Cancer involves dynamic changes caused by (epi)genetic alterations such as mutations or abnormal DNA methylation patterns which occur in cancer driver genes. These driver genes are divided into oncogenes and tumor suppressors depending on their function and mechanism of action. Discovering driver genes in different cancer (sub)types is important not only for increasing current understanding of carcinogenesis but also from prognostic and therapeutic perspectives. We have previously developed a framework called Moonlight which uses a systems biology multi-omics approach for prediction of driver genes. Here, we present an important development in Moonlight2 by incorporating a DNA methylation layer which provides epigenetic evidence for deregulated expression profiles of driver genes. To this end, we present a novel functionality called Gene Methylation Analysis (GMA) which investigates abnormal DNA methylation patterns to predict driver genes. This is achieved by integrating the tool EpiMix which is designed to detect such aberrant DNA methylation patterns in a cohort of patients and further couples these patterns with gene expression changes. To showcase GMA, we applied it to three cancer (sub)types (basal-like breast cancer, lung adenocarcinoma, and thyroid carcinoma) where we discovered 33, 190, and 263 epigenetically driven genes, respectively. A subset of these driver genes had prognostic effects with expression levels significantly affecting survival of the patients. Moreover, a subset of the driver genes demonstrated therapeutic potential as drug targets. This study provides a framework for exploring the driving forces behind cancer and provides novel insights into the landscape of three cancer sub(types) by integrating gene expression and methylation data.

Accurately labeling large datasets is important for biomedical machine learning yet challenging while modern data augmentation methods may generate noise in the training data, which may deteriorate machine learning model performance. Existing approaches addressing noisy training data typically rely on strict modeling assumptions, classification models and well-curated dataset. To address these, we propose a novel reliability-based training-data-cleaning method employing inductive conformal prediction (ICP). This method uses a small set of well-curated training data and leverages ICP-calculated reliability metrics to selectively correct mislabeled data and outliers within vast quantities of noisy training data. The efficacy is validated across three classification tasks with distinct modalities: filtering drug-induced-liver-injury (DILI) literature with free-text title and abstract, predicting ICU admission of COVID-19 patients through CT radiomics and electronic health records, and subtyping breast cancer using RNA-sequencing data. Varying levels of noise to the training labels were introduced via label permutation. Our training-data-cleaning method significantly enhanced the downstream classification performance (paired t-tests, p ≤ 0 . 05 among 30 random train/test partitions): significant accuracy enhancement in 86 out of 96 DILI experiments (up to 11.4% increase from 0.812 to 0.905), significant AUROC and AUPRC enhancements in all 48 COVID-19 experiments (up to 23.8% increase from 0.597 to 0.739 for AUROC, and 69.8% increase from 0.183 to 0.311 for AUPRC), and significant accuracy and macro-average F1-score improvements in 47 out of 48 RNA-sequencing experiments (up to 74.6% increase from 0.351 to 0.613 for accuracy, and 89.0% increase from 0.267 to 0.505 for F1-score). The improvement can be both statistically and clinically significant for information retrieval, disease diagnosis and prognosis. The method offers the potential to substantially boost classification performance in biomedical machine learning tasks without necessitating an excessive volume of well-curated training data or strong data distribution and modeling assumptions in existing semi-supervised learning methods.

MOTIVATION: Drug-target interaction (DTI) prediction is a relevant but challenging task in the drug repurposing field. In-silico approaches have drawn particular attention as they can reduce associated costs and time commitment of traditional methodologies. Yet, current state-of-the-art methods present several limitations: existing DTI prediction approaches are computationally expensive, thereby hindering the ability to use large networks and exploit available datasets and, the generalization to unseen datasets of DTI prediction methods remains unexplored, which could potentially improve the development processes of DTI inferring approaches in terms of accuracy and robustness. RESULTS: In this work, we introduce GeNNius (Graph Embedding Neural Network Interaction Uncovering System), a Graph Neural Network (GNN)-based method that outperforms state-of-the-art models in terms of both accuracy and time efficiency across a variety of datasets. We also demonstrated its prediction power to uncover new interactions by evaluating not previously known DTIs for each dataset. We further assessed the generalization capability of GeNNius by training and testing it on different datasets, showing that this framework can potentially improve the DTI prediction task by training on large datasets and testing on smaller ones. Finally, we investigated qualitatively the embeddings generated by GeNNius, revealing that the GNN encoder maintains biological information after the graph convolutions while diffusing this information through nodes, eventually distinguishing protein families in the node embedding space. AVAILABILITY AND IMPLEMENTATION: GeNNius code is available at https://github.com/ubioinformat/GeNNius.

