Understanding why only some breast cancers become metastatic, and predicting metastatic risks at earlier cancer stages, are two major goals of breast cancer research. These goals are clearly connected and synergistic, but are rarely integrated within a single study. Knowledge discovery-oriented research has identified critical biological pathways related to metastasis, but the interplay between these pathways remains elusive, resulting in models that lack prediction accuracy. Conversely, complex machine learning models achieve high prediction accuracy without detailed biological knowledge, making the explanation challenging. Here, we propose a novel computational framework that addresses both goals simultaneously. To support knowledge discovery, we construct gene regulatory networks (GRNs) to model the cellular states of metastatic and non-metastatic patients. To enable explainable metastasis prediction, we introduce a dysregulation score based on the GRN models. Experimental results demonstrate that our method not only identified significant metastasis-associated GRN changes, but also revealed the loss of co-regulation among key biological processes in metastatic patients. Leveraging the dysregulation score, our model-free classifier outperformed complex machine learning models under rigorous evaluation. This work bridges a significant gap between knowledge discovery and accurate, explainable prediction by employing carefully designed knowledge models and knowledge-based prediction, with potential applicability in other disease contexts.

Interpretability of machine learning (ML) models represents the extent to which a model's decision-making process can be understood by model developers and/or end users. Transcriptomics-based cancer prognosis models, for example, while achieving good accuracy, are usually hard to interpret, due to the high-dimensional feature space and the complexity of models. As interpretability is critical for the transparency and fairness of ML models, several algorithms have been proposed to improve the interpretability of arbitrary classifiers. However, evaluation of these algorithms often requires substantial domain knowledge. Here, we propose a breast cancer metastasis prediction model using a very small number of biologically interpretable features, and a simple yet novel model interpretation approach that can provide personalized interpretations. In addition, we contributed, to the best of our knowledge, the first method to quantitatively compare different interpretation algorithms. Experimental results show that our model not only achieved competitive prediction accuracy, but also higher inter-classifier interpretation consistency than state-of-the-art interpretation methods. Importantly, our interpretation results can improve the generalizability of the prediction models. Overall, this work provides several novel ideas to construct and evaluate interpretable ML models that can be valuable to both the cancer machine learning community and related application domains.

MOTIVATION: Metastasis prediction is a well-known problem in breast cancer research. As breast cancer is a complex and heterogeneous disease with many molecular subtypes, predictive models trained for one cohort often perform poorly on other cohorts, and a combined model may be suboptimal for individual patients. Furthermore, attempting to develop subtype-specific models is hindered by the ambiguity and stereotypical definitions of subtypes. RESULTS: Here, we propose a personalized approach by relaxing the definition of breast cancer subtypes. We assume that each patient belongs to a distinct subtype, defined implicitly by a set of patients with similar molecular characteristics, and construct a different predictive model for each patient, using as training data, only the patients defining the subtype. To increase robustness, we also develop a committee-based prediction method by pooling together multiple personalized models. Using both intra- and inter-dataset validations, we show that our approach can significantly improve the prediction accuracy of breast cancer metastasis compared with several popular approaches, especially on those hard-to-learn cases. Furthermore, we find that breast cancer patients belonging to different canonical subtypes tend to have different predictive models and gene signatures, suggesting that metastasis in different canonical subtypes are likely governed by different molecular mechanisms. AVAILABILITY AND IMPLEMENTATION: Source code implemented in MATLAB and Java available at www.cs.utsa.edu/∼jruan/PCC/.

DNA methylation is a common epigenetic marker that regulates gene expression. A robust and cost-effective way for measuring whole genome methylation is Methyl-CpG binding domain-based capture followed by sequencing (MBDCap-seq). In this study, we proposed BIMMER, a Hidden Markov Model (HMM) for differential Methylation Regions (DMRs) identification, where HMMs were proposed to model the methylation status in normal and cancer samples in the first layer and another HMM was introduced to model the relationship between differential methylation and methylation statuses in normal and cancer samples. To carry out the prediction for BIMMER, an Expectation-Maximization algorithm was derived. BIMMER was validated on the simulated data and applied to real MBDCap-seq data of normal and cancer samples. BIMMER revealed that 8.83% of the breast cancer genome are differentially methylated and the majority are hypo-methylated in breast cancer.

