Average AUROC Research Articles

GenHPF: General Healthcare Predictive Framework for Multi-task Multi-source Learning.

Despite the remarkable progress in the development of predictive models for healthcare, applying these algorithms on a large scale has been challenging. Algorithms trained on a particular task, based on specific data formats available in a set of medical records, tend to not generalize well to other tasks or databases in which the data fields may differ. To address this challenge, we propose General Healthcare Predictive Framework (GenHPF), which is applicable to any EHR with minimal preprocessing for multiple prediction tasks. GenHPF resolves heterogeneity in medical codes and schemas by converting EHRs into a hierarchical textual representation while incorporating as many features as possible. To evaluate the efficacy of GenHPF, we conduct multi-task learning experiments with single-source and multi-source settings, on three publicly available EHR datasets with different schemas for 12 clinically meaningful prediction tasks. Our framework significantly outperforms baseline models that utilize domain knowledge in multi-source learning, improving average AUROC by 1.2%P in pooled learning and 2.6%P in transfer learning while also showing comparable results when trained on a single EHR dataset. Furthermore, we demonstrate that self-supervised pretraining using multi-source datasets is effective when combined with GenHPF, resulting in a 0.6 pretraining. By eliminating the need for preprocessing and feature engineering, we believe that this work offers a solid framework for multi-task and multi-source learning that can be leveraged to speed up the scaling and usage of predictive algorithms in healthcare.1.

MT-EpiPred: Multitask Learning for Prediction of Small-Molecule Epigenetic Modulators.

Epigenetic modulators play an increasingly crucial role in the treatment of various diseases. In this case, it is imperative to systematically investigate the activity of these agents and understand their influence on the entire epigenetic regulatory network rather than solely concentrate on individual targets. This work introduces MT-EpiPred, a multitask learning method capable of predicting the activity of compounds against 78 epigenetic targets. MT-EpiPred demonstrated outstanding performance, boasting an average auROC of 0.915 and the ability to handle few-shot targets. In comparison to the existing method, MT-EpiPred not only expands the target pool but also achieves superior predictive performance with the same data set. MT-EpiPred was then applied to predict the epigenetic target of a newly synthesized compound (1), where the molecular target was unknown. The method identified KDM4D as a potential target, which was subsequently validated through an in vitro enzyme inhibition assay, revealing an IC50 of 4.8 μM. The MT-EpiPred method has been implemented in the web server MT-EpiPred (http://epipred.com), providing free accessibility. In summary, this work presents a convenient and accurate tool for discovering novel small-molecule epigenetic modulators, particularly in the development of selective inhibitors and evaluating the impact of these inhibitors over a broad epigenetic network.

A deep learning model for predicting risks of crop pests and diseases from sequential environmental data

Crop pests reduce productivity, so managing them through early detection and prevention is essential. Data from various modalities are being used to predict crop diseases by applying machine learning methodology. In particular, because growth environment data is relatively easy to obtain, many attempts are made to predict pests and diseases using it. In this paper, we propose a model that predicts diseases through previous growth environment information of crops, including air temperature, relative humidity, dew point, and CO2 concentration, using deep learning techniques. Using large-scale public data on crops of strawberry, pepper, grape, tomato, and paprika, we showed the model can predict the risk score of crop pests and diseases. It showed high predictive performance with an average AUROC of 0.917, and based on the predicted results, it can help prevent pests or post-processing. This environmental data-based crop disease prediction model and learning framework are expected to be universally applicable to various facilities and crops for disease/pest prevention.

A Multimodal Deep Learning Framework for Predicting PPI-Modulator Interactions.

Protein-protein interactions (PPIs) are essential for various biological processes and diseases. However, most existing computational methods for identifying PPI modulators require either target structure or reference modulators, which restricts their applicability to novel PPI targets. To address this challenge, we propose MultiPPIMI, a sequence-based deep learning framework that predicts the interaction between any given PPI target and modulator. MultiPPIMI integrates multimodal representations of PPI targets and modulators and uses a bilinear attention network to capture intermolecular interactions. Experimental results on our curated benchmark data set show that MultiPPIMI achieves an average AUROC of 0.837 in three cold-start scenarios and an AUROC of 0.994 in the random-split scenario. Furthermore, the case study shows that MultiPPIMI can assist molecular docking simulations in screening inhibitors of Keap1/Nrf2 PPI interactions. We believe that the proposed method provides a promising way to screen PPI-targeted modulators.

