Supervised promoter recognition: a benchmark framework.

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Deep learning is widely used for identifying genomic regulatory sequences like promoters.
• Previous studies reported improved performance of deep learning models for promoter recognition.
• Lack of standardized datasets and benchmarking procedures hinders accurate performance comparison of these models.

Purpose of the Study:

• To introduce a framework (SUPR REF) for systematic training, validation, testing, and comparison of promoter recognition models.
• To develop biologically relevant benchmark datasets for evaluating deep learning models.
• To reassess the performance of existing deep learning models using standardized benchmarks.

Main Methods:

• Development of the Supervised Promoter Recognition Framework (SUPR REF).
• Creation of standardized, biologically relevant benchmark datasets for promoter recognition.
• Systematic evaluation and comparison of deep learning models on these new datasets.

Main Results:

• Deep learning models demonstrate low reliability for ab initio promoter recognition on eukaryotic genomic sequences.
• Overall performance of current models remains insufficient, particularly for RNA Polymerase II core promoters.
• Cross-validation results from small datasets require cautious interpretation due to their observational nature.

Conclusions:

• The reliability of deep learning for promoter recognition needs significant improvement.
• Standardized benchmarking is crucial for accurate model assessment in genomics.
• Further research is needed to enhance the performance and reliability of computational promoter prediction tools.