Abstract
The cooperativity of transcription factors (TFs) is a widespread phenomenon in the gene regulation system. However, the interaction patterns between TF binding motifs remain elusive. The recent high-throughput assays, CAP-SELEX, have identified over 600 composite DNA sites (i.e. heterodimeric motifs) bound by cooperative TF pairs. However, there are over 25,000 inferentially effective heterodimeric TFs in human cell. It is not practically feasible to validate all heterodimeric motifs due to cost and labour. Therefore, it is highly demanding to develop a fast and accurate computational tool for heterodimeric motif synthesis. We introduce DeepMotifSyn, a deep-learning-based tool for synthesizing heterodimeric motifs from monomeric motif pairs. Specifically, DeepMotifSyn is composed of heterodimeric motif generator and evaluator. The generator is a U-Net-based neural network that can synthesize heterodimeric motifs from aligned motif pairs. The evaluator is a machine-learning-based model that can score the generated heterodimeric motif candidates based on the motif sequence features. Systematic evaluations on CAP-SELEX data illustrates that DeepMotif-Syn significantly outperforms the current state-of-the-art predictors. In addition, DeepMotifSyn can synthesize multiple heterodimeric motifs with different orientation and spacing settings. Such a feature can address the shortcomings of previous models. We believe Deep-MotifSyn is a more practical and reliable model than current predictors on heterodimeric motif synthesis.
Introduction
In the complex regulatory landscape of the cell, transcription factors (TFs) act as critical switches for gene expression. While individual TF binding preferences (monomeric motifs) are well-documented, TFs frequently operate in pairs—forming heterodimers—to bind composite DNA sites. Understanding these interactions is vital for decoding the "operating system" of the genome.
Currently, experimental validation of these heterodimeric motifs via high-throughput assays like CAP-SELEX has identified approximately 600 motifs. However, with over 25,000 potential TF pairings in the human genome, experimental coverage is limited by cost and labor. Previous computational methods often lacked generalizability, struggled with complex overlaps, or failed to account for multiple binding configurations. DeepMotifSyn addresses these gaps by leveraging computer vision architectures to synthesize heterodimeric motifs.
Methods: The DeepMotifSyn Framework
DeepMotifSyn employs a two-stage generative and evaluative pipeline to predict heterodimeric motifs from monomeric pairs. The process begins by generating various alignments of monomeric motifs, varying their orientation and spacing.
1. The Generator (U-Net)
The core of the synthesis process is a U-Net-based convolutional neural network, an architecture typically used for medical image segmentation. The generator takes two aligned monomeric motifs as input. It utilizes a "contracting" path to extract high-level interaction patterns and an "expansive" path to reconstruct the final heterodimeric motif. Skip connections are employed to preserve sequence details from the non-overlapping regions. To handle varying motif lengths, the authors implemented a custom Masked Mean Square Error (MaskMSE) loss function.
2. The Evaluator (XGBoost)
Because TFs can bind in numerous orientations and spacings, the generator produces a large pool of candidates. An XGBoost-based evaluator then filters these candidates. It scores them based on 784 features, including Information Content Matrices (ICM) and positional entropy, to identify the most biologically plausible binding configurations.
Datasets and Evaluation
The study primarily utilizes data from CAP-SELEX experiments, consisting of 614 validated heterodimeric motifs from 313 TF pairs. To test the evaluator's precision, a "negative" set of over 368,000 possible alignments was generated, representing the vast space of non-functional configurations.
Key Results
DeepMotifSyn demonstrates significant improvements over existing state-of-the-art models:
- Accuracy: The U-Net generator achieved an average motif distance of 0.17, outperforming the previous standard (IOHMM, 0.26).
- Recovery: In cross-validation, the top 10 predictions recovered 70% of validated motifs, rising to 90% within the top 30.
- Robustness: The XGBoost evaluator achieved a near-perfect ROC-AUC of 0.995 and a PR-AUC of 0.431, effectively identifying true binding modes within a large pool of candidates.
- Generalization: By training on 45 different DNA-binding families, the model learned a general "DNA grammar," allowing it to predict motifs for families with sparse experimental data.
Conclusion
DeepMotifSyn represents a significant step forward in computational genomics by treating motif synthesis as a spatial generative task. Unlike previous models, it can predict multiple valid binding modes, reflecting the flexibility of protein-DNA interactions. Future work aims to extend this framework to hetero-multimeric motifs involving three or more TFs, further unraveling the complexity of gene regulation.
The software is open-source and available on GitHub.
Summary updated