SURG-04. FAST DETECTION OF GLIOMA INFILTRATION USING LABEL-FREE OPTICAL MICROSCOPY AND DEEP NEURAL NETWORKS

Abstract INTRODUCTION Fast and accurate detection of tumor infiltration is a central challenge in the comprehensive management of diffuse glioma patients. Despite our best real-time tumor detection strategies, such as intraoperative MRI and fluorescence-guided surgery, dense and safely resectable tumor infiltration remains at the surgical margin in over 85% of diffuse glioma patients. Patterns of treatment failure demonstrate that glioma recurrence occurs at the surgical margin in the majority of patients due to residual tumor burden. Objective Here, we present FastGlioma, an artificial intelligence (AI)-based intraoperative diagnostic system for fast (< 10 seconds) detection of diffuse glioma infiltration at microscopic resolution without the need for tumor-specific markers. Method FastGlioma is trained using large-scale, self-supervised visual feature learning on stimulated Raman histology (SRH), a rapid, label-free, optical microscopy technique, and a GPT-style whole slide SRH tumor infiltration scoring method. We pushed the performance limits of FastGlioma by investigating the trade-off between imaging speed versus accuracy for microscopic tumor detection. RESULTS In a large prospective external testing cohort of diffuse glioma patients (180 patients, 935 surgical margin specimens) who underwent intraoperative SRH imaging, we demonstrate that FastGlioma was able to detect and quantify microscopic tumor infiltration with an average AUROC of 92.1 +/- 0.0%, performing on par with three expert neuropathologists on the same task. FastGlioma is over 10X faster than previous SRH-based methods and 100X faster than conventional H&E histology. Finally, we demonstrate that FastGlioma was > 20% more accurate at detecting dense tumor infiltration when compared to radiologic features or 5-aminolevulinic fluorescence (98.0% versus 77.8% accuracy) for both IDH-mutant diffuse gliomas and IDH-wildtype glioblastomas. CONCLUSION Our results demonstrate how intraoperative optical imaging and deep neural networks can achieve fast and accurate tumor detection, unlocking the role of AI in improving the surgical management of brain tumor patients.

Unsupervised anomaly detection in images using attentional normalizing flows

Image-based anomaly detection has been widely used in practice, but it is still a challenging task due to the irregularity of anomalies. Existing representation-based methods have achieved high accuracy metrics in image-based anomaly detection, but they are weak in capturing anomalous regions, resulting in small inter-class variance between the latent distributions of normal and anomalous data. Furthermore, most current neural networks have low parameter efficiency, which limits their inference performance. To address these issues, we propose a novel normalizing flow called AFlow. Our network represents multi-scale features with multi-scale convolutions instead of inputs of different resolutions, significantly reducing the network parameters and improving the fine-grained representation in features. Then, an attention module is introduced into the normalizing flow to further select the multi-scale features along feature channels, making the network focus on relevant regions with more object details to refine the objective of maximum likelihood estimation. Experiments on two public datasets show that the proposed method can achieve more accurate prediction results and higher parameter efficiency than previous detection models: the average AUROC of the proposed method reaches 99.75% and 97.2% on MVTec AD and BTAD datasets, respectively.

Developing A Baseline Metabolomic Signature Associated with COVID-19 Severity: Insights from Prospective Trials Encompassing 13 U.S. Centers.

Metabolic disease is a significant risk factor for severe COVID-19 infection, but the contributing pathways are not yet fully elucidated. Using data from two randomized controlled trials across 13 U.S. academic centers, our goal was to characterize metabolic features that predict severe COVID-19 and define a novel baseline metabolomic signature. Individuals (n = 133) were dichotomized as having mild or moderate/severe COVID-19 disease based on the WHO ordinal scale. Blood samples were analyzed using the Biocrates platform, providing 630 targeted metabolites for analysis. Resampling techniques and machine learning models were used to determine metabolomic features associated with severe disease. Ingenuity Pathway Analysis (IPA) was used for functional enrichment analysis. To aid in clinical decision making, we created baseline metabolomics signatures of low-correlated molecules. Multivariable logistic regression models were fit to associate these signatures with severe disease on training data. A three-metabolite signature, lysophosphatidylcholine a C17:0, dihydroceramide (d18:0/24:1), and triacylglyceride (20:4_36:4), resulted in the best discrimination performance with an average test AUROC of 0.978 and F1 score of 0.942. Pathways related to amino acids were significantly enriched from the IPA analyses, and the mitogen-activated protein kinase kinase 5 (MAP2K5) was differentially activated between groups. In conclusion, metabolites related to lipid metabolism efficiently discriminated between mild vs. moderate/severe disease. SDMA and GABA demonstrated the potential to discriminate between these two groups as well. The mitogen-activated protein kinase kinase 5 (MAP2K5) regulator is differentially activated between groups, suggesting further investigation as a potential therapeutic pathway.

