gprege

gprege models continuous gene expression trajectories in time-series data using Gaussian Process (GP) regression with Radial Basis Function (RBF) kernels to analyze temporal dynamics, incorporate noise modeling, and support filtering and differential expression assessment.

Key Features:

• Gaussian Process Regression: Uses GP regression to estimate continuous trajectories from gene expression time series while accommodating biological replicates and missing values.
• Kernel Configuration — Two Models: Fits two distinct RBF-kernel GP models: one initialized with a short lengthscale, signal variance set to the observed variance, and zero noise variance that is optimized; and a second with fixed hyperparameters (zero inverse-width, zero signal variance, and noise variance equal to the observed variance).
• Optimization: Optimizes the first GP model using scaled conjugate gradients implemented via netlab.
• Noise Handling: Incorporates a noise diagonal into the RBF kernel to account for data variability.
• Uncertainty Quantification: Provides confidence intervals along estimated expression curves.
• Filtering and Differential Expression: Implements methods to filter inactive (quiet) genes and to quantify differential expression in time series expressed as ratios.
• Performance Assessment: Evaluates rankings with Receiver Operating Characteristic (ROC) curve analysis and comparative assessments against the hierarchical Bayesian model BATS, reporting superior performance on simulated and experimental data.

Scientific Applications:

• Temporal trajectory estimation: Reconstruction of continuous gene expression trajectories from time-series data including microarray experiments.
• Handling replicates and missing data: Analysis of time-series datasets containing biological replicates and missing observations.
• Differential expression over time: Identification and quantification of temporally varying gene expression using ratio-based measures.
• Benchmarking and validation: Comparative evaluation of time-series inference methods using ROC analysis on simulated and experimental datasets.

Methodology:

Fits two RBF-kernel Gaussian Process models (one initialized with short lengthscale, signal variance set to observed variance and zero noise then optimized with scaled conjugate gradients via netlab; the other with fixed hyperparameters: zero inverse-width, zero signal variance, noise variance equal to observed variance), incorporates a noise diagonal into the kernel, and assesses performance using ROC curves and comparisons to BATS.