LRSA: A new computational method for analyzing time course microarray data

Motivation: Time course data obtained from biological samples subject to specific treatments can be very useful for revealing complex and novel biological phenomena. Although an increasing number of time course microarray datasets becomes available, most of them contain few biological replicates and time points. So far there are few computational methods that can effectively reveal differentially expressed genes and their patterns in such data. Results: We have proposed a new two-step nonparametric statistical procedure, LRSA, to reveal differentially expressed genes and their expression trends in temporal microarray data. We have also employed external controls as a surrogate to estimate false discovery rates and thus to guide the discovery of differentially expressed genes. Our results showed that LRSA reveals substantially more differentially expressed genes and have much lower than two other methods, STEM and ANOVA, in both real data and the simulated data. Our computational results are confirmed using real-time PCRs. Contact: wuw2@upmc.edu

💡 Research Summary

The paper introduces LRSA (Local Regression and Spectral Analysis), a two‑step non‑parametric framework designed for time‑course microarray experiments that typically suffer from few biological replicates and a limited number of time points. In the first step, each gene’s expression trajectory across time is fitted with a local quadratic regression (implemented via the R package locfit). The optimal smoothing bandwidth is chosen automatically using a generalized cross‑validation (GCV) criterion, allowing the method to adapt to varying data density without user‑defined parameters. To assess differential expression (DE), the authors compute simultaneous 95 % confidence bands for the fitted curves using the Faraway‑Sun approach, which explicitly accounts for heteroscedasticity—an important feature of temporal gene expression where variance often changes over time. A gene is declared DE at a particular time point if its observed expression lies outside the simultaneous band and the fold‑change exceeds a 2‑fold threshold. Multiple‑testing correction is performed in a Bonferroni‑like fashion, adjusting the confidence level either across time points or across genes, depending on the experimental focus.

A novel aspect of LRSA is the use of external control probes (408 bacterial spots, 68 unique sequences) embedded on each array as a surrogate for true null hypotheses. By counting how many of these controls are mistakenly called DE, the authors define an “External‑Control FDR” (EC‑FDR), providing a practical, data‑driven estimate of false discovery rates when the ground truth is unknown.

In the second step, the set of DE genes identified in step one is clustered using spectral clustering. Rather than clustering raw expression values, LRSA clusters the smoothed fitted values ˆf(t) evaluated at 31 equally spaced pseudo‑time points between 0 and 30 days. This reduces noise and enables the algorithm to capture subtle temporal patterns. Correlation coefficients serve as the affinity matrix, and the eigen‑vectors of the normalized Laplacian are fed into a k‑means step, following the Ng‑Jordan‑Weiss formulation. The authors argue that spectral clustering better respects the connectivity structure of high‑dimensional gene expression data than traditional K‑means.

The method was evaluated on a real hypoxia experiment (rats exposed to low oxygen for 0, 1, 3, 7, 14, 30 days, with three biological replicates per condition, though some replicates were lost) and on simulated data that mimics the same experimental design. Compared with two widely used approaches—STEM (which clusters genes after a simple fold‑change filter) and a classic ANOVA‑based pipeline—the LRSA pipeline identified substantially more DE genes while maintaining an EC‑FDR close to zero. For example, without multiple‑testing adjustment and using only a 2‑fold cutoff, LRSA detected 1,525 DE genes versus 237 for ANOVA; STEM’s EC‑FDR was 0.07, indicating many false positives among its “significant” genes, whereas LRSA’s EC‑FDR was essentially 0. The authors also present four illustrative genes (HO‑2, MAPK1, NFKB1, KCNMA1) whose smooth temporal profiles and confidence bands are clearly visualized. Validation by quantitative RT‑PCR confirmed the microarray‑derived expression trends for selected genes, supporting the biological relevance of LRSA’s findings.

Key strengths of LRSA include (1) flexibility to model arbitrary, non‑linear expression trajectories without imposing parametric forms; (2) explicit handling of heteroscedasticity through simultaneous confidence bands, which improves statistical power; (3) a pragmatic, array‑based estimate of false discovery rates via external controls; and (4) robust clustering of temporal patterns using spectral methods. Limitations are acknowledged: the simultaneous band construction can become unstable when the number of replicates per time point is extremely low (≤2), and the method’s performance may be sensitive to the choice of bandwidth, although the GCV procedure mitigates this. Future work could explore Bayesian bandwidth selection and incorporation of additional external controls to further stabilize FDR estimation. Overall, LRSA offers a statistically rigorous and biologically interpretable solution for extracting dynamic gene expression signatures from sparse time‑course microarray data.