M5U-GEPred: prediction of RNA 5-methyluridine sites based on sequence-derived and graph embedding features.

5-Methyluridine (m5U) is one of the most common post-transcriptional RNA modifications, which is involved in a variety of important biological processes and disease development. The precise identification of the m5U sites allows for a better understanding of the biological processes of RNA and contributes to the discovery of new RNA functional and therapeutic targets. Here, we present m5U-GEPred, a prediction framework, to combine sequence characteristics and graph embedding-based information for m5U identification. The graph embedding approach was introduced to extract the global information of training data that complemented the local information represented by conventional sequence features, thereby enhancing the prediction performance of m5U identification. m5U-GEPred outperformed the state-of-the-art m5U predictors built on two independent species, with an average AUROC of 0.984 and 0.985 tested on human and yeast transcriptomes, respectively. To further validate the performance of our newly proposed framework, the experimentally validated m5U sites identified from Oxford Nanopore Technology (ONT) were collected as independent testing data, and in this project, m5U-GEPred achieved reasonable prediction performance with ACC of 91.84%. We hope that m5U-GEPred should make a useful computational alternative for m5U identification.

Predictive models of long COVID

The cause and symptoms of long COVID are poorly understood. It is challenging to predict whether a given COVID-19 patient will develop long COVID in the future. We used electronic health record (EHR) data from the National COVID Cohort Collaborative to predict the incidence of long COVID. We trained two machine learning (ML) models - logistic regression (LR) and random forest (RF). Features used to train predictors included symptoms and drugs ordered during acute infection, measures of COVID-19 treatment, pre-COVID comorbidities, and demographic information. We assigned the 'long COVID' label to patients diagnosed with the U09.9 ICD10-CM code. The cohorts included patients with (a) EHRs reported from data partners using U09.9 ICD10-CM code and (b) at least one EHR in each feature category. We analysed three cohorts: all patients (n=2,190,579; diagnosed with long COVID=17,036), inpatients (149,319; 3,295), and outpatients (2,041,260; 13,741). LR and RF models yielded median AUROC of 0.76 and 0.75, respectively. Ablation study revealed that drugs had the highest influence on the prediction task. The SHAP method identified age, gender, cough, fatigue, albuterol, obesity, diabetes, and chronic lung disease as explanatory features. Models trained on data from one N3C partner and tested on data from the other partners had average AUROC of 0.75. ML-based classification using EHR information from the acute infection period is effective in predicting long COVID. SHAP methods identified important features for prediction. Cross-site analysis demonstrated the generalizability of the proposed methodology. NCATS U24 TR002306, NCATS UL1 TR003015, Axle Informatics Subcontract: NCATS-P00438-B, NIH/NIDDK/OD, PSR2015-1720GVALE_01, G43C22001320007, and Director, Office of Science, Office of Basic Energy Sciences of the U.S. Department of Energy Contract No. DE-AC02-05CH11231.

Merit: multi-level graph embedding refinement framework for large-scale graph

The development of the Internet and big data has led to the emergence of graphs as an important data representation structure in various real-world scenarios. However, as data size increases, computational complexity and memory requirements pose significant challenges for graph embedding. To address this challenge, this paper proposes a multilevel embedding refinement framework (MERIT) based on large-scale graphs, using spectral distance-constrained graph coarsening algorithms and an improved graph convolutional neural network model that addresses the over-smoothing problem by incorporating initial values and identity mapping. Experimental results on large-scale datasets demonstrate the effectiveness of MERIT, with an average AUROC score 8% higher than other baseline methods. Moreover, in a node classification task on a large-scale graph with 126,825 nodes and 22,412,658 edges, the framework improves embedding quality while enhancing the runtime by 25 times. The experimental findings highlight the superior efficiency and accuracy of the proposed approach compared to other graph embedding methods.