Machine learning (ML) applications in medical artificial intelligence (AI) systems have shifted from traditional and statistical methods to increasing application of deep learning models. This survey navigates the current landscape of multimodal ML, focusing on its profound impact on medical image analysis and clinical decision support systems. Emphasizing challenges and innovations in addressing multimodal representation, fusion, translation, alignment, and co-learning, the paper explores the transformative potential of multimodal models for clinical predictions. It also highlights the need for principled assessments and practical implementation of such models, bringing attention to the dynamics between decision support systems and healthcare providers and personnel. Despite advancements, challenges such as data biases and the scarcity of "big data" in many biomedical domains persist. We conclude with a discussion on principled innovation and collaborative efforts to further the mission of seamless integration of multimodal ML models into biomedical practice.

OBJECTIVE: While there are currently approaches to handle unstructured clinical data, such as manual abstraction and structured proxy variables, these methods may be time-consuming, not scalable, and imprecise. This article aims to determine whether selective prediction, which gives a model the option to abstain from generating a prediction, can improve the accuracy and efficiency of unstructured clinical data abstraction. MATERIALS AND METHODS: We trained selective classifiers (logistic regression, random forest, support vector machine) to extract 5 variables from clinical notes: depression (n = 1563), glioblastoma (GBM, n = 659), rectal adenocarcinoma (DRA, n = 601), and abdominoperineal resection (APR, n = 601) and low anterior resection (LAR, n = 601) of adenocarcinoma. We varied the cost of false positives (FP), false negatives (FN), and abstained notes and measured total misclassification cost. RESULTS: The depression selective classifiers abstained on anywhere from 0% to 97% of notes, and the change in total misclassification cost ranged from -58% to 9%. Selective classifiers abstained on 5%-43% of notes across the GBM and colorectal cancer models. The GBM selective classifier abstained on 43% of notes, which led to improvements in sensitivity (0.94 to 0.96), specificity (0.79 to 0.96), PPV (0.89 to 0.98), and NPV (0.88 to 0.91) when compared to a non-selective classifier and when compared to structured proxy variables. DISCUSSION: We showed that selective classifiers outperformed both non-selective classifiers and structured proxy variables for extracting data from unstructured clinical notes. CONCLUSION: Selective prediction should be considered when abstaining is preferable to making an incorrect prediction.

Drug-induced liver injury describes the adverse effects of drugs that damage the liver. Life-threatening results were also reported in severe cases. Therefore, liver toxicity is an important assessment for new drug candidates. These reports are documented in research papers that contain preliminary in vitro and in vivo experiments. Conventionally, data extraction from publications relies on resource-demanding manual labeling, which restricts the efficiency of the information extraction. The development of natural language processing techniques enables the automatic processing of biomedical texts. Herein, based on around 28,000 papers (titles and abstracts) provided by the Critical Assessment of Massive Data Analysis challenge, this study benchmarked model performances on filtering liver-damage-related literature. Among five text embedding techniques, the model using term frequency-inverse document frequency (TF-IDF) and logistic regression outperformed others with an accuracy of 0.957 on the validation set. Furthermore, an ensemble model with similar overall performances was developed with a logistic regression model on the predicted probability given by separate models with different vectorization techniques. The ensemble model achieved a high accuracy of 0.954 and an F1 score of 0.955 in the hold-out validation data in the challenge. Moreover, important words in positive/negative predictions were identified via model interpretation. The prediction reliability was quantified with conformal prediction, which provides users with a control over the prediction uncertainty. Overall, the ensemble model and TF-IDF model reached satisfactory classification results, which can be used by researchers to rapidly filter literature that describes events related to liver injury induced by medications.

MOTIVATION: Gene regulatory networks describe the regulatory relationships among genes, and developing methods for reverse engineering these networks is an ongoing challenge in computational biology. The majority of the initially proposed methods for gene regulatory network discovery create a network of genes and then mine it in order to uncover previously unknown regulatory processes. More recent approaches have focused on inferring modules of co-regulated genes, linking these modules with regulatory genes and then mining them to discover new molecular biology. RESULTS: In this work we analyze module-based network approaches to build gene regulatory networks, and compare their performance to single gene network approaches. In the process, we propose a novel approach to estimate gene regulatory networks drawing from the module-based methods. We show that generating modules of co-expressed genes which are predicted by a sparse set of regulators using a variational Bayes method, and then building a bipartite graph on the generated modules using sparse regression, yields more informative networks than previous single and module-based network approaches as measured by: (i) the rate of enriched gene sets, (ii) a network topology assessment, (iii) ChIP-Seq evidence and (iv) the KnowEnG Knowledge Network collection of previously characterized gene-gene interactions. AVAILABILITY AND IMPLEMENTATION: The code is written in R and can be downloaded from https://github.com/mikelhernaez/linker. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

