Treatment of missing data in Bayesian network structure learning: an application to linked biomedical and social survey data.

BACKGROUND: Availability of linked biomedical and social science data has risen dramatically in past decades, facilitating holistic and systems-based analyses. Among these, Bayesian networks have great potential to tackle complex interdisciplinary problems, because they can easily model inter-relations between variables. They work by encoding conditional independence relationships discovered via advanced inference algorithms. One challenge is dealing with missing data, ubiquitous in survey or biomedical datasets. Missing data is rarely addressed in an advanced way in Bayesian networks; the most common approach is to discard all samples containing missing measurements. This can lead to biased estimates. Here, we examine how Bayesian network structure learning can incorporate missing data. METHODS: We use a simulation approach to compare a commonly used method in frequentist statistics, multiple imputation by chained equations (MICE), with one specific for Bayesian network learning, structural expectation-maximization (SEM). We simulate multiple incomplete categorical (discrete) data sets with different missingness mechanisms, variable numbers, data amount, and missingness proportions. We evaluate performance of MICE and SEM in capturing network structure. We then apply SEM combined with community analysis to a real-world dataset of linked biomedical and social data to investigate associations between socio-demographic factors and multiple chronic conditions in the US elderly population. RESULTS: We find that applying either method (MICE or SEM) provides better structure recovery than doing nothing, and SEM in general outperforms MICE. This finding is robust across missingness mechanisms, variable numbers, data amount and missingness proportions. We also find that imputed data from SEM is more accurate than from MICE. Our real-world application recovers known inter-relationships among socio-demographic factors and common multimorbidities. This network analysis also highlights potential areas of investigation, such as links between cancer and cognitive impairment and disconnect between self-assessed memory decline and standard cognitive impairment measurement. CONCLUSION: Our simulation results suggest taking advantage of the additional information provided by network structure during SEM improves the performance of Bayesian networks; this might be especially useful for social science and other interdisciplinary analyses. Our case study show that comorbidities of different diseases interact with each other and are closely associated with socio-demographic factors.

Dynamic modulation of phosphoprotein expression in ovarian cancer xenograft models.

BACKGROUND: The dynamic changes that occur in protein expression after treatment of a cancer in vivo are poorly described. In this study we measure the effect of chemotherapy over time on the expression of a panel of proteins in ovarian cancer xenograft models. The objective was to identify phosphoprotein and other protein changes indicative of pathway activation that might link with drug response. METHODS: Two xenograft models, platinum-responsive OV1002 and platinum-unresponsive HOX424, were used. Treatments were carboplatin and carboplatin-paclitaxel. Expression of 49 proteins over 14 days post treatment was measured by quantitative immunofluorescence and analysed by AQUA. RESULTS: Carboplatin treatment in the platinum-sensitive OV1002 model triggered up-regulation of cell cycle, mTOR and DDR pathways, while at late time points WNT, invasion, EMT and MAPK pathways were modulated. Estrogen receptor-alpha (ESR1) and ERBB pathways were down-regulated early, within 24 h from treatment administration. Combined carboplatin-paclitaxel treatment triggered a more extensive response in the OV1002 model modulating expression of 23 of 49 proteins. Therefore the cell cycle and DDR pathways showed similar or more pronounced changes than with carboplatin alone. In addition to expression of pS6 and pERK increasing, components of the AKT pathway were modulated with pAKT increasing while its regulator PTEN was down-regulated early. WNT signaling, EMT and invasion markers were modulated at later time points. Additional pathways were also observed with the NFκB and JAK/STAT pathways being up-regulated. ESR1 was down-regulated as was HER4, while further protein members of the ERBB pathway were upregulated late. By contrast, in the carboplatin-unresponsive HOX 424 xenograft, carboplatin only modulated expression of MLH1 while carboplatin-paclitaxel treatment modulated ESR1 and pMET. CONCLUSIONS: Thirteen proteins were modulated by carboplatin and a more robust set of changes by carboplatin-paclitaxel. Early changes included DDR and cell cycle regulatory proteins associating with tumor volume changes, as expected. Changes in ESR1 and ERBB signaling were also observed. Late changes included components of MAPK signaling, EMT and invasion markers and coincided in time with reversal in tumor volume reduction. These results suggest potential therapeutic roles for inhibitors of such pathways that may prolong chemotherapeutic effects.

Novel Monte Carlo approach quantifies data assemblage utility and reveals power of integrating molecular and clinical information for cancer prognosis.

