DNALongBench: A Benchmark Suite for Long-Range DNA Prediction Tasks.

Area of Science:

• Genomics
• Computational Biology
• Bioinformatics

Background:

• Modeling long-range DNA dependencies is vital for understanding genome structure and function.
• Current methods struggle to capture extensive dependencies, spanning millions of base pairs, crucial for tasks like 3D chromatin folding.
• A lack of comprehensive benchmarks hinders the evaluation of models addressing these long-range dependencies.

Purpose of the Study:

• Introduce DNALongBench, a novel benchmark dataset designed to evaluate deep learning models on genomics tasks involving long-range DNA dependencies.
• Facilitate standardized comparisons and rigorous assessments of emerging DNA sequence-based models.

Main Methods:

• DNALongBench includes five key genomics tasks: enhancer-target gene interaction, expression quantitative trait loci, 3D genome organization, regulatory sequence activity, and transcription initiation signals.
• The benchmark considers DNA dependencies up to 1 million base pairs.
• Evaluated five methods: a task-specific expert model, a convolutional neural network (CNN), and three fine-tuned DNA foundation models (HyenaDNA, Caduceus-Ph, Caduceus-PS).

Main Results:

• Performance of the evaluated methods on DNALongBench was assessed.
• This evaluation provides a baseline for future model development and comparison.

Conclusions:

• DNALongBench serves as a standardized resource for evaluating deep learning models on long-range DNA dependency tasks.
• The benchmark is expected to accelerate the development and validation of advanced computational genomics tools.