Methyl-CpG binding domain-based capture followed by sequencing (MBDCap-seq) is a cost-effective method for genome-wide methylation analyses especially in CpG-rich regions. In this study, we developed BIMMER, a BI-layer hidden Markov model for differential Methylation Regions (DMRs) identification BIMMER using MBDCap-seq samples derived from two different phenotypes. BIMMER models and generates a posterior probability for a 100bp bin to be a methylation site in either normal or disease samples by its first hidden layer, and then integrate these posterior probabilities in the second hidden layer to obtain the posterior probability of bin-specific differential methylation between the normal and disease samples. Based on these posterior probabilities, the decisions on the methylation and differential statuses for each bin can be calculated. Simulated results showed 94.3% area under precision-recall curve for BIMMER (BIMMER is programmed in Java and available by request).

BACKGROUND: RNA polymerase II (PolII) is essential in gene transcription and ChIP-seq experiments have been used to study PolII binding patterns over the entire genome. However, since PolII enriched regions in the genome can be very long, existing peak finding algorithms for ChIP-seq data are not adequate for identifying such long regions. METHODS: Here we propose an enriched region detection method for ChIP-seq data to identify long enriched regions by combining a signal denoising algorithm with a false discovery rate (FDR) approach. The binned ChIP-seq data for PolII are first processed using a non-local means (NL-means) algorithm for purposes of denoising. Then, a FDR approach is developed to determine the threshold for marking enriched regions in the binned histogram. RESULTS: We first test our method using a public PolII ChIP-seq dataset and compare our results with published results obtained using the published algorithm HPeak. Our results show a high consistency with the published results (80-100%). Then, we apply our proposed method on PolII ChIP-seq data generated in our own study on the effects of hormone on the breast cancer cell line MCF7. The results demonstrate that our method can effectively identify long enriched regions in ChIP-seq datasets. Specifically, pertaining to MCF7 control samples we identified 5,911 segments with length of at least 4 Kbp (maximum 233,000 bp); and in MCF7 treated with E2 samples, we identified 6,200 such segments (maximum 325,000 bp). CONCLUSIONS: We demonstrated the effectiveness of this method in studying binding patterns of PolII in cancer cells which enables further deep analysis in transcription regulation and epigenetics. Our method complements existing peak detection algorithms for ChIP-seq experiments.

SUMMARY: Differential Identification using Mixtures Ensemble (DIME) is a package for identification of biologically significant differential binding sites between two conditions using ChIP-seq data. It considers a collection of finite mixture models combined with a false discovery rate (FDR) criterion to find statistically significant regions. This leads to a more reliable assessment of differential binding sites based on a statistical approach. In addition to ChIP-seq, DIME is also applicable to data from other high-throughput platforms. AVAILABILITY AND IMPLEMENTATION: DIME is implemented as an R-package, which is available at http://www.stat.osu.edu/~statgen/SOFTWARE/DIME. It may also be downloaded from http://cran.r-project.org/web/packages/DIME/.

BACKGROUND: DNA methylation has been shown to play an important role in the silencing of tumor suppressor genes in various tumor types. In order to have a system-wide understanding of the methylation changes that occur in tumors, we have developed a differential methylation hybridization (DMH) protocol that can simultaneously assay the methylation status of all known CpG islands (CGIs) using microarray technologies. A large percentage of signals obtained from microarrays can be attributed to various measurable and unmeasurable confounding factors unrelated to the biological question at hand. In order to correct the bias due to noise, we first implemented a quantile regression model, with a quantile level equal to 75%, to identify hypermethylated CGIs in an earlier work. As a proof of concept, we applied this model to methylation microarray data generated from breast cancer cell lines. However, we were unsure whether 75% was the best quantile level for identifying hypermethylated CGIs. In this paper, we attempt to determine which quantile level should be used to identify hypermethylated CGIs and their associated genes. RESULTS: We introduce three statistical measurements to compare the performance of the proposed quantile regression model at different quantile levels (95%, 90%, 85%, 80%, 75%, 70%, 65%, 60%), using known methylated genes and unmethylated housekeeping genes reported in breast cancer cell lines and ovarian cancer patients. Our results show that the quantile levels ranging from 80% to 90% are better at identifying known methylated and unmethylated genes. CONCLUSIONS: In this paper, we propose to use a quantile regression model to identify hypermethylated CGIs by incorporating probe effects to account for noise due to unmeasurable factors. Our model can efficiently identify hypermethylated CGIs in both breast and ovarian cancer data.

