DNALONGBENCH

DNALONGBENCH provides a benchmark dataset and standardized resource for evaluating deep learning models that account for long-range dependencies in DNA sequences, covering tasks with dependencies up to 1 million base pairs.

Key Features:

• Long-range DNA prediction benchmark: Covers five genomics tasks with long-range dependencies up to 1 million base pairs.
• Task coverage: Includes enhancer-target gene interaction, expression quantitative trait loci, 3D genome organization, regulatory sequence activity, and transcription initiation signals.
• Standardized evaluation framework: Enables comprehensive comparisons and rigorous evaluations of DNA sequence-based deep learning models.
• Multi-model assessment: Benchmarked using a task-specific expert model, a convolutional neural network (CNN)-based model, and three fine-tuned DNA foundation models (HyenaDNA, Caduceus-Ph, and Caduceus-PS).

Scientific Applications:

• Enhancer-target gene interaction prediction: Evaluates models on identifying regulatory connections between distal enhancers and their target genes.
• Expression quantitative trait loci analysis: Assesses prediction of genetic variants affecting gene expression across long genomic distances.
• 3D genome organization prediction: Tests modeling of three-dimensional chromatin folding and structure across millions of base pairs.
• Regulatory sequence activity prediction: Measures inference of functional activity from distal regulatory DNA elements.
• Transcription initiation signal prediction: Evaluates detection of signals governing transcription start site selection and activity.

Methodology:

DNALONGBENCH was constructed as a benchmark dataset covering five genomics tasks with long-range dependencies up to 1 million base pairs, and was assessed using five methods: a task-specific expert model, a convolutional neural network (CNN)-based model, and three fine-tuned DNA foundation models (HyenaDNA, Caduceus-Ph, and Caduceus-PS).