Differentially expressed heterogeneous overdispersion genes testing for count data

Abstract

The mRNA-seq data analysis is a powerful technology for inferring information from biological systems of interest. Specifically, the sequenced RNA fragments are aligned with genomic reference sequences, and we count the number of sequence fragments corresponding to each gene for each condition. A gene is identified as differentially expressed (DE) if the difference in its count numbers between conditions is statistically significant. Several statistical analysis methods have been developed to detect DE genes based on RNA-seq data. However, the existing methods could suffer decreasing power to identify DE genes arising from overdispersion and limited sample size, where overdispersion refers to the empirical phenomenon that the variance of read counts is larger than the mean of read counts. We propose a new differential expression analysis procedure: heterogeneous overdispersion genes testing (DEHOGT) based on heterogeneous overdispersion modeling and a post-hoc inference procedure. DEHOGT integrates sample information from all conditions and provides a more flexible and adaptive overdispersion modeling for the RNA-seq read count. DEHOGT adopts a gene-wise estimation scheme to enhance the detection power of differentially expressed genes when the number of replicates is limited as long as the number of conditions is large. DEHOGT is tested on the synthetic RNA-seq read count data and outperforms two popular existing methods, DESeq2 and EdgeR, in detecting DE genes. We apply the proposed method to a test dataset using RNAseq data from microglial cells. DEHOGT tends to detect more differently expressed genes potentially related to microglial cells under different stress hormones treatments.

Fig 1. The overdispersion in the real RNA-seq count data.

Fig 2. The false negative rate from different methods when the read counts follow the quasi-Poisson distribution with different overdispersion levels θgQP in simulation setting 1.

The bars represents the standard deviation of the false negative rate over repeated experiments.

Fig 3. The false negative rate from different methods when the read counts follow the negative binomial distribution with different overdispersion levels θgNB in simulation setting 1.

The bars represents the standard deviation of the false negative rate over repeated experiments.

Fig 4. The area under the ROC curve from different methods when the read counts follow the quasi-Poisson distribution with different overdispersion levels θgQP in simulation setting 1.

Fig 5. The area under the ROC curve from different methods when the read counts follow the negative binomial distribution with different overdispersion levels θgNB in simulation setting 1.

Fig 6. The false negative rate from different methods when the read count follow the quasi-Poisson distribution with different overdispersion levels θ^QP.

The variance of FNR obtained from repeated experiments is illustrated using the bars.

Fig 7. The false negative rate from different methods when the read count follows a negative binomial distribution with different overdispersion levels θ^NB.

Fig 8. The AUC from different methods when the read count follows the quasi-Poisson distribution with different overdispersion levels θ^QP.

The variance of FNR obtained from repeated experiments is illustrated using the bars.

Fig 9. The AUC from different methods when the read counts follow the negative binomial distribution with different overdispersion levels θ^NB.

Fig 10. The ROC curve from different methods when the read counts follow the negative binomial distribution with θgNB∈(1,2) and quasi-Poisson distribution with θgQP∈(50,100).

Fig 11. The dispersion of genewise RNA read counts.

Each dot corresponds to a sample count from a specific gene.

Fig 12. The selected DE genes from DEHOGT, DESeq2, EdgeR under treatment comparison dexh3 versus control.

Fig 13. The selected DE genes from DEHOGT, DESeq2, EdgeR under treatment comparison dexh6 versus control.

Fig 14. The selected DE genes from DEHOGT, DESeq2, EdgeR under treatment comparison cortvh3 vs corth3.

Fig 15. The selected DE genes from DEHOGT, DESeq2, EdgeR under treatment comparison corth6 versus control.

Fig 16. The selected DE genes from DEHOGT, DESeq2, EdgeR under treatment comparison dexvh3 versus dexh3.

Fig 17. The selected DE genes from DEHOGT, DESeq2, EdgeR under treatment comparison dexvl3 vs dexl3.

Fig 18. The selected DE genes from DEHOGT, DESeq2, EdgeR under treatment comparison corth3 versus control.

Fig 19. The rank of p-value of selected genes under treatment comparison dexh3 versus control, a shorter bar indicates a smaller p-value (more significantly differently expressed).

Fig 20. The rank of p-value of selected genes under treatment comparison dexh6 versus control, and shorter bar indicates a smaller p-value.

Fig 21. The rank of p-value of selected genes under treatment comparison corth3 versus control, and shorter bar indicates a smaller p-value.

Fig 22. The rank of p-value of selected genes under treatment comparison dexvh versus dexh, and shorter bar indicates a smaller p-value.

Fig 23. The rank of p-value of selected genes under treatment comparison dexvl versus dexl, and shorter bar indicates a smaller p-value.

Fig 24. The rank of p-value of selected genes under treatment comparison cortvh versus corth, and shorter bar indicates a smaller p-value.