Analysis of neuropsychological and neuroradiological features for diagnosis of Alzheimer's disease and mild cognitive impairment

BackgroundAge-related neurodegenerative diseases are constantly increasing with prediction that in 2050 over 60 % of population will suffer from some level of cognitive impairment. A cure for the Alzheimer’s disease (AD) does not exist, so early diagnosis is of a great importance. Machine learning techniques can help in early diagnosis with deep medical data processing, disease understanding, intervention analysis and knowledge discovery for achieving better medical decision making. MethodsIn this paper, we analyze the dataset consisting of 90 individuals and 482 input features. We investigate the achieved AD prediction performances using seven classifiers and five feature selection algorithms. We pay special focus on analyzing performance by utilizing only a subset of best ranked attributes to establish the minimum amount of input features that ensure acceptable performance. We also investigate the significance of neuropsychological (NP) and neuroradiological (NR) attributes for the AD diagnosis. ResultsThe accuracy for the whole set of attributes ranged between 66.22 % and 81.00 %, and the weighted average AUROC was between 76.3 % and 95.0 %. The best results were achieved by the naive Bayes classifier and the Relief feature selection algorithm. Additionally, Support Vector Machines classifier shows the most stable results since it depends the least on the feature selection algorithm which is used. As the main result of this paper, we compare the performance of models trained with automatically selected features to models trained with hand-selected features performed by medical experts (NP and NR features). ConclusionsThe results reveal that unlike the NR attributes, the NP attributes achieve a good performance that is comparable to the full set of attributes, which suggests that they possess a high predictive power for AD diagnosis.

Predicting successful placements for youth in child welfare with machine learning

Out-of-home placement decisions have extremely high stakes for the present and future well-being of children in care because some placement types, and multiple placements, are associated with poor outcomes. We propose that a clinical decision support system (CDSS) using existing data about children and their previous placement success could inform future placement decision-making for their peers. The objective of this study was to test the feasibility of developing machine learning models to predict the best level of care placement (i.e., the placement with the highest likelihood of doing well in treatment) based on each youth’s behavioral health needs and characteristics. We developed machine learning models to predict the probability of each youth’s treatment success in psychiatric residential care (i.e., Psychiatric Residential Treatment Facility [PRTF]) versus any other placement (AUROCs > 0.70) using data collected in standard care at a behavioral health organization. Placement recommendations based on these machine learning models distinguished between youth who did well in residential care versus non-residential care (e.g., 80% of those who received care in the recommended setting with the highest predicted likelihood of success had above average risk-adjusted outcomes). Then we developed and validated machine learning models to predict the probability of each youth’s treatment success across specific placement types in a state-wide system, achieving an average AUROC score of >0.75. Machine learning models based on risk-adjusted behavioral health and functional data show promise in predicting positive placement outcomes and informing future placement decisions for youth in care. Related ethical considerations are discussed.

MDAlmc: A Novel Low-rank Matrix Completion Model for MiRNADisease Association Prediction by Integrating Similarities among MiRNAs and Diseases.

The importance of microRNAs (miRNAs) has been emphasized by an increasing number of studies, and it is well-known that miRNA dysregulation is associated with a variety of complex diseases. Revealing the associations between miRNAs and diseases are essential to disease prevention, diagnosis, and treatment. However, traditional experimental methods in validating the roles of miRNAs in diseases could be very expensive, labor-intensive and time-consuming. Thus, there is a growing interest in predicting miRNA-disease associations by computational methods. Though many computational methods are in this category, their prediction accuracy needs further improvement for downstream experimental validation. In this study, we proposed a novel model to predict miRNA-disease associations by low-rank matrix completion (MDAlmc) integrating miRNA functional similarity, disease semantic similarity, and known miRNA-disease associations. In the 5-fold cross-validation, MDAlmc achieved an average AUROC of 0.8709 and AUPRC of 0.4172, better than those of previous models. Among the case studies of three important human diseases, the top 50 predicted miRNAs of 96% (breast tumors), 98% (lung tumors), and 90% (ovarian tumors) have been confirmed by previous literatures. And the unconfirmed miRNAs were also validated to be potential disease-associated miRNAs. MDAlmc is a valuable computational resource for miRNA-disease association prediction.