MOTIVATION: Estimating the future course of patients with cancer lesions is invaluable to physicians; however, current clinical methods fail to effectively use the vast amount of multimodal data that is available for cancer patients. To tackle this problem, we constructed a multimodal neural network-based model to predict the survival of patients for 20 different cancer types using clinical data, mRNA expression data, microRNA expression data and histopathology whole slide images (WSIs). We developed an unsupervised encoder to compress these four data modalities into a single feature vector for each patient, handling missing data through a resilient, multimodal dropout method. Encoding methods were tailored to each data type-using deep highway networks to extract features from clinical and genomic data, and convolutional neural networks to extract features from WSIs. RESULTS: We used pancancer data to train these feature encodings and predict single cancer and pancancer overall survival, achieving a C-index of 0.78 overall. This work shows that it is possible to build a pancancer model for prognosis that also predicts prognosis in single cancer sites. Furthermore, our model handles multiple data modalities, efficiently analyzes WSIs and represents patient multimodal data flexibly into an unsupervised, informative representation. We thus present a powerful automated tool to accurately determine prognosis, a key step towards personalized treatment for cancer patients. AVAILABILITY AND IMPLEMENTATION: https://github.com/gevaertlab/MultimodalPrognosis.

Aberrant DNA methylation disrupts normal gene expression in cancer and broadly contributes to oncogenesis. We previously developed MethylMix, a model-based algorithmic approach to identify epigenetically regulated driver genes. MethylMix identifies genes where methylation likely executes a functional role by using transcriptomic data to select only methylation events that can be linked to changes in gene expression. However, given that proteins more closely link genotype to phenotype recent high-throughput proteomic data provides an opportunity to more accurately identify functionally relevant abnormal methylation events. Here we present a MethylMix analysis that refines nominations for epigenetic driver genes by leveraging quantitative high-throughput proteomic data to select only genes where DNA methylation is predictive of protein abundance. Applying our algorithm across three cancer cohorts we find that using protein abundance data narrows candidate nominations, where the effect of DNA methylation is often buffered at the protein level. Next, we find that MethylMix genes predictive of protein abundance are enriched for biological processes involved in cancer including functions involved in epithelial and mesenchymal transition. Moreover, our results are also enriched for tumor markers which are predictive of clinical features like tumor stage and we find clustering using MethylMix genes predictive of protein abundance captures cancer subtypes.

Summary: DNA methylation is an important mechanism regulating gene transcription, and its role in carcinogenesis has been extensively studied. Hyper and hypomethylation of genes is a major mechanism of gene expression deregulation in a wide range of diseases. At the same time, high-throughput DNA methylation assays have been developed generating vast amounts of genome wide DNA methylation measurements. We developed MethylMix, an algorithm implemented in R to identify disease specific hyper and hypomethylated genes. Here we present a new version of MethylMix that automates the construction of DNA-methylation and gene expression datasets from The Cancer Genome Atlas (TCGA). More precisely, MethylMix 2.0 incorporates two major updates: the automated downloading of DNA methylation and gene expression datasets from TCGA and the automated preprocessing of such datasets: value imputation, batch correction and CpG sites clustering within each gene. The resulting datasets can subsequently be analyzed with MethylMix to identify transcriptionally predictive methylation states. We show that the Differential Methylation Values created by MethylMix can be used for cancer subtyping. Availability and implementation: MethylMix 2.0 was implemented as an R package and is available in bioconductor. https://www.bioconductor.org/packages/release/bioc/html/MethylMix.html.

BACKGROUND: The current state-of-the-art in cancer diagnosis and treatment is not ideal; diagnostic tests are accurate but invasive, and treatments are "one-size fits-all" instead of being personalized. Recently, miRNA's have garnered significant attention as cancer biomarkers, owing to their ease of access (circulating miRNA in the blood) and stability. There have been many studies showing the effectiveness of miRNA data in diagnosing specific cancer types, but few studies explore the role of miRNA in predicting treatment outcome. METHODS: Here we go a step further, using tissue miRNA and clinical data across 21 cancers from the 'The Cancer Genome Atlas' (TCGA) database. We use machine learning techniques to create an accurate pan-cancer diagnosis system, and a prediction model for treatment outcomes. Finally, using these models, we create a web-based tool that diagnoses cancer and recommends the best treatment options. RESULTS: We achieved 97.2% accuracy for classification using a support vector machine classifier with radial basis. The accuracies improved to 99.9-100% when climbing up the embryonic tree and classifying cancers at different stages. We define the accuracy as the ratio of the total number of instances correctly classified to the total instances. The classifier also performed well, achieving greater than 80% sensitivity for many cancer types on independent validation datasets. Many miRNAs selected by our feature selection algorithm had strong previous associations to various cancers and tumor progression. Then, using miRNA, clinical and treatment data and encoding it in a machine-learning readable format, we built a prognosis predictor model to predict the outcome of treatment with 85% accuracy. We used this model to create a tool that recommends personalized treatment regimens. Both the diagnosis and prognosis model, incorporating semi-supervised learning techniques to improve their accuracies with repeated use, were uploaded online for easy access. CONCLUSION: Our research is a step towards the final goal of diagnosing cancer and predicting treatment recommendations using non-invasive blood tests.

