion (RVO) is a common retinal vascular disease that can cause severe visual impairment. Many observational studies have shown that type 2 diabetes (T2DM) is associated with RVO, but it remains unknown if the association is causal. The present study aimed to perform Mendelian randomization (MR) analyses to evaluate the causal contribution of genetically predicted T2DM to RVO. We obtained summary-level data from a genome-wide association study meta-analysis including 48,286 cases and 250,671 controls for T2DM and from a genome wide association study of 372 cases and 182,573 controls in the FinnGen project for RVO. To verify the robustness of the results, an independent validation dataset for T2DM (12,931 cases and 57,196 controls) was used. In addition to the main MR analysis using the inverse variance weighted (fixed effect) approach, sensitivity analyses and multivariable MR adjusting for common risk factors of RVO were conducted. Genetically predicted T2DM was found to be causally associated with RVO risk (odds ratio (OR)=2.823, 95% confidence interval (CI): 2.072-3.847, P=4.868×10-11). This association was supported by sensitivity analyses using the weighted median (OR=2.415, 95% CI: 1.411-4.132, P=1.294×10-3), weighted mode (OR=2.370, 95% CI: 1.321-4.252, P=5.159×10-3), maximum likelihood (OR=2.871, 95% CI: 2.100-3.924, P=3.719×10-11), MR-PRESSO (OR=2.823, 95% CI: 2.135-3.733, P=5.150×10-10), and MR-Egger (OR=2.441, 95% CI: 1.149-5.184, P=2.335×10-2) methods. In addition, this association persisted in multivariable MR after accounting for common RVO risk factors (OR=1.748, 95% CI: 1.238-2.467, P=1.490×10-3). The MR analyses using the validation dataset obtained consistent results. This study indicates that genetically predicted T2DM may have a causal contribution to RVO. Future studies are required to elucidate the underlying mechanisms.

O-GlyThr: Prediction of human O-linked threonine glycosites using multi-feature fusion

O-linked glycosylation is one of the most complex post-translational modifications (PTM) of human proteins modulating various cellular metabolic and signaling pathways. Unlike N-glycosylation, the O-glycosylation has non-specific sequence features and unstable glycan core structure, which makes identification of O-glycosites more challenging either by experimental or computational methods. Biochemical experiments to identify O-glycosites in batches are technically and economically demanding. Therefore, development of computation-based methods is greatly warranted. This study constructed a prediction model based on feature fusion for O-glycosites linked to the threonine residues in Homo sapiens. In the training model, we collected and sorted out high-quality human protein data with O-linked threonine glycosites. Seven feature coding methods were fused to represent the sample sequence. By comparison of different algorithms, random forest was selected as the final classifier to construct the classification model. Through 5-fold cross-validation, the proposed model, namely O-GlyThr, performed satisfactorily on both training set (AUC: 0.9308) and independent validation dataset (AUC: 0.9323). Compared with previously published predictors, O-GlyThr achieved the highest ACC of 0.8475 on the independent test dataset. These results demonstrated the high competency of our predictor in identifying O-glycosites on threonine residues. Furthermore, a user-friendly webserver named O-GlyThr (http://cbcb.cdutcm.edu.cn/O-GlyThr/) was developed to assist glycobiologists in the research associated with glycosylation structure and function.

Design of a Convolutional Neural Network as a Deep Learning Tool for the Automatic Classification of Small-Bowel Cleansing in Capsule Endoscopy.

Background and objectives: Capsule endoscopy (CE) is a non-invasive method to inspect the small bowel that, like other enteroscopy methods, requires adequate small-bowel cleansing to obtain conclusive results. Artificial intelligence (AI) algorithms have been seen to offer important benefits in the field of medical imaging over recent years, particularly through the adaptation of convolutional neural networks (CNNs) to achieve more efficient image analysis. Here, we aimed to develop a deep learning model that uses a CNN to automatically classify the quality of intestinal preparation in CE. Methods: A CNN was designed based on 12,950 CE images obtained at two clinical centers in Porto (Portugal). The quality of the intestinal preparation was classified for each image as: excellent, ≥90% of the image surface with visible mucosa; satisfactory, 50-90% of the mucosa visible; and unsatisfactory, <50% of the mucosa visible. The total set of images was divided in an 80:20 ratio to establish training and validation datasets, respectively. The CNN prediction was compared with the classification established by consensus of a group of three experts in CE, currently considered the gold standard to evaluate cleanliness. Subsequently, how the CNN performed in diagnostic terms was evaluated using an independent validation dataset. Results: Among the images obtained, 3633 were designated as unsatisfactory preparation, 6005 satisfactory preparation, and 3312 with excellent preparation. When differentiating the classes of small-bowel preparation, the algorithm developed here achieved an overall accuracy of 92.1%, with a sensitivity of 88.4%, a specificity of 93.6%, a positive predictive value of 88.5%, and a negative predictive value of 93.4%. The area under the curve for the detection of excellent, satisfactory, and unsatisfactory classes was 0.98, 0.95, and 0.99, respectively. Conclusions: A CNN-based tool was developed to automatically classify small-bowel preparation for CE, and it was seen to accurately classify intestinal preparation for CE. The development of such a system could enhance the reproducibility of the scales used for such purposes.

