Gene Regulatory Network Inference via Conditional Inference Trees and Forests

Trees are classical data structures allowing effectively classifying and predicting responses. Due to versatility and high performance in classification and prediction, there exist plenty of tree-based methods including popular Conditional Inference Tree (CIT) and Forests (CIF), Random Forests (RF), Randomized Trees (RT), randomized C4.5, etc. In this work we assessed the performance of CIT and CIF methods in correct gene regulatory network (GRN) prediction from expression data by using reference golden standard built from real transcriptional regulatory network of E. coli. The synthetic microarray expression data was obtained from DREAM4 challenge. The performance of each network inference method was assessed via Area Under Receiver Operating Characteristic (AUROC) and Area Under Precision Recall (AUPR) metrics.
Our preliminary results show that CIT and CIF successfully predict directed GRNs at acceptable performance rates although not optimal (the best AUROC at 0.68 and AUPR at 0.13 for CIF and the best AUROC at 0.58 and AUPR at 0.18 for CIT). Surprisingly by using the current aggregation scheme of feature importance that prefers features with the highest number of observations, a single CIT was a better performer compared to CIFs in all 5 networks. Nevertheless, the CIFs showed an overall 10% improvement in AUROC. A single CIT has 24% and CIFs have 27% lower overall performance compared to the best performer of DREAM4 Challenge based on cumulative areas of PR and ROC curves.
We plan to test other feature importance aggregation techniques in a single tree and in tree ensembles in order to outperform the top DREAM4 algorithms. In addition the effects of expression data standardization to unit variance will be presented. In future, the developed CIF framework will be used to perform data integration analysis of multi-omics datasets.