A crucial and limiting factor in data reuse is the lack of accurate, structured, and complete descriptions of data, known as metadata. Towards improving the quantity and quality of metadata, we propose a novel metadata prediction framework to learn associations from existing metadata that can be used to predict metadata values. We evaluate our framework in the context of experimental metadata from the Gene Expression Omnibus (GEO). We applied four rule mining algorithms to the most common structured metadata elements (sample type, molecular type, platform, label type and organism) from over 1.3million GEO records. We examined the quality of well supported rules from each algorithm and visualized the dependencies among metadata elements. Finally, we evaluated the performance of the algorithms in terms of accuracy, precision, recall, and F-measure. We found that PART is the best algorithm outperforming Apriori, Predictive Apriori, and Decision Table. All algorithms perform significantly better in predicting class values than the majority vote classifier. We found that the performance of the algorithms is related to the dimensionality of the GEO elements. The average performance of all algorithm increases due of the decreasing of dimensionality of the unique values of these elements (2697 platforms, 537 organisms, 454 labels, 9 molecules, and 5 types). Our work suggests that experimental metadata such as present in GEO can be accurately predicted using rule mining algorithms. Our work has implications for both prospective and retrospective augmentation of metadata quality, which are geared towards making data easier to find and reuse.

Accurate lung nodule segmentation from computed tomography (CT) images is of great importance for image-driven lung cancer analysis. However, the heterogeneity of lung nodules and the presence of similar visual characteristics between nodules and their surroundings make it difficult for robust nodule segmentation. In this study, we propose a data-driven model, termed the Central Focused Convolutional Neural Networks (CF-CNN), to segment lung nodules from heterogeneous CT images. Our approach combines two key insights: 1) the proposed model captures a diverse set of nodule-sensitive features from both 3-D and 2-D CT images simultaneously; 2) when classifying an image voxel, the effects of its neighbor voxels can vary according to their spatial locations. We describe this phenomenon by proposing a novel central pooling layer retaining much information on voxel patch center, followed by a multi-scale patch learning strategy. Moreover, we design a weighted sampling to facilitate the model training, where training samples are selected according to their degree of segmentation difficulty. The proposed method has been extensively evaluated on the public LIDC dataset including 893 nodules and an independent dataset with 74 nodules from Guangdong General Hospital (GDGH). We showed that CF-CNN achieved superior segmentation performance with average dice scores of 82.15% and 80.02% for the two datasets respectively. Moreover, we compared our results with the inter-radiologists consistency on LIDC dataset, showing a difference in average dice score of only 1.98%.

UNLABELLED: DNA methylation is an important mechanism regulating gene transcription, and its role in carcinogenesis has been extensively studied. Hyper and hypomethylation of genes is an alternative mechanism to deregulate gene expression in a wide range of diseases. At the same time, high-throughput DNA methylation assays have been developed generating vast amounts of genome wide DNA methylation measurements. Yet, few tools exist that can formally identify hypo and hypermethylated genes that are predictive of transcription and thus functionally relevant for a particular disease. To accommodate this lack of tools, we developed MethylMix, an algorithm implemented in R to identify disease specific hyper and hypomethylated genes. MethylMix is based on a beta mixture model to identify methylation states and compares them with the normal DNA methylation state. MethylMix introduces a novel metric, the 'Differential Methylation value' or DM-value defined as the difference of a methylation state with the normal methylation state. Finally, matched gene expression data are used to identify, besides differential, transcriptionally predictive methylation states by focusing on methylation changes that effect gene expression. AVAILABILITY AND IMPLEMENTATION: MethylMix was implemented as an R package and is available in bioconductor.

The Center for Expanded Data Annotation and Retrieval is studying the creation of comprehensive and expressive metadata for biomedical datasets to facilitate data discovery, data interpretation, and data reuse. We take advantage of emerging community-based standard templates for describing different kinds of biomedical datasets, and we investigate the use of computational techniques to help investigators to assemble templates and to fill in their values. We are creating a repository of metadata from which we plan to identify metadata patterns that will drive predictive data entry when filling in metadata templates. The metadata repository not only will capture annotations specified when experimental datasets are initially created, but also will incorporate links to the published literature, including secondary analyses and possible refinements or retractions of experimental interpretations. By working initially with the Human Immunology Project Consortium and the developers of the ImmPort data repository, we are developing and evaluating an end-to-end solution to the problems of metadata authoring and management that will generalize to other data-management environments.