Identification of atypical sleep microarchitecture biomarkers in children with autism spectrum disorder.

Sleep disorders are one of the most frequent comorbidities in children with autism spectrum disorder (ASD). However, the link between neurodevelopmental effects in ASD children with their underlying sleep microarchitecture is not well understood. An improved understanding of etiology of sleep difficulties and identification of sleep-associated biomarkers for children with ASD can improve the accuracy of clinical diagnosis. To investigate whether machine learning models can identify biomarkers for children with ASD based on sleep EEG recordings. Sleep polysomnogram data were obtained from the Nationwide Children' Health (NCH) Sleep DataBank. Children (ages: 8-16 yrs) with 149 autism and 197 age-matched controls without neurodevelopmental diagnosis were selected for analysis. An additional independent age-matched control group (n = 79) selected from the Childhood Adenotonsillectomy Trial (CHAT) was also used to validate the models. Furthermore, an independent smaller NCH cohort of younger infants and toddlers (age: 0.5-3 yr.; 38 autism and 75 controls) was used for additional validation. We computed periodic and non-periodic characteristics from sleep EEG recordings: sleep stages, spectral power, sleep spindle characteristics, and aperiodic signals. Machine learning models including the Logistic Regression (LR) classifier, Support Vector Machine (SVM), and Random Forest (RF) model were trained using these features. We determined the autism class based on the prediction score of the classifier. The area under the receiver operating characteristics curve (AUC), accuracy, sensitivity, and specificity were used to evaluate the model performance. In the NCH study, RF outperformed two other models with a 10-fold cross-validated median AUC of 0.95 (interquartile range [IQR], [0.93, 0.98]). The LR and SVM models performed comparably across multiple metrics, with median AUC 0.80 [0.78, 0.85] and 0.83 [0.79, 0.87], respectively. In the CHAT study, three tested models have comparable AUC results: LR: 0.83 [0.76, 0.92], SVM: 0.87 [0.75, 1.00], and RF: 0.85 [0.75, 1.00]. Sleep spindle density, amplitude, spindle-slow oscillation (SSO) coupling, aperiodic signal's spectral slope and intercept, as well as the percentage of REM sleep were found to be key discriminative features in the predictive models. Our results suggest that integration of EEG feature engineering and machine learning can identify sleep-based biomarkers for ASD children and produce good generalization in independent validation datasets. Microstructural EEG alterations may help reveal underlying pathophysiological mechanisms of autism that alter sleep quality and behaviors. Machine learning analysis may reveal new insight into the etiology and treatment of sleep difficulties in autism.

Accurate prediction of pan-cancer types using machine learning with minimal number of DNA methylation sites.

DNA methylation analysis has been applied to determine the primary site of cancer; however, robust and accurate prediction of cancer types with a minimum number of sites is still a significant scientific challenge. To build an accurate and robust cancer type prediction tool with a minimum number of DNA methylation sites, we internally benchmarked different DNA methylation site selection and ranking procedures, as well as different classification models. We used The Cancer Genome Atlas dataset (26 cancer types with 8296 samples) to train and test models and used an independent dataset (17 cancer types with 2738 samples) for model validation. A deep neural network model using a combined feature selection procedure (named MethyDeep) can predict 26 cancer types using 30 methylation sites with superior performance compared with the known methods for both primary and metastatic cancers in independent validation datasets. In conclusion, MethyDeep is an accurate and robust cancer type predictor with the minimum number of DNA methylation sites; it could help the cost-effective clarification of cancer of unknown primary patients and the liquid biopsy-based early screening of cancers.