Expanding potential targets of herbal chemicals by node2vec based on herb–drug interactions

BackgroundThe identification of chemical–target interaction is key to pharmaceutical research and development, but the unclear materials basis and complex mechanisms of traditional medicine (TM) make it difficult, especially for low-content chemicals which are hard to test in experiments. In this research, we aim to apply the node2vec algorithm in the context of drug-herb interactions for expanding potential targets and taking advantage of molecular docking and experiments for verification.MethodsRegarding the widely reported risks between cardiovascular drugs and herbs, Salvia miltiorrhiza (Danshen, DS) and Ligusticum chuanxiong (Chuanxiong, CX), which are widely used in the treatment of cardiovascular disease (CVD), and approved drugs for CVD form the new dataset as an example. Three data groups DS-drug, CX-drug, and DS-CX-drug were applied to serve as the context of drug-herb interactions for link prediction. Three types of datasets were set under three groups, containing information from chemical-target connection (CTC), chemical-chemical connection (CCC) and protein–protein interaction (PPI) in increasing steps. Five algorithms, including node2vec, were applied as comparisons. Molecular docking and pharmacological experiments were used for verification.ResultsNode2vec represented the best performance with average AUROC and AP values of 0.91 on the datasets “CTC, CCC, PPI”. Targets of 32 herbal chemicals were identified within 43 predicted edges of herbal chemicals and drug targets. Among them, 11 potential chemical-drug target interactions showed better binding affinity by molecular docking. Further pharmacological experiments indicated caffeic acid increased the thermal stability of the protein GGT1 and ligustilide and low-content chemical neocryptotanshinone induced mRNA change of FGF2 and MTNR1A, respectively.ConclusionsThe analytical framework and methods established in the study provide an important reference for researchers in discovering herb–drug interactions, alerting clinical risks, and understanding complex mechanisms of TM.

TEINet: a deep learning framework for prediction of TCR–epitope binding specificity

The adaptive immune response to foreign antigens is initiated by T-cell receptor (TCR) recognition on the antigens. Recent experimental advances have enabled the generation of a large amount of TCR data and their cognate antigenic targets, allowing machine learning models to predict the binding specificity of TCRs. In this work, we present TEINet, a deep learning framework that utilizes transfer learning to address this prediction problem. TEINet employs two separately pretrained encoders to transform TCR and epitope sequences into numerical vectors, which are subsequently fed into a fully connected neural network to predict their binding specificities. A major challenge for binding specificity prediction is the lack of a unified approach to sampling negative data. Here, we first assess the current negative sampling approaches comprehensively and suggest that the Unified Epitope is the most suitable one. Subsequently, we compare TEINet with three baseline methods and observe that TEINet achieves an average AUROC of 0.760, which outperforms baseline methods by 6.4-26%. Furthermore, we investigate the impacts of the pretraining step and notice that excessive pretraining may lower its transferability to the final prediction task. Our results and analysis show that TEINet can make an accurate prediction using only the TCR sequence (CDR3$\beta $) and the epitope sequence, providing novel insights to understand the interactions between TCRs and epitopes.

Gorge: graph convolutional networks on heterogeneous multi-relational graphs for polypharmacy side effect prediction

Determining the side effects of multidrug combinations is a very important issue in drug risk studies. However, designing clinical trials to determine frequencies is often time-consuming and expensive, and previous work has often been limited to using the target protein of a drug without screening. Although this alleviates the sparsity of the raw data to some extent, blindly introducing proteins as auxiliary information can lead to a large amount of noisy information being added, which in turn leads to less efficient models. For this reason, we propose a new method called Gorge (graph convolutional networks on heterogeneous multi-relational graphs for polypharmacy side effect prediction). Specifically, we designed two protein auxiliary pathways directly related to drugs and combined these two auxiliary pathways with a multi-relational graph of drug side effects, which both alleviates data sparsity and filters noisy data. Then, we introduce a query-aware attention mechanism that generates different attention pathways for drug entities based on different drug pairs, fine-grained to determine the extent of information delivery. Finally, we output the exact frequency of drug side effects occurring through a tensor factorization decoder, in contrast to most existing methods that can only predict the presence or association of side effects, but not their frequency. We found that Gorge achieves excellent performance on real-world datasets (average AUROC of 0.822 and average AUPR of 0.775), outperforming existing methods. Further examination provides literature evidence for highly ranked predictions.

