Time-Series Gene-Expression Forecasting

From 36 hours of a gene's time course, how well can the next 12 hours be forecast - and are deep sequence models better than a simple 24-hour cosine? This project forecasts 72,000 circadian gene time courses of 12 mouse tissues (GEO GSE54650) with persistence, seasonal-naive and cosinor baselines, ridge autoregression, an LSTM and a temporal CNN, validated on held-out tissues and on another lab's liver (GSE11923). Only rhythmic genes can be forecast (skill 0.42 for the cosinor, 0.41 for the LSTM and TCN, 0.46 for their ensemble), persistence is worse than the mean, scaling with the future fakes skill, and another lab's liver halves it.

Category: Bioinformatics - Transcriptomics and Gene-Expression Analysis

Transcriptomics & Gene-Expression Analysis Project B20 from ProTechSC Academy, with graphical abstract, video demonstration and project guidance.