Enabling data analysis in large data depositories for high throughput experimental data such as gene microarrays and ChIP-seq is challenging. In this paper, we discuss three methods for integrating QUEST, a data depository for epigenetic experiments, with a web-based data analysis platform GenePattern. These methods are universal and can serve as an exemplary implementation resolving the dilemma facing many similar database systems in integrating data analysis tools.

MOTIVATION: Antibody-based Chromatin Immunoprecipitation assay followed by high-throughput sequencing technology (ChIP-seq) is a relatively new method to study the binding patterns of specific protein molecules over the entire genome. ChIP-seq technology allows scientist to get more comprehensive results in shorter time. Here, we present a non-linear normalization algorithm and a mixture modeling method for comparing ChIP-seq data from multiple samples and characterizing genes based on their RNA polymerase II (Pol II) binding patterns. RESULTS: We apply a two-step non-linear normalization method based on locally weighted regression (LOESS) approach to compare ChIP-seq data across multiple samples and model the difference using an Exponential-Normal(K) mixture model. Fitted model is used to identify genes associated with differential binding sites based on local false discovery rate (fdr). These genes are then standardized and hierarchically clustered to characterize their Pol II binding patterns. As a case study, we apply the analysis procedure comparing normal breast cancer (MCF7) to tamoxifen-resistant (OHT) cell line. We find enriched regions that are associated with cancer (P < 0.0001). Our findings also imply that there may be a dysregulation of cell cycle and gene expression control pathways in the tamoxifen-resistant cells. These results show that the non-linear normalization method can be used to analyze ChIP-seq data across multiple samples. AVAILABILITY: Data are available at http://www.bmi.osu.edu/~khuang/Data/ChIP/RNAPII/.

BACKGROUND: DNA methylation plays an important role in the process of tumorigenesis. Identifying differentially methylated genes or CpG islands (CGIs) associated with genes between two tumor subtypes is thus an important biological question. The methylation status of all CGIs in the whole genome can be assayed with differential methylation hybridization (DMH) microarrays. However, patient samples or cell lines are heterogeneous, so their methylation pattern may be very different. In addition, neighboring probes at each CGI are correlated. How these factors affect the analysis of DMH data is unknown. RESULTS: We propose a new method for identifying differentially methylated (DM) genes by identifying the associated DM CGI(s). At each CGI, we implement four different mixed effect and generalized least square models to identify DM genes between two groups. We compare four models with a simple least square regression model to study the impact of incorporating random effects and correlations. CONCLUSIONS: We demonstrate that the inclusion (or exclusion) of random effects and the choice of correlation structures can significantly affect the results of the data analysis. We also assess the false discovery rate of different models using CGIs associated with housekeeping genes.

MOTIVATION: In the human genome, 'CpG islands', CG-rich regions located in or near gene promoters, are normally unmethylated. However, in cancer cells, CpG islands frequently gain methylation, resulting in silencing of growth-limiting tumor suppressor genes. To our knowledge, the potential relationship between CpG island hypermethylation, transcription factor (TF) binding in local promoter regions and transcriptional control has not been previously explored in a genome-wide context. RESULTS: In this study, we utilized bioinformatics tools and TF binding site(TFBs) databases to globally analyze sequences methylated in a laboratory model for the development of drug-resistant cancer. Our results demonstrated that four TFBS were enriched in hypermethylated sequences. More interestingly, overrepresentation of these TFBS was observed in hyper-/hypo-methylated sequences where signi.cant changes in methylation levels were observed in drug-resistant cancer cells. In summary, we believe that these.ndings offer a means to further explore the relationship between DNA methylation and gene expression in drug resistance and tumorigenesis.