Guild-Level Microbiome Signature Associated with COVID-19 Severity and Prognosis.

Coronavirus disease 2019 (COVID-19) severity has been associated with alterations of the gut microbiota. However, the relationship between gut microbiome alterations and COVID-19 prognosis remains elusive. Here, we performed a genome-resolved metagenomic analysis on fecal samples from 300 in-hospital COVID-19 patients, collected at the time of admission. Among the 2,568 high quality metagenome-assembled genomes (HQMAGs), redundancy analysis identified 33 HQMAGs which showed differential distribution among mild, moderate, and severe/critical severity groups. Co-abundance network analysis determined that the 33 HQMAGs were organized as two competing guilds. Guild 1 harbored more genes for short-chain fatty acid biosynthesis, and fewer genes for virulence and antibiotic resistance, compared with Guild 2. Based on average abundance difference between the two guilds, the guild-level microbiome index (GMI) classified patients from different severity groups (average AUROC [area under the receiver operating curve] = 0.83). Moreover, age-adjusted partial Spearman's correlation showed that GMIs at admission were correlated with 8 clinical parameters, which are predictors for COVID-19 prognosis, on day 7 in hospital. In addition, GMI at admission was associated with death/discharge outcome of the critical patients. We further validated that GMI was able to consistently classify patients with different COVID-19 symptom severities in different countries and differentiated COVID-19 patients from healthy subjects and pneumonia controls in four independent data sets. Thus, this genome-based guild-level signature may facilitate early identification of hospitalized COVID-19 patients with high risk of more severe outcomes at time of admission. IMPORTANCE Previous reports on the associations between COVID-19 and gut microbiome have been constrained by taxonomic-level analysis and overlook the interaction between microbes. By applying a genome-resolved, reference-free, guild-based metagenomic analysis, we demonstrated that the relationship between gut microbiota and COVID-19 is genome-specific instead of taxon-specific or even species-specific. Moreover, the COVID-19-associated genomes were not independent but formed two competing guilds, with Guild 1 potentially beneficial and Guild 2 potentially more detrimental to the host based on comparative genomic analysis. The dominance of Guild 2 over Guild 1 at time of admission was associated with hospitalized COVID-19 patients at high risk for more severe outcomes. Moreover, the guild-level microbiome signature is not only correlated with the symptom severity of COVID-19 patients, but also differentiates COVID-19 patients from pneumonia controls and healthy subjects across different studies. Here, we showed the possibility of using genome-resolved and guild-level microbiome signatures to identify hospitalized COVID-19 patients with a high risk of more severe outcomes at the time of admission.

P209 Delayed IBD diagnosis increases risk of requiring biologics in first 4-years - an Inception cohort study

Abstract Background Delays in diagnosis can be due to patient and health-system related factors. The impact of such delays in Inflammatory Bowel Diseases (IBD), particularly on biological therapies, continues to be of interest. The aim of the study was to investigate if delays could predict need for biologics and report on related adverse outcomes (AOs). Methods New patients referred for suspected IBD to a single tertiary care centre between Jan 2013 to Dec 2017 were identified using EMR. For this study, delay-types and cut-off times for each type were set based on best average hospital waiting times. The reasons for delays until start of treatment were recorded; data on the start of biologics, duration, and reasons for switch (if required) were recorded. Data was analysed using multiple Pearson correlations and Cox proportional Hazard model to determine whether any delay-types could predict need for biologics and compare the risk between patients with or without delay. Results 105 patients were identified using stringent criteria (M=58; median age=32y) with a median follow-up of 55 months. 65, 27 and 13 patients had a final diagnosis of Ulcerative colitis, Crohn’s disease, and Unclassified colitis respectively, and were analysed collectively. A model that combined delay in seeking medical attention, and time taken from specialty review to index endoscopy exhibited the best predictive power with average AUROC of 0.67. When individual delay-types were analysed, it was noted that most delay-types increased the risk of biologic use: a) delay in seeking medical attention [OR-1.08(1.03 1.17)] b) delay from GP referral to specialty review [OR-1.10(1.0 1.23)] c) delay from specialty review to index endoscopy [OR-1.09(1.01 1.21)] On further analysis, there was a statistically significant increase noted in the risk of requiring biologics at any time during the follow-up period. Fig 1A illustrates survival without biologics when there is a delay of >2 weeks in seeking medical attention (p=0.042). Figure 1B illustrates survival without biologics when there is delay of >6 weeks from GP referral to index endoscopy (p=0.021). Figure 1A- Patient-related delay of >2weeks and risk of needing biologics (p=0.042) Figure 1B- Health-system related delay of >6weeks and risk of needing biologics (p=0.021) Conclusion The risk of requiring biologic therapy is higher in those patients who have faced any type of delay in their diagnosis of IBD. A patient-related delay of >2weeks and cumulative health-system related delay of >6weeks increase the risk of requiring biologic therapy at any time during their follow-up. Both primary and secondary care must work to reduce any delays as this has long-term implications for patients and economic implications for health care providers.

