GkmExplain

GkmExplain interprets predictive sequence patterns learned by Support Vector Machines (SVMs) using gapped k-mer kernels to attribute nucleotide-level importance within regulatory DNA sequences and transcription factor binding models.

Key Features:

• Interpretation of Complex Models: Attributes sequence features learned by gapped k-mer SVMs (gkm-SVMs), including models with nonlinear components, and addresses limitations of deltaSVM, in-silico mutagenesis (ISM), and SHAP.
• Computational Efficiency: Computes feature attributions substantially faster than SHAP by several orders of magnitude while maintaining accuracy.
• Theoretical Foundation: Has formal connections to Integrated Gradients for feature attribution.
• Accuracy and Reliability: Demonstrates high accuracy in simulations with regulatory DNA sequences and avoids pitfalls associated with deltaSVM and ISM.
• Application to Regulatory Variant Identification: Applied to gkm-SVM models trained on in vivo transcription factor (TF) binding data to recover consolidated, non-redundant TF motifs, and produces mutation impact scores that outperform deltaSVM and ISM in chromatin accessibility models.

Scientific Applications:

• Regulatory DNA Sequence Analysis: Interpreting gapped k-mer patterns to elucidate functional signals in non-coding genomic regions.
• Transcription Factor Binding Studies: Recovering and validating transcription factor motifs from in vivo TF binding gkm-SVM models.
• Genetic Variant Impact Assessment: Scoring mutation impacts to prioritize regulatory genetic variants, including within chromatin accessibility models.

Methodology:

GkmExplain uses a feature attribution approach leveraging theoretical connections to Integrated Gradients to compute nucleotide-level importance scores for gapped k-mer SVM models.