Predicting gene expression from sequence: a reexamination

Abstract

Although much of the information regarding genes' expressions is encoded in the genome, deciphering such information has been very challenging. We reexamined Beer and Tavazoie's (BT) approach to predict mRNA expression patterns of 2,587 genes in Saccharomyces cerevisiae from the information in their respective promoter sequences. Instead of fitting complex Bayesian network models, we trained naïve Bayes classifiers using only the sequence-motif matching scores provided by BT. Our simple models correctly predict expression patterns for 79% of the genes, based on the same criterion and the same cross-validation (CV) procedure as BT, which compares favorably to the 73% accuracy of BT. The fact that our approach did not use position and orientation information of the predicted binding sites but achieved a higher prediction accuracy, motivated us to investigate a few biological predictions made by BT. We found that some of their predictions, especially those related to motif orientations and positions, are at best circumstantial. For example, the combinatorial rules suggested by BT for the PAC and RRPE motifs are not unique to the cluster of genes from which the predictive model was inferred, and there are simpler rules that are statistically more significant than BT's ones. We also show that CV procedure used by BT to estimate their method's prediction accuracy is inappropriate and may have overestimated the prediction accuracy by about 10%.

Figure 1. Training and Test Set Classification Accuracy for Naïve Bayes Method Using Motif Scores Only

Classification accuracies for training sets increases with the number of top motifs selected in models, while test set accuracies only increase when model sizes are small. Including too many features will overfit the training set and thus decrease the test set accuracies. 100 random repeats of 5-fold CVs were performed, and the curves display the mean accuracies. The error bars denote the maximum and minimum accuracy achieved in the 100 random repeats.

Figure 2. Motif Logos of M198 and RAP1

These two TFBMs are very similar, except that M198 is one position longer than RAP1 on the right end. Compared to RAP1, M198 can help distinguish genes in cluster 1 from other genes in a higher statistical significance, without using any position or orientation constraints.