bAIcis: A Novel Bayesian Network Structural Learning Algorithm and Its Comprehensive Performance Evaluation Against Open-Source Software

Abstract

Structural learning of Bayesian networks (BNs) from observational data has gained increasing applied use and attention from various scientific and industrial areas. The mathematical theory of BNs and their optimization is well developed. Although there are several open-source BN learners in the public domain, none of them are able to handle both small and large feature space data and recover network structures with acceptable accuracy. bAIcis Ò is a novel BN learning and simulation software from BERG. It was developed with the goal of learning BNs from ''Big Data'' in health care, often exceeding hundreds of thousands features when research is conducted in genomics or multi-omics. This article provides a comprehensive performance evaluation of bAIcis and its comparison with the open-source BN learners. The study investigated synthetic datasets of discrete, continuous, and mixed data in small and large feature space, respectively. The results demonstrated that bAIcis outperformed the publicly available algorithms in structure recovery precision in almost all of the evaluated settings, achieving the true positive rates of 0.9 and precision of 0.8. In addition, bAIcis supports all data types, including continuous, discrete, and mixed variables. It is effectively parallelized on a distributed system and can work with datasets of thousands of features that are infeasible for any of the publicly available tools with a desired level of recovery accuracy.

Bayesian Model Infers Drug Repurposing Candidates for Treatment of COVID-19

“…Network based dynamic and temporal graphs have already been deployed to deconvolute epidemiological data associated with SARS-CoV-2 and other approaches such as influence diffusion algorithms hold promise in complex datasets [7,8]. Herein, we applied Bayesian networks [9], differential expression analysis and drug-enrichment analysis to a publicly available longitudinal dataset of proteomics and viral titer from Caco-2 cells infected with SARS-CoV-2 [10]. Utilizing this BN-based approach, a de novo map of host proteins was established as well as drug enrichment signatures were defined.…”

Synthetic data generation with probabilistic Bayesian Networks

“…Most of the recent work in the field has been aimed at improving the BN modeling scalability, handling mixed data types (including various types of biological data), incorporating latent variables, and developing more robust software interfaces and visualization options (Andrews, Ramsey, and Cooper 2018;Andrews, Ramsey, and Cooper 2019;Chen et al 2019;Hong et al 2018;Jabbari et al 2017;Ogarrio, Spirtes, and J 2016;Raghu et al 2018;Ramsey et al 2017;Sedgewick et al 2016;Spirtes and Zhang 2016;Xing et al 2017;Xing et al 2018;Yu, Liu, and Li 2019;Zhang et al 2018;Zhang et al 2019). In view of these developments, the ability to objectively assess the performance of BN modeling is of the utmost importance.…”

“…In view of these developments, the ability to objectively assess the performance of BN modeling is of the utmost importance. Currently, there are three principal venues for accomplishing this: (i) using well-established (in machine learning community) predefined benchmark models/datasets, such as Alarm or Asia, (ii) using various specialized biologically-oriented benchmark datasets, both real and simulated, such as DREAM (Eicher et al 2019;Wang, Audenaert, and Michoel 2019;Xing et al 2017), and (iii) developing more-or-less realistic simulation frameworks (Andrews, Ramsey, and Cooper 2018;Han et al 2017;Ellis and Wong 2008;Isci et al 2011;Tasaki et al 2015;Zhang et al 2019). The first approach is necessarily limited and does not generalize to the modern high-dimensional biological data.…”

“…In this study, we concentrate on the third approach, namely constructing robust, generalizable, and mathematically rigorous simulation frameworks. Currently, these frameworks typically involve (i) specifying the synthetic network structures, and (ii) utilizing linear SEM (structural equation modeling) methodology to generate the actual synthetic data (Isci et al 2011;Tasaki et al 2015;Zhang et al 2019).…”

“…In the broader context of heterogeneous biological data and networks, it is important to distinguish between the notions of causality, correlation, association and dependence / conditional independence. Establishing causality (Cooper et al 2015;Glymour, Zhang, and Spirtes 2019;Pearl 2009;Zhang et al 2018) is usually perceived as the ultimate goal in the biological network analyses (Ramsey et al 2017;Zeng, Jiang, and Neapolitan 2016;Zhang et al 2019). However, establishing (and even defining) causality is often equivocal, for both theoretical (Glymour, Zhang, and Spirtes 2019;Zhang et al 2018) and practical/biological reasons (e.g., ambiguity of the causal interpretation of the relationship between two correlated gene expression measurements, or two single nucleotide polymorphisms in linkage disequilibrium) (Tasaki et al 2015).…”