A Pipeline for Integrated Theory and Data-Driven Modeling of Biomedical Data

Abstract

Genome sequencing technologies have the potential to transform clinical decision making and biomedical research by enabling high-throughput measurements of the genome at a granular level. However, to truly understand mechanisms of disease and predict the effects of medical interventions, high-throughput data must be integrated with demographic, phenotypic, environmental, and behavioral data from individuals. Further, effective knowledge discovery methods must infer relationships between these data types. We recently proposed a pipeline (CausalMGM) to achieve this. CausalMGM uses probabilistic graphical models to infer the relationships between variables in the data; however, CausalMGM's graphical structure learning algorithm can only handle small datasets efficiently. We propose a new methodology (piPref-Div) that selects the most informative variables for CausalMGM, enabling it to scale. We validate the efficacy of piPref-Div against other feature selection methods and demonstrate how the use of the full pipeline improves breast cancer outcome prediction and provides biologically interpretable views of gene expression data.

Fig. 1:

Pipeline proposed in this work to learn graphical model structure from mixed clinical and omics datasets.

Fig. 2:

Illustration of procedure to limit tested parameter range. Figure originally appeared in [7]

Fig. 3:

Subsampling procedure to determine empirical probabilities for every edge in the correlation graph. B (λ, S) returns a correlation graph computed upon dataset S with threshold λ. Figure originally appeared in [7]

Fig. 4:

Cluster Simulation to generate simulated datasets. Purple nodes are master regulators of a cluster, blue nodes are causal parents of the target variable, and the beige node is the target variable.

Fig. 5:

Heatmap of correlation of prior knowledge between sources. Each cell is the percentage of gene-gene pairs in the prior source of the row that are also in the prior source in the column.

Fig. 6:

Heatmap of overlapping prior knowledge between sources. Each cell is the correlation between the probabilities given by each source for all gene-gene pairs in the prior source of the row that are also in the prior source in the column.

Fig. 7:

Predicted Weight vs. Net Reliability for each prior knowledge source in simulated experiments for piPref-Div for (left) 50 samples and (right) 200 samples.

Fig. 8:

Accuracy of predicted clusters for varying amount and reliability of prior knowledge. Sample size was set to 50 (left) and 200 (right). Star represents the best possible performance if optimal correlation thresholds were selected.

Fig. 9:

Accuracy of predicted clusters for varying amount and reliability of prior knowledge on large datasets. Sample size was set to 50 (left) and 200 (right).

Fig. 10:

AUC of predicting five-year metastasis-free and relapse-free survival using several feature selection methods on six independent breast cancer microarray datasets.

Fig. 11:

Stability of learned models for predicting five-year metastasis-free and relapse-free survival using several feature selection methods on six independent breast cancer microarray datasets.

Fig. 12:

Graphical model of breast cancer subtype. Size of each edge represents the number of times a similar cluster was selected to be related to Subtype in each of the cross-validation folds.