Abstract 5359: Translating state-of-the-art deep learning predictions of treatment efficacy to clinical practice

Abstract Despite the clinical success of immune checkpoint blockade therapies, in many patients the disease becomes unresponsive or resistant to treatment. Attempts to predict treatment efficacy and identify determinants of response have suffered from limited accuracy. Here, we introduce a state-of-the-art deep learning method that better captures the complex relationships among the host, the immune system, and tumor biology. Our method, an explainable deep learning framework, is based on transformer architecture that combines data with different feature sets and clinical endpoints for survival prediction or classification. The framework includes(1) a new loss function based on Harrell’s concordance-index, (2) an explainability module providing importance score between features based on mutual contribution to predictions, and (3) a transfer learning strategy that enables leveraging diverse clinical data sets in the public or private domain. Using seven data sets comprising more than 140,000 patients from immunooncologic (IO), targeted, and chemotherapy treatments, our method consistently outperformed other methods previously described in the literature, including CoxPH and random survival forest (RF). For example, our framework achieved a concordance index of 0.66 (±0.04) vs. 0.60 (±0.04) of the second-best method (RF) on MYSTIC IO arms using only clinical data. Utilizing our explainability module, we identified key features driving response prediction that are consistent with those in previous publications. For example, in the Chowell et al. data set, we identified albumin, neutrophil-to-lymphocyte ratio, prior-chemo-treatment, and tumor mutation burden as the most important features. We also identified in sparse mutation calls of 469 genes from the Samstein et al. data set, functional modules of several genes, each with a strong predictive power. For example, the functional module with the highest hazard ratio (HR) (1.43) was the gene pair KEAP1 and STK11, which are well-established drivers of resistance to multiple therapies. We further validated these results across multiple data sets: The Cancer Genome Atlas HR = 1.23), GENIE/anti-PD-L1 (HR = 1.06), and MYSTIC/anti-PD-L1/CTLA-4 (HR = 1.39). Another example is a functional module related to adaptive immunity, comprising the genes AKT2, BTK, CDC73, HLA-B, IKBKE, INPPL1, RFWD2, TRAF2, and WHSC1. This module stratified patients with an HR of 0.58 within the Samstein et al. data set and 0.42 within the independent validation data set. We describe a new deep learning method with state-of-the-art performance in survival prediction and the potential to uncover biological and clinical insights related to disease response and resistance. Our framework simplifies the process of translating complex AI models to clinical practice and may accelerate the identification of targetable drivers of resistance and response. Citation Format: Gustavo Arango-Argoty, Etai Jacob. Translating state-of-the-art deep learning predictions of treatment efficacy to clinical practice. [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2023; Part 1 (Regular and Invited Abstracts); 2023 Apr 14-19; Orlando, FL. Philadelphia (PA): AACR; Cancer Res 2023;83(7_Suppl):Abstract nr 5359.

Abstract 1478: Broad downregulation of normal-specific super-enhancer associated genes in lung adenocarcinoma

Abstract Background Lung adenocarcinoma (LUAD) is the most common histologic subtype of lung cancer and a highly heterogeneous disease whose mechanism is not fully uncovered. Recently, Super-enhancers (SEs), tissue-specific clusters of enhancer elements, were found to drive cell identity genes and play an important role in tumorigenesis. However, current studies focus on SE in lung tumors but largely neglect those specific in normal lung tissues, which also provide critical data for understanding transformation of normal cells into cancer cells. For the first time, this study characterized normal-specific SE-associated genes (SEGs) in LUAD by dysregulation in tumor, prognostic significance and tumor microenvironment (TME) interactions via a comprehensive analysis of 22 bulk and single-cell transcriptomic datasets. Methods We analyzed TCGA datasets of 10 cancer types, and additional datasets comprising 4 datasets of LUAD tumor and normal paired samples, 5 datasets of LUAD tumors with survival information, and 3 single-cell transcriptomic datasets. For each dataset, standard methods were used for data preprocessing and normalization according to its generation platforms (e.g., microarray, bulk RNA-seq and scRNA-seq). For downstream analyses, paired t-test and linear model were used to identify dysregulated SEGs; Cox Proportional Hazards Regression was used to evaluate their prognostic significance and LASSO was used to select features for building multivariate prognostic models; bulk data deconvolution and correlation analysis were applied to assess the association of TME cell composition and SEG expression; Gene set enrichment analysis was applied for functional analysis. Results We found normal-lung SEGs were significantly disproportionally downregulated in LUAD tumors (Chi-squared test, p &amp;lt; 0.01), and associated with poor prognosis of patients. The patterns were consistently found in discovery (TCGA) and independent validation datasets. Interestingly, the effect appears specific to lung as such downregulation of normal-lung SEGs was also found in lung squamous cell carcinoma but not observed in most of other 8 cancer types. The downregulated SEGs were significantly involved in cell differentiation, apoptosis, proliferation, and stem cell pluripotency, and were enriched in tumor suppressors. Also, their downregulation correlate with reduced endothelial cells and lower immune activity with reduced B and CD8 T cells and inhibited immune target genes. Conclusions Our study characterized the genes associated to normal-specific SEs in LUAD, shedding sheds light on their role in tumorigenesis of lung cancers. Their broad downregulation indicates a significant effect of SE aberrancy in the transformation of normal cell into cancer cells with a modulation of immune microenvironment; further functional studies are needed to elucidate the underlying mechanism. Citation Format: Guoshuai Cai, Yihan Liu, Xiang Cui, Christopher Amos, Feifei Xiao. Broad downregulation of normal-specific super-enhancer associated genes in lung adenocarcinoma [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2023; Part 1 (Regular and Invited Abstracts); 2023 Apr 14-19; Orlando, FL. Philadelphia (PA): AACR; Cancer Res 2023;83(7_Suppl):Abstract nr 1478.

