Bayesian Prediction Model Research Articles

IGM‐based perceptual multimodal medical image fusion using free energy motivated adaptive PCNN

AbstractMultimodal medical image fusion merges two medical images to produce a visual enhanced fused image, to provide more accurate comprehensive pathological information to doctors for better diagnosis and treatment. In this article, we present a perceptual multimodal medical image fusion method with free energy (FE) motivated adaptive pulse coupled neural network (PCNN) by employing Internal Generative Mechanism (IGM). First, source images are divided into predicted layers and detail layers with Bayesian prediction model. Then to retain human visual system inspired features, FE is used to motivate the PCNN for processing detail layers, and large firing times are selected as coefficients. The predicted layers are fused with the averaging strategy as activity level measurement. Finally, the fused image is reconstructed by merging coefficients in both fused layers. Experimental results visually and quantitatively show that the proposed fusion strategy is superior to the state‐of‐the‐art methods.

Invited Commentary

Using data from the United Network for Organ Sharing (UNOS) publicly available, deidentified Scientific Registry of Transplant Recipients (SRTR) database, Crawford and colleagues [1Crawford T.C. Magruder J.T. Grimm J.C. et al.A comprehensive risk score to predict prolonged hospital length of stay after heart transplantation.Ann Thorac Surg. 2018; 105: 83-91Abstract Full Text Full Text PDF PubMed Scopus (14) Google Scholar] performed a multivariate logistic regression analysis to develop a predictive risk score for prolonged postoperative hospital length of stay in a cohort of isolated, adult heart transplantation recipients who survived to discharge from 2003 to 2012. Notable exclusions from their analysis included patients with missing postoperative length of stay data and all patients who died during the postoperative period during their index hospitalization before discharge. Using an approximation of the 90th percentile of length of stays for the entire cohort, the authors defined a prolonged postoperative hospital length of stay as a posttransplant hospital length of stay greater than 30 days. Using the c-statistic to evaluate for discrimination, the authors found their model to be modestly accurate at best (c-statistic 0.61). Multivariate logistic regression has been the cornerstone on which cardiac surgery risk estimation models have been developed over the last 3 decades. The methodology assumes fixed, weighted combinations of independent variables (input) and linearly relates them to the dependent (output) variables under evaluation. However, using logistic regression to model complex postoperative outcomes is insufficient as this method is unable to correctly evaluate the nonlinear and indirect relationships that characterize more complicated clinical scenarios, leading to more often than not inaccurately published risk prediction models. One potential way of more accurately assessing these nonlinear relationships is through the use of advanced machine learning statistical methods. Machine learning techniques, coupled with readily available large electronic datasets and open-source statistical software, can integrate massive feature sets to model many interrelated, nonlinear relationships and higher order interactions. Examples of machine learning methods arriving at more accurate clinical risk prediction models are present throughout the medical literature but are also now beginning to dot the cardiac surgery evidence base as well. Examples include the Bayesian network approach of Loghmanpour and colleagues [2Loghmanpour N.A. Kanwar M.K. Druzdzel M.J. Benza R.L. Murali S. Antaki J.F. A new Bayesian network-based risk stratification model for prediction of short-term and long-term LVAD mortality.ASAIO J. 2015; 61: 313-323Crossref PubMed Scopus (34) Google Scholar] for short- and long-term mortality risk in left ventricular assist device patients using the Interagency Registry for Mechanically Assisted Circulatory Support (INTERMACS) data; and the more recent publication by Allyn and colleagues [3Allyn J. Allou N. Augustin P. et al.A comparison of a machine learning model with EuroSCORE II in predicting mortality after elective cardiac surgery: a decision curve analysis.PLoS One. 2017; 12: e0169772Crossref PubMed Scopus (95) Google Scholar] of an ensemble machine learning risk prediction model that outperformed the European System for Cardiac Operative Risk Evaluation (EuroSCORE) II in a cohort of 6,520 cardiac surgery patients from a single center in France. The idea behind building a risk prediction model with forced linear relationships results in publication of many inaccurate models. Instead, we should strive to assess the true nature of the underlying relationships between clinical-level features through machine learning methodologies in hopes of developing more accurate risk stratification and prediction scores that will help us more precisely inform patient counseling and surgeon practice. A Comprehensive Risk Score to Predict Prolonged Hospital Length of Stay After Heart TransplantationThe Annals of Thoracic SurgeryVol. 105Issue 1PreviewProlonged hospital length of stay (PLOS) after heart transplantation increases cost and morbidity. To better inform care, we developed a risk score to identify patients at risk for PLOS after heart transplantation. Full-Text PDF

A Hierarchical Bayesian Model for the Identification of PET Markers Associated to the Prediction of Surgical Outcome after Anterior Temporal Lobe Resection.

