Kolmogorov-Arnold networks for genomic tasks

Area of Science:

• Computational Biology
• Machine Learning
• Genomics

Background:

• Kolmogorov-Arnold networks (KANs) offer a novel architecture as an alternative to multilayer perceptrons (MLPs).
• Previous research has explored KAN integration in computer vision and natural language processing, but not in genomics.

Purpose of the Study:

• To investigate the efficacy of KANs, specifically linear KANs (LKANs) and convolutional KANs (CKANs), as replacements for MLPs in deep learning models for genomic sequence analysis.
• To evaluate their performance in genomic sequence classification and generation tasks.

Main Methods:

• Implemented LKANs and CKANs within baseline deep learning architectures.
• Tested models on three genomic benchmark datasets: Genomic Benchmarks, Genome Understanding Evaluation, and Flipon Benchmark.
• Conducted ablation studies to assess the impact of KAN layer count on performance.

Main Results:

• LKANs demonstrated superior performance compared to baseline models and CKANs across most datasets.
• CKANs achieved comparable results but faced challenges with scalability concerning a large number of parameters.
• Ablation analysis confirmed a positive correlation between the number of KAN layers and model performance.

Conclusions:

• Linear KANs present a promising approach for enhancing deep learning model performance in genomics, particularly with a reduced parameter count.
• Further research is needed to explore the full potential of KANs within various state-of-the-art deep learning architectures for genomics.