The impact of different negative training data on regulatory sequence predictions

Area of Science:

• Genomics
• Computational Biology
• Bioinformatics

Background:

• Regulatory DNA elements (promoters, enhancers) are vital for gene regulation and human variation.
• Understanding the functional encoding of these regions is limited.
• Machine learning, including gapped k-mer support vector machines (gkm-SVMs) and convolutional neural networks (CNNs), shows promise for deciphering DNA encoding.

Purpose of the Study:

• To investigate the impact of negative sequence selection on the performance of machine learning models trained on regulatory DNA.
• To compare different strategies for generating negative training datasets.

Main Methods:

• Trained gapped k-mer support vector machines (gkm-SVMs) and convolutional neural networks (CNNs) on open chromatin data.
• Compared two negative training data approaches: genomic background sequences and sequence shuffles of positive sequences.
• Evaluated model performance on predicting cell-type activity, cell-type specificity, and quantitative activity.

Main Results:

• Genomic background sequences generally led to better model performance compared to shuffled sequences.
• Models trained on highly shuffled sequences underperformed on complex prediction tasks and learned artificial sequence features.
• CNNs outperformed gkm-SVMs on larger datasets, while gkm-SVMs provided robust results for typical dataset sizes without extensive tuning.

Conclusions:

• Negative sequence selection significantly impacts the performance of machine learning models for regulatory element prediction.
• Genomic background sequences are a more effective choice for negative training data.
• The choice of model and negative data strategy should consider dataset size and prediction task complexity.