CRMnet: A deep learning model for predicting gene expression from large regulatory sequence datasets.

Area of Science:

• Computational Biology
• Genomics
• Machine Learning

Background:

• Large-scale gene expression datasets offer opportunities for deep learning applications.
• Understanding gene regulation requires accurate prediction of expression from DNA sequences.
• Deep neural networks can model complex dependencies in regulatory sequences.

Purpose of the Study:

• To design and train a novel deep-learning model (CRMnet) for predicting gene expression in *Saccharomyces cerevisiae*.
• To interpret the model to identify key regulatory regions and transcription factor binding sites.
• To evaluate the model's performance against existing benchmarks and assess its training efficiency.

Main Methods:

• Development of a novel deep-learning architecture (CRMnet) for gene expression prediction.
• Training the model on large datasets of gene promoter sequences.
• Utilizing model saliency maps to interpret informative genomic regions.
• Comparing model performance (Pearson correlation, MSE) and training times (GPU, TPU).

Main Results:

• CRMnet achieved superior predictive performance, with a Pearson correlation coefficient of 0.971 and a mean squared error of 3.200.
• Model interpretation successfully identified informative genomic regions, correlating with known yeast motifs.
• The model demonstrated the ability to locate transcription factor binding sites modulating gene expression.
• Training times were evaluated on high-performance computing resources (GPUs, TPUs).

Conclusions:

• CRMnet represents a significant advancement in predicting gene expression from DNA sequences.
• Model interpretability facilitates biological discovery by pinpointing regulatory elements.
• The developed model aids in understanding the complex regulatory code governing gene expression.
• The study highlights the practical feasibility of training such models on large datasets.