BACKGROUND: Hypermethylation of promoter CpG islands is strongly correlated to transcriptional gene silencing and epigenetic maintenance of the silenced state. As well as its role in tumor development, CpG island methylation contributes to the acquisition of resistance to chemotherapy. Differential Methylation Hybridisation (DMH) is one technique used for genome-wide DNA methylation analysis. The study of such microarray data sets should ideally account for the specific biological features of DNA methylation and the non-symmetrical distribution of the ratios of unmethylated and methylated sequences hybridised on the array. We have therefore developed a novel algorithm tailored to this type of data, Methylation Linear Discriminant Analysis (MLDA). RESULTS: MLDA was programmed in R (version 2.7.0) and the package is available at CRAN 1. This approach utilizes linear regression models of non-normalised hybridisation data to define methylation status. Log-transformed signal intensities of unmethylated controls on the microarray are used as a reference. The signal intensities of DNA samples digested with methylation sensitive restriction enzymes and mock digested are then transformed to the likelihood of a locus being methylated using this reference. We tested the ability of MLDA to identify loci differentially methylated as analysed by DMH between cisplatin sensitive and resistant ovarian cancer cell lines. MLDA identified 115 differentially methylated loci and 23 out of 26 of these loci have been independently validated by Methylation Specific PCR and/or bisulphite pyrosequencing. CONCLUSION: MLDA has advantages for analyzing methylation data from CpG island microarrays, since there is a clear rational for the definition of methylation status, it uses DMH data without between-group normalisation and is less influenced by cross-hybridisation of loci. The MLDA algorithm successfully identified differentially methylated loci between two classes of samples analysed by DMH using CpG island microarrays.

BACKGROUND: Non-biological signal (or noise) has been the bane of microarray analysis. Hybridization effects related to probe-sequence composition and DNA dye-probe interactions have been observed in differential methylation hybridization (DMH) microarray experiments as well as other effects inherent to the DMH protocol. RESULTS: We suggest two models to correct for non-biologically relevant probe signal with an overarching focus on probe-sequence composition. The estimated effects are evaluated and the strengths of the models are considered in the context of DMH analyses. CONCLUSION: The majority of estimated parameters were statistically significant in all considered models. Model selection for signal correction is based on interpretation of the estimated values and their biological significance.

BACKGROUND: In order to recapitulate tumor progression pathways using epigenetic data, we developed novel clustering and pathway reconstruction algorithms, collectively referred to as heritable clustering. This approach generates a progression model of altered DNA methylation from tumor tissues diagnosed at different developmental stages. The samples act as surrogates for natural progression in breast cancer and allow the algorithm to uncover distinct epigenotypes that describe the molecular events underlying this process. Furthermore, our likelihood-based clustering algorithm has great flexibility, allowing for incomplete epigenotype or clinical phenotype data and also permitting dependencies among variables. RESULTS: Using this heritable clustering approach, we analyzed methylation data obtained from 86 primary breast cancers to recapitulate pathways of breast tumor progression. Detailed annotation and interpretation are provided to the optimal pathway recapitulated. The result confirms the previous observation that aggressive tumors tend to exhibit higher levels of promoter hypermethylation. CONCLUSION: Our results indicate that the proposed heritable clustering algorithms are a useful tool for stratifying both methylation and clinical variables of breast cancer. The application to the breast tumor data illustrates that this approach can select meaningful progression models which may aid the interpretation of pathways having biological and clinical significance. Furthermore, the framework allows for other types of biological data, such as microarray gene expression or array CGH data, to be integrated.

MOTIVATION: To detect and select patterns of transcription factor binding sites (TFBSs) which distinguish genes directly regulated by estrogen receptor-alpha (ERalpha), we developed an innovative mixture model-based discriminate analysis for identifying ordered TFBS pairs. RESULTS: Biologically, our proposed new algorithm clearly suggests that TFBSs are not randomly distributed within ERalpha target promoters (P-value < 0.001). The up-regulated targets significantly (P-value < 0.01) possess TFBS pairs, (DBP, MYC), (DBP, MYC/MAX heterodimer), (DBP, USF2) and (DBP, MYOGENIN); and down-regulated ERalpha target genes significantly (P-value < 0.01) possess TFBS pairs, such as (DBP, c-ETS1-68), (DBP, USF2) and (DBP, MYOGENIN). Statistically, our proposed mixture model-based discriminate analysis can simultaneously perform TFBS pattern recognition, TFBS pattern selection, and target class prediction; such integrative power cannot be achieved by current methods. AVAILABILITY: The software is available on request from the authors. CONTACT: [email protected] SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

One of the longest standing problems in medical image analysis is that of the automated recognition of chromosomes from images of a metaphase spread of a cell. This process of visualizing and categorizing the chromosomes within a cell, called karyotyping, is a key factor in many medical procedures. It is a labor intensive activity, and hence, is a great candidate for automation. However, there are many sources of uncertainty in this problem domain, making complete karyotyping a difficult problem. We describe how fuzzy logic is being inserted into a complete karyotyping system to deal with uncertainty in similar chromosome classes.>