Comprehensive Study on Molecular Supervised Learning with Graph Neural Networks.

This work considers strategies to develop accurate and reliable graph neural networks (GNNs) for molecular property predictions. Prediction performance of GNNs is highly sensitive to the change in various parameters due to the inherent challenges in molecular machine learning, such as a deficient amount of data samples and bias in data distribution. Comparative studies with well-designed experiments are thus important to clearly understand which GNNs are powerful for molecular supervised learning. Our work presents a number of ablation studies along with a guideline to train and utilize GNNs for both molecular regression and classification tasks. First, we validate that using both atomic and bond meta-information improves the prediction performance in the regression task. Second, we find that the graph isomorphism hypothesis proposed by [Xu, K.; et al How powerful are graph neural networks? 2018, arXiv:1810.00826. arXiv.org e-Print archive. https://arxiv.org/abs/1810.00826] is valid for the regression task. Surprisingly, however, the findings above do not hold for the classification tasks. Beyond the study on model architectures, we test various regularization methods and Bayesian learning algorithms to find the best strategy to achieve a reliable classification system. We demonstrate that regularization methods penalizing predictive entropy might not give well-calibrated probability estimation, even though they work well in other domains, and Bayesian learning methods are capable of developing reliable prediction systems. Furthermore, we argue the importance of Bayesian learning in virtual screening by showing that well-calibrated probability estimation may lead to a higher success rate.

Meta-path Based Prioritization of Functional Drug Actions with Multi-Level Biological Networks.

Functional drug actions refer to drug-affected GO terms. They aid in the investigation of drug effects that are therapeutic or adverse. Previous studies have utilized the linkage information between drugs and functions in molecular level biological networks. Since the current knowledge of molecular level mechanisms of biological functions is still limited, such previous studies were incomplete. We expected that the multi-level biological networks would allow us to more completely investigate the functional drug actions. We constructed multi-level biological networks with genes, GO terms, and diseases. Meta-paths were utilized to extract the features of each GO term. We trained 39 SVM models to prioritize the functional drug actions of the various 39 drugs. Through the multi-level networks, more functional drug actions were utilized for the 39 models and inferred by the models. Multi-level based features improved the performance of the models, and the average AUROC value in the cross-validation was 0.86. Moreover, 60% of the candidates were true.

Publisher Correction: CODA: Integrating multi-level context-oriented directed associations for analysis of drug effects.

A correction to this article has been published and is linked from the HTML and PDF versions of this paper. The error has been fixed in the paper.

Author Correction: CODA: Integrating multi-level context-oriented directed associations for analysis of drug effects.

A correction to this article has been published and is linked from the HTML version of this paper. The error has been fixed in the paper.

CODA: Integrating multi-level context-oriented directed associations for analysis of drug effects.

In silico network-based methods have shown promising results in the field of drug development. Yet, most of networks used in the previous research have not included context information even though biological associations actually do appear in the specific contexts. Here, we reconstruct an anatomical context-specific network by assigning contexts to biological associations using protein expression data and scientific literature. Furthermore, we employ the context-specific network for the analysis of drug effects with a proximity measure between drug targets and diseases. Distinct from previous context-specific networks, intercellular associations and phenomic level entities such as biological processes are included in our network to represent the human body. It is observed that performances in inferring drug-disease associations are increased by adding context information and phenomic level entities. In particular, hypertension, a disease related to multiple organs and associated with several phenomic level entities, is analyzed in detail to investigate how our network facilitates the inference of drug-disease associations. Our results indicate that the inclusion of context information, intercellular associations, and phenomic level entities can contribute towards a better prediction of drug-disease associations and provide detailed insight into understanding of how drugs affect diseases in the human body.

Context-based resolution of semantic conflicts in biological pathways.

BACKGROUND: Interactions between biological entities such as genes, proteins and metabolites, so called pathways, are key features to understand molecular mechanisms of life. As pathway information is being accumulated rapidly through various knowledge resources, there are growing interests in maintaining the integrity of the heterogeneous databases. METHODS: Here, we defined conflict as a status where two contradictory pieces of evidence (i.e. 'A increases B' and 'A decreases B') coexist in a same pathway. This conflict damages unity so that inference of simulation on the integrated pathway network might be unreliable. We defined rule and rule group. A rule consists of interaction of two entities, meta-relation (increase or decrease), and contexts terms about tissue specificity or environmental conditions. The rules, which have the same interaction, are grouped into a rule group. If the rules don't have a unanimous meta-relation, the rule group and the rules are judged as being conflicting. RESULTS: This analysis revealed that almost 20% of known interactions suffer from conflicting information and conflicting information occurred much more frequently in the literature than the public database. CONCLUSIONS: By identifying and resolving the conflicts, we expect that pathway databases can be cleaned and used for better secondary analyses such as gene/protein annotation, network dynamics and qualitative/quantitative simulation.

Anomalies in network bridges involved in bile Acid metabolism predict outcomes of colorectal cancer patients.

Biomarkers prognostic for colorectal cancer (CRC) would be highly desirable in clinical practice. Proteins that regulate bile acid (BA) homeostasis, by linking metabolic sensors and metabolic enzymes, also called bridge proteins, may be reliable prognostic biomarkers for CRC. Based on a devised metric, "bridgeness," we identified bridge proteins involved in the regulation of BA homeostasis and identified their prognostic potentials. The expression patterns of these bridge proteins could distinguish between normal and diseased tissues, suggesting that these proteins are associated with CRC pathogenesis. Using a supervised classification system, we found that these bridge proteins were reproducibly prognostic, with high prognostic ability compared to other known markers.