We develop an integrative Bayesian predictive modeling framework that identifies individual pathological brain states based on the selection of fluoro-deoxyglucose positron emission tomography (PET) imaging biomarkers and evaluates the association of those states with a clinical outcome. We consider data from a study on temporal lobe epilepsy (TLE) patients who subsequently underwent anterior temporal lobe resection. Our modeling framework looks at the observed profiles of regional glucose metabolism in PET as the phenotypic manifestation of a latent individual pathologic state, which is assumed to vary across the population. The modeling strategy we adopt allows the identification of patient subgroups characterized by latent pathologies differentially associated to the clinical outcome of interest. It also identifies imaging biomarkers characterizing the pathological states of the subjects. In the data application, we identify a subgroup of TLE patients at high risk for post-surgical seizure recurrence after anterior temporal lobe resection, together with a set of discriminatory brain regions that can be used to distinguish the latent subgroups. We show that the proposed method achieves high cross-validated accuracy in predicting post-surgical seizure recurrence.

Variable selection models for genomic selection using whole-genome sequence data and singular value decomposition

BackgroundNon-linear Bayesian genomic prediction models such as BayesA/B/C/R involve iteration and mostly Markov chain Monte Carlo (MCMC) algorithms, which are computationally expensive, especially when whole-genome sequence (WGS) data are analyzed. Singular value decomposition (SVD) of the genotype matrix can facilitate genomic prediction in large datasets, and can be used to estimate marker effects and their prediction error variances (PEV) in a computationally efficient manner. Here, we developed, implemented, and evaluated a direct, non-iterative method for the estimation of marker effects for the BayesC genomic prediction model.MethodsThe BayesC model assumes a priori that markers have normally distributed effects with probability uppi and no effect with probability (1 − uppi ). Marker effects and their PEV are estimated by using SVD and the posterior probability of the marker having a non-zero effect is calculated. These posterior probabilities are used to obtain marker-specific effect variances, which are subsequently used to approximate BayesC estimates of marker effects in a linear model. A computer simulation study was conducted to compare alternative genomic prediction methods, where a single reference generation was used to estimate marker effects, which were subsequently used for 10 generations of forward prediction, for which accuracies were evaluated.ResultsSVD-based posterior probabilities of markers having non-zero effects were generally lower than MCMC-based posterior probabilities, but for some regions the opposite occurred, resulting in clear signals for QTL-rich regions. The accuracies of breeding values estimated using SVD- and MCMC-based BayesC analyses were similar across the 10 generations of forward prediction. For an intermediate number of generations (2 to 5) of forward prediction, accuracies obtained with the BayesC model tended to be slightly higher than accuracies obtained using the best linear unbiased prediction of SNP effects (SNP-BLUP model). When reducing marker density from WGS data to 30 K, SNP-BLUP tended to yield the highest accuracies, at least in the short term.ConclusionsBased on SVD of the genotype matrix, we developed a direct method for the calculation of BayesC estimates of marker effects. Although SVD- and MCMC-based marker effects differed slightly, their prediction accuracies were similar. Assuming that the SVD of the marker genotype matrix is already performed for other reasons (e.g. for SNP-BLUP), computation times for the BayesC predictions were comparable to those of SNP-BLUP.

Modeling heterogeneous (co)variances from adjacent-SNP groups improves genomic prediction for milk protein composition traits

