This release substantially improves the accuracy, speed, and memory footprint of Cufflinks and Cuffdiff. It is recommended for all users. Those who wish to see the impact of the accuracy improvements should look at the new benchmarking section. In addition to numerous bugfixes, the main changes are as follows:

- Cuffdiff now includes a new statistical test. Prior versions used a delta method based test, which lacked power for designs with more than a few replicates. The new test directly samples from the beta negative binomial model for each transcript in each condition in order to estimate the null distribution of its log fold change under the null hypothesis. This test is substantially more powerful, resulting in improved accuracy over all experimental designs, particularly those with more than three replicates. A similarly improved test is now used for isoform switching. The benchmarking page shows the improvements in detail.
- Prior versions of Cuffdiff reported the FPKM for each gene and transcript that maximizes the joint likelihood of the reads from all replicates pooled together. In version 2.1, Cuffdiff instead reports the mean of the maximum likelihood estimates from each replicate processed independently. As shown in the benchmarking section, these two methods report nearly identical values. However, the new method is faster and simpler to compute, and will enable new features for future releases.
- The high and low confidence intervals reported by Cufflinks and Cuffdiff are now constructed from the samples generated from the beta negative binomial model, rather than estimated as twice the standard deviation. This better reflects the underlying distribution of the FPKM.
- The library normalization system in Cuffdiff 2 has been overhauled, and several new normalization-related options have been added:
- The new
`–library-norm-method`option now sets which method should be used to compute scaling factors for the library sizes. The default method`geometric`is the same as prior releases of Cuffdiff (and the same as DESeq). The optional modes`quartile`and`classic-fpkm`are also supported. - The new
`–dispersion-method`option controls how the variance model should be computed for each condition. The default mode pooled computes a mean-variance model for each condition that has multiple replicates, averages these models to generate a “pooled” average, and uses it for all conditions. This policy is borrowed from DESeq. Alternative models blind and per-condition are also supported. Prior versions of Cuffdiff used the method`per-condition`. - Several bugs for quartile normalization have been fixed.
- Quartile normalization is no longer supported in Cufflinks, just in Cuffdiff. Cufflinks only supports the
`classic-fpkm`mode. - All library size normalization is now conducted through the internal scaling factor. The external scaling factor should always be set to 1.0.
- Library sizes and dispersions are now computed only on fragment counts from compatible fragments. Prior versions counted intronic and other structurally incompatible fragments in some sections of the code.

- The new
- An optimized sampling procedure drastically improves running time for Cuffdiff. Cufflinks also benefits from this change. The improvements are particularly noticeable on deeply sequenced libraries.
- The range of p-values that users should expect from Cuffdiff has changed. Because the test is now based on explicit sampling from the beta negative binomial, users will not see values less than 10^-5 by default. The test_stat field of Cuffdiff’s output still contains the delta method’s test statistic, but this test statistic is not used to compute p-values. It is preserved for backward compatibility with some functions in CummeRbund. Some extraneous temporary output files have been removed after Cuffmerge runs.