Current clinical practice in cancer stratifies patients based on tumour histology to determine prognosis. Molecular profiling has been hailed as the path towards personalised care, but molecular data are still typically analysed independently of known clinical information. Conventional clinical and histopathological data, if used, are added only to improve a molecular prediction, placing a high burden upon molecular data to be informative in isolation. Here, we develop a novel Monte Carlo analysis to evaluate the usefulness of data assemblages. We applied our analysis to varying assemblages of clinical data and molecular data in an ovarian cancer dataset, evaluating their ability to discriminate one-year progression-free survival (PFS) and three-year overall survival (OS). We found that Cox proportional hazard regression models based on both data types together provided greater discriminative ability than either alone. In particular, we show that proteomics data assemblages that alone were uninformative (p = 0.245 for PFS, p = 0.526 for OS) became informative when combined with clinical information (p = 0.022 for PFS, p = 0.048 for OS). Thus, concurrent analysis of clinical and molecular data enables exploitation of prognosis-relevant information that may not be accessible from independent analysis of these data types.

Relationship between differentially expressed mRNA and mRNA-protein correlations in a xenograft model system.

Differential mRNA expression studies implicitly assume that changes in mRNA expression have biological meaning, most likely mediated by corresponding changes in protein levels. Yet studies into mRNA-protein correspondence have shown notoriously poor correlation between mRNA and protein expression levels, creating concern for inferences from only mRNA expression data. However, none of these studies have examined in particular differentially expressed mRNA. Here, we examined this question in an ovarian cancer xenograft model. We measured protein and mRNA expression for twenty-nine genes in four drug-treatment conditions and in untreated controls. We identified mRNAs differentially expressed between drug-treated xenografts and controls, then analysed mRNA-protein expression correlation across a five-point time-course within each of the four experimental conditions. We evaluated correlations between mRNAs and their protein products for mRNAs differentially expressed within an experimental condition compared to those that are not. We found that differentially expressed mRNAs correlate significantly better with their protein product than non-differentially expressed mRNAs. This result increases confidence for the use of differential mRNA expression for biological discovery in this system, as well as providing optimism for the usefulness of inferences from mRNA expression in general.

An analytical approach differentiates between individual and collective cancer invasion.

Tumour cells employ a variety of mechanisms to invade their environment and to form metastases. An important property is the ability of tumour cells to transition between individual cell invasive mode and collective mode. The switch from collective to individual cell invasion in the breast was shown recently to determine site of subsequent metastasis. Previous studies have suggested a range of invasion modes from single cells to large clusters. Here, we use a novel image analysis method to quantify and categorise invasion. We have developed a process using automated imaging for data collection, unsupervised morphological examination of breast cancer invasion using cognition network technology (CNT) to determine how many patterns of invasion can be reliably discriminated. We used Bayesian network analysis to probabilistically connect morphological variables and therefore determine that two categories of invasion are clearly distinct from one another. The Bayesian network separated individual and collective invading cell groups based on the morphological measurements, with the level of cell-cell contact the most discriminating morphological feature. Smaller invading groups were typified by smoother cellular surfaces than those invading collectively in larger groups. Interestingly, elongation was evident in all invading cell groups and was not a specific feature of single cell invasion as a surrogate of epithelial-mesenchymal transition. In conclusion, the combination of cognition network technology and Bayesian network analysis provides an insight into morphological variables associated with transition of cancer cells between invasion modes. We show that only two morphologically distinct modes of invasion exist.

Cancer systems biology.

Cancer is a complex and heterogeneous disease, not only at a genetic and biochemical level, but also at a tissue, organism, and population level. Multiple data streams, from reductionist biochemistry in vitro to high-throughput "-omics" from clinical material, have been generated with the hope that they encode useful information about phenotype and, ultimately, tumour behaviour in response to drugs. While these data stand alone in terms of the biology they represent, there is the enticing prospect that if incorporated into systems biology models, they can help understand complex systems behaviour and provide a predictive framework as an additional tool in understanding how tumours change and respond to treatment over time. Since these biological data are heterogeneous and frequently qualitative rather than quantitative, at the present time a single systems biology approach is unlikely to be effective; instead, different computational and mathematical approaches should be tailored to different types of data, and to each other, in order to test and re-test hypotheses. In time, these models might converge and result in usable tractable models which accurately represent human cancer. Likewise, biologists and clinicians need to understand what the requirements of systems biology are so that compatible data are produced for computational modelling. In this review, we describe some theoretical approaches (data-driven and process-driven) and experimental methodologies which are being used in cancer research and the clinical context where they might be applied.