BackgroundAccurate genomic prediction requires a large reference population, which is problematic for traits that are expensive to measure. Traits related to milk protein composition are not routinely recorded due to costly procedures and are considered to be controlled by a few quantitative trait loci of large effect. The amount of variation explained may vary between regions leading to heterogeneous (co)variance patterns across the genome. Genomic prediction models that can efficiently take such heterogeneity of (co)variances into account can result in improved prediction reliability. In this study, we developed and implemented novel univariate and bivariate Bayesian prediction models, based on estimates of heterogeneous (co)variances for genome segments (BayesAS). Available data consisted of milk protein composition traits measured on cows and de-regressed proofs of total protein yield derived for bulls. Single-nucleotide polymorphisms (SNPs), from 50K SNP arrays, were grouped into non-overlapping genome segments. A segment was defined as one SNP, or a group of 50, 100, or 200 adjacent SNPs, or one chromosome, or the whole genome. Traditional univariate and bivariate genomic best linear unbiased prediction (GBLUP) models were also run for comparison. Reliabilities were calculated through a resampling strategy and using deterministic formula.ResultsBayesAS models improved prediction reliability for most of the traits compared to GBLUP models and this gain depended on segment size and genetic architecture of the traits. The gain in prediction reliability was especially marked for the protein composition traits β-CN, κ-CN and β-LG, for which prediction reliabilities were improved by 49 percentage points on average using the MT-BayesAS model with a 100-SNP segment size compared to the bivariate GBLUP. Prediction reliabilities were highest with the BayesAS model that uses a 100-SNP segment size. The bivariate versions of our BayesAS models resulted in extra gains of up to 6% in prediction reliability compared to the univariate versions.ConclusionsSubstantial improvement in prediction reliability was possible for most of the traits related to milk protein composition using our novel BayesAS models. Grouping adjacent SNPs into segments provided enhanced information to estimate parameters and allowing the segments to have different (co)variances helped disentangle heterogeneous (co)variances across the genome.

A multi-paradigm framework to assess the impacts of climate change on end-use energy demand.

Projecting the long-term trends in energy demand is an increasingly complex endeavor due to the uncertain emerging changes in factors such as climate and policy. The existing energy-economy paradigms used to characterize the long-term trends in the energy sector do not adequately account for climate variability and change. In this paper, we propose a multi-paradigm framework for estimating the climate sensitivity of end-use energy demand that can easily be integrated with the existing energy-economy models. To illustrate the applicability of our proposed framework, we used the energy demand and climate data in the state of Indiana to train a Bayesian predictive model. We then leveraged the end-use demand trends as well as downscaled future climate scenarios to generate probabilistic estimates of the future end-use demand for space cooling, space heating and water heating, at the individual household and building level, in the residential and commercial sectors. Our results indicated that the residential load is much more sensitive to climate variability and change than the commercial load. Moreover, since the largest fraction of the residential energy demand in Indiana is attributed to heating, future warming scenarios could lead to reduced end-use demand due to lower space heating and water heating needs. In the commercial sector, the overall energy demand is expected to increase under the future warming scenarios. This is because the increased cooling load during hotter summer months will likely outpace the reduced heating load during the more temperate winter months.

Genome-wide regression models considering general and specific combining ability predict hybrid performance in oilseed rape with similar accuracy regardless of trait architecture.

Genomic prediction using the Brassica 60k genotyping array is efficient in oilseed rape hybrids. Prediction accuracy is more dependent on trait complexity than on the prediction model. In oilseed rape breeding programs, performance prediction of parental combinations is of fundamental importance. Due to the phenomenon of heterosis, per se performance is not a reliable indicator for F1-hybrid performance, and selection of well-paired parents requires the testing of large quantities of hybrid combinations in extensive field trials. However, the number of potential hybrids, in general, dramatically exceeds breeding capacity and budget. Integration of genomic selection (GS) could substantially increase the number of potential combinations that can be evaluated. GS models can be used to predict the performance of untested individuals based only on their genotypic profiles, using marker effects previously predicted in a training population. This allows for a preselection of promising genotypes, enabling a more efficient allocation of resources. In this study, we evaluated the usefulness of the Illumina Brassica 60k SNP array for genomic prediction and compared three alternative approaches based on a homoscedastic ridge regression BLUP and three Bayesian prediction models that considered general and specific combining ability (GCA and SCA, respectively). A total of 448 hybrids were produced in a commercial breeding program from unbalanced crosses between 220 paternal doubled haploid lines and five male-sterile testers. Predictive ability was evaluated for seven agronomic traits. We demonstrate that the Brassica 60k genotyping array is an adequate and highly valuable platform to implement genomic prediction of hybrid performance in oilseed rape. Furthermore, we present first insights into the application of established statistical models for prediction of important agronomical traits with contrasting patterns of polygenic control.

Signatures of positive selection reveal a universal role of chromatin modifiers as cancer driver genes

Tumors are composed of an evolving population of cells subjected to tissue-specific selection, which fuels tumor heterogeneity and ultimately complicates cancer driver gene identification. Here, we integrate cancer cell fraction, population recurrence, and functional impact of somatic mutations as signatures of selection into a Bayesian model for driver prediction. We demonstrate that our model, cDriver, outperforms competing methods when analyzing solid tumors, hematological malignancies, and pan-cancer datasets. Applying cDriver to exome sequencing data of 21 cancer types from 6,870 individuals revealed 98 unreported tumor type-driver gene connections. These novel connections are highly enriched for chromatin-modifying proteins, hinting at a universal role of chromatin regulation in cancer etiology. Although infrequently mutated as single genes, we show that chromatin modifiers are altered in a large fraction of cancer patients. In summary, we demonstrate that integration of evolutionary signatures is key for identifying mutational driver genes, thereby facilitating the discovery of novel therapeutic targets for cancer treatment.

