Transforming L1000 profiles to RNA-seq-like profiles with deep learning.

Area of Science:

• Computational biology
• Genomics
• Machine learning

Background:

• The L1000 technology provides high-throughput gene expression profiles for over 30,000 perturbations.
• It measures 978 landmark genes, inferring an additional 11,350, but lacks full genome coverage.
• This limitation hinders comprehensive knowledge discovery and integration with other transcriptomic data.

Observation:

• The L1000 dataset, with over 3 million profiles, is valuable for drug discovery and understanding small molecule mechanisms.
• Current L1000 data only measures 978 landmark genes, with the rest computationally inferred.
• This incomplete coverage restricts the analysis of half of human protein-coding genes.

Findings:

• A novel two-step deep learning model converts L1000 profiles into RNA-seq-like profiles covering 23,614 genes.
• The model utilizes a CycleGAN for initial transformation and a neural network for full genome extrapolation.
• High accuracy was achieved with 0.914 Pearson's correlation and 1.167 RMSE on paired L1000/RNA-seq data.

Implications:

• This method significantly expands the utility of existing L1000 data for biological research.
• Enables deeper insights into drug mechanisms and identification of novel drug and target candidates.
• Provides researchers with downloadable RNA-seq-like profiles for signature and reverse gene searches.