WSMD: weakly-supervised motif discovery in transcription factor ChIP-seq data

Abstract

Although discriminative motif discovery (DMD) methods are promising for eliciting motifs from high-throughput experimental data, due to consideration of computational expense, most of existing DMD methods have to choose approximate schemes that greatly restrict the search space, leading to significant loss of predictive accuracy. In this paper, we propose Weakly-Supervised Motif Discovery (WSMD) to discover motifs from ChIP-seq datasets. In contrast to the learning strategies adopted by previous DMD methods, WSMD allows a "global" optimization scheme of the motif parameters in continuous space, thereby reducing the information loss of model representation and improving the quality of resultant motifs. Meanwhile, by exploiting the connection between DMD framework and existing weakly supervised learning (WSL) technologies, we also present highly scalable learning strategies for the proposed method. The experimental results on both real ChIP-seq datasets and synthetic datasets show that WSMD substantially outperforms former DMD methods (including DREME, HOMER, XXmotif, motifRG and DECOD) in terms of predictive accuracy, while also achieving a competitive computational speed.

Figure 1

An overview of object detection and discriminative motif discovery. (a) Object detection. (b) Discriminative motif discovery.

Figure 2

3-fold cross-validation test performance on three reference-free evaluation criteria over 134 datasets. The performances of six methods on same dataset were plotted on one horizontal bar while differing in colors. In this way, the lines with different colors in one horizontal bar present the performance archived by corresponding tools, and the height of box with different colors can show the performance improvement of corresponding tools compared with the one performing more poorly. (a) 3-fold cross-validation test performance on AUC over 134 datasets. (b) 3-fold cross-validation test performance on Fisher’s Exact Test score over 134 datasets. (c) 3-fold cross-validation test performance on Minimal Hyper-Geometric score over 134 datasets.

Figure 3

Performance comparison of different refinement and extension strategies. For each IC value, we show the average performance obtained by using each tools over 10 distinct synthetic datasets. (a) Performance comparison of different refinement strategies. (b) Performance comparison of different extension strategies.

Figure 4

Comparison of running time (seconds) for DREME, HOMER, motifRG, DECOD and WSMD.