퍼지집합과 베이지안 확률 기법을 이용한 암설사면지형 분포지역 추출에 관한 연구

퍼지집합과 베이지안 확률 기법을 이용한 암설사면지형 분포지역 추출에 관한 연구

COMPARISON OF ADJACENCY AND DISTANCE-BASED APPROACHES FOR SPATIAL ANALYSIS OF MULTIMODAL TRAFFIC CRASH DATA

Abstract. Many studies have utilized the spatial correlations among traffic crash data to develop crash prediction models with the aim to investigate the influential factors or predict crash counts at different sites. The spatial correlation have been observed to account for heterogeneity in different forms of weight matrices which improves the estimation performance of models. But very rarely have the weight matrices been compared for the prediction accuracy for estimation of crash counts. This study was targeted at the comparison of two different approaches for modelling the spatial correlations among crash data at macro-level (County). Multivariate Full Bayesian crash prediction models were developed using Decay-50 (distance-based) and Queen-1 (adjacency-based) weight matrices for simultaneous estimation crash counts of four different modes: vehicle, motorcycle, bike, and pedestrian. The goodness-of-fit and different criteria for accuracy at prediction of crash count reveled the superiority of Decay-50 over Queen-1. Decay-50 was essentially different from Queen-1 with the selection of neighbors and more robust spatial weight structure which rendered the flexibility to accommodate the spatially correlated crash data. The consistently better performance of Decay-50 at prediction accuracy further bolstered its superiority. Although the data collection efforts to gather centroid distance among counties for Decay-50 may appear to be a downside, but the model has a significant edge to fit the crash data without losing the simplicity of computation of estimated crash count.

Genomic Prediction of Autogamous and Allogamous Plants by SNPs and Haplotypes

The implementation of single‐nucleotide polymorphism (SNP)‐based genomic selection has demonstrated great predictive potential in plants. However, its application is sometimes limited to the biallelism of the marker. In this context, the use of haplotype blocks as multiallelic markers might improve genomic prediction. This study was performed to compare the predictive ability of Bayesian genomic prediction models using haplotypes (confidence interval and four‐gamete), individual SNPs, and sets of SNPs selected according to haplotype construction. The use of haplotype matrices increased the predictive ability and selection coincidence with the phenotypic selection for the maize (Zea mays L.) breeding population. However, this was not observed for the rice (Oryza sativa L.) population, in which the use of the nonreduced SNP matrix was more efficient. Overall, the use of reduced SNP matrices did not lead to better predictive abilities. No difference was observed between the genomic prediction methods used. We found that the use of haplotypes has potential to increase predictive ability of genomic prediction in breeding populations of allogamous plants or plants with high multiallelism.

Perceptual medical image fusion with internal generative mechanism

Medical image fusion is the process of integrating two medical images with a visual enhanced single fused image, to attain a resultant image richer in information to aid medical practitioners in better diagnosis. A perceptual medical image fusion method is proposed by employing Internal Generative Mechanism. First, source images are divided into a predicted layer and a detail layer with a Bayesian prediction model. Then, the detail layer is merged with the energy of Tchebichef moments for blocks while the predicted layer is fused using the averaging strategy as activity level measurement. The fused image is finally obtained by merging coefficients in both fused layers. Experimental results prove that the proposed fusion algorithm is superior to the previously developed methods.

A Machine Learning Decision-Support System Improves the Internet of Things’ Smart Meter Operations

An Internet of Things' (IoT) connected society and system represents a tremendous paradigm shift. We present a framework for a decision-support system (DSS) that operates within the IoT ecosystem. The DSS leverages advanced analytics of electric smart meter (ESM) network communication-quality data to improve cost predictions for smart meter field operations and provide actionable decision recommendations regarding whether to send a technician to a customer location to resolve an ESM issue. The model is empirically evaluated using data sets from a commercial network. We demonstrate the efficiency of our approach with a complete Bayesian network prediction model and compare with three machine learning prediction model classifiers: 1) Naïve Bayes; 2) random forest; and 3) decision tree. Results demonstrate that our approach generates statistically noteworthy estimations and that the DSS will improve the cost efficiency of ESM network operations and maintenance.