PD22-07 A MULTICENTER STUDY PRESENTING AN EXPLAINABLE AI APPROACH TO AUTOMATIC SEGMENTATION OF RENAL TUMORS ON COMPUTED TOMOGRAPHY

You have accessJournal of UrologyCME1 Apr 2023PD22-07 A MULTICENTER STUDY PRESENTING AN EXPLAINABLE AI APPROACH TO AUTOMATIC SEGMENTATION OF RENAL TUMORS ON COMPUTED TOMOGRAPHY Annemarie Uhlig, Sophie Bachanek, Paul Romario Würzberg, Manuel Nietert, Lutz Trojan, Joachim Lotz, and Johannes Uhlig Annemarie UhligAnnemarie Uhlig More articles by this author , Sophie BachanekSophie Bachanek More articles by this author , Paul Romario WürzbergPaul Romario Würzberg More articles by this author , Manuel NietertManuel Nietert More articles by this author , Lutz TrojanLutz Trojan More articles by this author , Joachim LotzJoachim Lotz More articles by this author , and Johannes UhligJohannes Uhlig More articles by this author View All Author Informationhttps://doi.org/10.1097/JU.0000000000003295.07AboutPDF ToolsAdd to favoritesDownload CitationsTrack CitationsPermissionsReprints ShareFacebookLinked InTwitterEmail Abstract INTRODUCTION AND OBJECTIVE: Automatic segmentation of renal tumors is a cornerstone for Artificial Intelligence (AI) based tumor diagnostics. This study presents an automatic segmentation algorithm and visualization method using CT-studies from renal tumor patients recruited in a multicenter setting. METHODS: Renal tumor patients diagnosed between 2018 and 2021 were retrospectively assessed for this study. Inclusion criteria were CT-imaging of renal tumors in corticomedullary (CM) or nephrogenic (NG) contrast media phase. Patients aged <18years, and those with cystic or infiltrative renal tumors (i.e. lymphomas) were excluded. Manual segmentation of the renal tumors was performed by a GU-radiologist on all axial CT slices. A convolutional neural network (UNET) was trained based on the radiologists´ manual segmentations. In an independent validation dataset, the accuracy of the UNETs predictions of renal tumor contours was compared to the references standard of manual segmentations and quantified using the DICE score. RESULTS: A total of n=394/ n=350 patients with renal tumors imaged in CM /NG phase, respectively, were included (median age 66y; 35% female; median tumor diameter 5.4cm). CT-studies from >20 radiological imaging centers were included with different imaging protocols and slice thickness. The UNET was trained on n=316 CM and n=294 NG contrast phase patients (n=7019 / n=6859 separate CT images). In the independent validation dataset (n=78 / n=56 patients with 1713 / 1298 CT images), the UNET achieved a DICE score of 0.88 and 0.90 for the corticomedullary and nephrogenic CM phase, respectively. The UNET predictions were visualized using a tile-based approach with color-coding and contour-lines that could be overlaid on CT-images to depict varying levels of prediction confidence. CONCLUSIONS: A UNET AI-approach yields a robust automatic delineation of renal tumors on CT-images acquired in clinical routine, irrespective of the contrast media phase. Using color-coding and contour-lines that could be overlaid on original CT-images provides an explainable approach to the UNETs predictions and might improve AI-acceptance in clinical practice. Source of Funding: This study received funding by the 2022 ESR Research Seed Grant. The 2022 ESR Research Seed Grants were kindly supported by an unrestricted, non-exclusive grant from GE Healthcare © 2023 by American Urological Association Education and Research, Inc.FiguresReferencesRelatedDetails Volume 209Issue Supplement 4April 2023Page: e667 Advertisement Copyright & Permissions© 2023 by American Urological Association Education and Research, Inc.MetricsAuthor Information Annemarie Uhlig More articles by this author Sophie Bachanek More articles by this author Paul Romario Würzberg More articles by this author Manuel Nietert More articles by this author Lutz Trojan More articles by this author Joachim Lotz More articles by this author Johannes Uhlig More articles by this author Expand All Advertisement PDF downloadLoading ...