Assess and validate predictive performance of models for in-hospital mortality in COVID-19 patients: A retrospective cohort study in the Netherlands comparing the value of registry data with high-granular electronic health records

PurposeTo assess, validate and compare the predictive performance of models for in-hospital mortality of COVID-19 patients admitted to the intensive care unit (ICU) over two different waves of infections. Our models were built with high-granular Electronic Health Records (EHR) data versus less-granular registry data. MethodsObservational study of all COVID-19 patients admitted to 19 Dutch ICUs participating in both the national quality registry National Intensive Care Evaluation (NICE) and the EHR-based Dutch Data Warehouse (hereafter EHR). Multiple models were developed on data from the first 24 h of ICU admissions from February to June 2020 (first COVID-19 wave) and validated on prospective patients admitted to the same ICUs between July and December 2020 (second COVID-19 wave). We assessed model discrimination, calibration, and the degree of relatedness between development and validation population. Coefficients were used to identify relevant risk factors. ResultsA total of 1533 patients from the EHR and 1563 from the registry were included. With high granular EHR data, the average AUROC was 0.69 (standard deviation of 0.05) for the internal validation, and the AUROC was 0.75 for the temporal validation. The registry model achieved an average AUROC of 0.76 (standard deviation of 0.05) in the internal validation and 0.77 in the temporal validation. In the EHR data, age, and respiratory-system related variables were the most important risk factors identified. In the NICE registry data, age and chronic respiratory insufficiency were the most important risk factors. ConclusionIn our study, prognostic models built on less-granular but readily-available registry data had similar performance to models built on high-granular EHR data and showed similar transportability to a prospective COVID-19 population. Future research is needed to verify whether this finding can be confirmed for upcoming waves.

DeepZF: improved DNA-binding prediction of C2H2-zinc-finger proteins by deep transfer learning.

Cys2His2 zinc-finger (C2H2-ZF) proteins are the largest class of human transcription factors and hence play central roles in gene regulation and cell function. C2H2-ZF proteins are characterized by a DNA-binding domain containing multiple ZFs. A subset of the ZFs bind diverse DNA triplets. Despite their central roles, little is known about which of their ZFs are binding and how the DNA-binding preferences are encoded in the amino acid sequence of each ZF. We present DeepZF, a deep-learning-based pipeline for predicting binding ZFs and their DNA-binding preferences given only the amino acid sequence of a C2H2-ZF protein. To the best of our knowledge, we compiled the first in vivo dataset of binding and non-binding ZFs for training the first ZF-binding classifier. Our classifier, which is based on a novel protein transformer, achieved an average AUROC of 0.71. Moreover, we took advantage of both in vivo and in vitro datasets to learn the recognition code of ZF-DNA binding through transfer learning. Our newly developed model, which is the first to utilize deep learning for the task, achieved an average Pearson correlation greater than 0.94 over each of the three DNA binding positions. Together, DeepZF outperformed extant methods in the task of C2H2-ZF protein DNA-binding preferences prediction: it achieved an average Pearson correlation of 0.42 in motif similarity compared with an average correlation smaller than 0.1 achieved by extant methods. By applying established interpretability techniques, we show that DeepZF inferred biologically relevant binding principles, such as the effect of amino acid residue positions on ZF DNA-binding potential. DeepZF code, model, and results are available via github.com/OrensteinLab/DeepZF. Supplementary data are available at Bioinformatics online.