Predictive models for pressure ulcers from intensive care unit electronic health records using Bayesian networks

BackgroundWe develop predictive models enabling clinicians to better understand and explore patient clinical data along with risk factors for pressure ulcers in intensive care unit patients from electronic health record data. Identifying accurate risk factors of pressure ulcers is essential to determining appropriate prevention strategies; in this work we examine medication, diagnosis, and traditional Braden pressure ulcer assessment scale measurements as patient features. In order to predict pressure ulcer incidence and better understand the structure of related risk factors, we construct Bayesian networks from patient features. Bayesian network nodes (features) and edges (conditional dependencies) are simplified with statistical network techniques. Upon reviewing a network visualization of our model, our clinician collaborators were able to identify strong relationships between risk factors widely recognized as associated with pressure ulcers.MethodsWe present a three-stage framework for predictive analysis of patient clinical data: 1) Developing electronic health record feature extraction functions with assistance of clinicians, 2) simplifying features, and 3) building Bayesian network predictive models. We evaluate all combinations of Bayesian network models from different search algorithms, scoring functions, prior structure initializations, and sets of features.ResultsFrom the EHRs of 7,717 ICU patients, we construct Bayesian network predictive models from 86 medication, diagnosis, and Braden scale features. Our model not only identifies known and suspected high PU risk factors, but also substantially increases sensitivity of the prediction - nearly three times higher comparing to logistical regression models - without sacrificing the overall accuracy. We visualize a representative model with which our clinician collaborators identify strong relationships between risk factors widely recognized as associated with pressure ulcers.ConclusionsGiven the strong adverse effect of pressure ulcers on patients and the high cost for treating pressure ulcers, our Bayesian network based model provides a novel framework for significantly improving the sensitivity of the prediction model. Thus, when the model is deployed in a clinical setting, the caregivers can suitably respond to conditions likely associated with pressure ulcer incidence.

Bayesian Correlation Prediction Model between Hydrogen-Induced Cracking in Structural Members

Background: A quantitative model was developed and applied for analyzing the correlation between hydrogen-induced corrosion cracking in both main cable wires and degraded stiffening of the girders of a cable suspension bridge, considering maintenance effects across time and space. Method: Bayesian inference is applied for predicting the correlations among the wires in the main cables owed to hydrogen-induced cracking (HIC) in the cable wires of a steel bridge, by using the improved hierarchical Bayesian models proposed here. Results: The simulated risk prediction under decreased strength of cable wires, due to the corrosion cracking, yields posterior distributions based on prior distributions and likelihoods. The Bayesian inference model can be applied to the design and maintenance of highly corroded and correlated components Data are updated through analyzed information from previous crack steps. A numerical example including not only reliability indices but also probabilities of failure for cable wires, damaged by HIC, is then presented. Compared with a conventional linear prediction model, the one herein developed provides highly improved convergence and closeness to the analyzed data. Conclusion: The proposed model can be used as a diagnostic or prognostic prediction tool for the performance of corroded bridge cable wires with crack propagation, allowing the development of maintenance plans for mechanical components and the overall structural system.

Prediction of annual water consumption in Guangdong Province based on Bayesian neural network

In the context of the implementation of the most stringent water resources management system, the role of water demand forecasting for regional water resources management is becoming increasingly significant. Based on the analysis of the influencing factors of water consumption in Guangdong Province, we made the forecast index system of annual water consumption, and constructed the forecast model of annual water consumption of BP neural network, then optimized the regularization BP neural network in utilization rate of water. The results showed that the average absolute percentage error of Bayesian neural network prediction model and BP neural network prediction model is 0.70% and 0.46% respectively. BP neural network model by Bayesian regularization is more ability to improve the accuracy of about 0.24%, more in line with the regional annual water demand forecast high precision requirements. Take the planning index value of Guangdong Province’s thirteen five plan into Bayesian neural network forecasting model, and its forecast value is 45.432 billion cubic meters, which will reach 456.04 billion cubic meters of red water in Guangdong Province in 2020.

Collapse risk evaluation method on Bayesian network prediction model and engineering application

Collapse was one of the typical common geological hazards during the construction of tunnels. The risk assessment of collapse was an effective way to ensure the safety of tunnels. We established a prediction model of collapse based on Bayesian Network. 76 large or medium collapses in China were analyzed. The variable set and range of the model were determined according to the statistics. A collapse prediction software was developed and its veracity was also evaluated. At last the software was used to predict tunnel collapses. It effectively evaded the disaster. Establishing the platform can be subsequent perfect. The platform can also be applied to the risk assessment of other tunnel engineering.

On the Bayesian calibration of computer model mixtures through experimental data, and the design of predictive models

For many real systems, several computer models may exist with different physics and predictive abilities. To achieve more accurate simulations/predictions, it is desirable for these models to be properly combined and calibrated. We propose the Bayesian calibration of computer model mixture method which relies on the idea of representing the real system output as a mixture of the available computer model outputs with unknown input dependent weight functions. The method builds a fully Bayesian predictive model as an emulator for the real system output by combining, weighting, and calibrating the available models in the Bayesian framework. Moreover, it fits a mixture of calibrated computer models that can be used by the domain scientist as a mean to combine the available computer models, in a flexible and principled manner, and perform reliable simulations. It can address realistic cases where one model may be more accurate than the others at different input values because the mixture weights, indicating the contribution of each model, are functions of the input. Inference on the calibration parameters can consider multiple computer models associated with different physics. The method does not require knowledge of the fidelity order of the models. We provide a technique able to mitigate the computational overhead due to the consideration of multiple computer models that is suitable to the mixture model framework. We implement the proposed method in a real-world application involving the Weather Research and Forecasting large-scale climate model.

Developing a Bayesian belief network model for prediction of R&D project success

The project success is critical to the business performance in the era of fierce competition and globalization. The basis for project success lies in the capabilities of managing risks effectively. Innovation has always been considerably risky; however, managing Research and Development (R&D) project risks has become even more important given today’s tight schedules and limited resources. Risk management has to be an integral part of the development process. The purpose of this research is to develop a model to assess and estimate the risk exposure of an R&D project. A risk quantification model based on the Bayesian belief network is proposed, which is effective in capturing the interaction between various risk factors. The aim of this model is to empower the project managers to predict the failure risk probability of R&D projects.

Identification of novel MRP3 inhibitors based on computational models and validation using an in vitro membrane vesicle assay

IntroductionMultidrug resistance-associated protein 3 (MRP3), an efflux transporter on the hepatic basolateral membrane, may function as a compensatory mechanism to prevent the accumulation of anionic substrates (e.g., bile acids) in hepatocytes. Inhibition of MRP3 may disrupt bile acid homeostasis and is one hypothesized risk factor for the development of drug-induced liver injury (DILI). Therefore, identifying potential MRP3 inhibitors could help mitigate the occurrence of DILI. MethodsBayesian models were developed using MRP3 transporter inhibition data for 86 structurally diverse drugs. The compounds were split into training and test sets of 57 and 29 compounds, respectively, and six models were generated based on distinct inhibition thresholds and molecular fingerprint methods. The six Bayesian models were validated against the test set and the model with the highest accuracy was utilized for a virtual screen of 1470 FDA-approved drugs from DrugBank. Compounds that were predicted to be inhibitors were selected for in vitro validation. The ability of these compounds to inhibit MRP3 transport at a concentration of 100μM was measured in membrane vesicles derived from stably transfected MRP3-over-expressing HEK-293 cells with [3H]-estradiol-17β-d-glucuronide (E217G; 10μM; 5min uptake) as the probe substrate. ResultsA predictive Bayesian model was developed with a sensitivity of 73% and specificity of 71% against the test set used to evaluate the six models. The area under the Receiver Operating Characteristic (ROC) curve was 0.710 against the test set. The final selected model was based on compounds that inhibited substrate transport by at least 50% compared to the negative control, and functional-class fingerprints (FCFP) with a circular diameter of six atoms, in addition to one-dimensional physicochemical properties. The in vitro screening of predicted inhibitors and non-inhibitors resulted in similar model performance with a sensitivity of 64% and specificity of 70%. The strongest inhibitors of MRP3-mediated E217G transport were fidaxomicin, suramin, and dronedarone. Kinetic assessment revealed that fidaxomicin was the most potent of these inhibitors (IC50=1.83±0.46μM). Suramin and dronedarone exhibited IC50 values of 3.33±0.41 and 47.44±4.41μM, respectively. ConclusionBayesian models are a useful screening approach to identify potential inhibitors of transport proteins. Novel MRP3 inhibitors were identified by virtual screening using the selected Bayesian model, and MRP3 inhibition was confirmed by an in vitro transporter inhibition assay. Information generated using this modeling approach may be valuable in predicting the potential for DILI and/or MRP3-mediated drug-drug interactions.