Tom Skelly

Abstract Motivation: Primary data analysis methods are of critical importance in second generation DNA sequencing. Improved methods have the potential to increase yield and reduce the error rates. Openly documented analysis tools enable the user to understand the primary data, this is important for the optimization and validity of their scientific work. Results: In this article, we describe Swift, a new tool for performing primary data analysis on the Illumina Solexa Sequencing Platform. Swift is the first tool, outside of the vendors own software, which completes the full analysis process, from raw images through to base calls. As such it provides an alternative to, and independent validation of, the vendor supplied tool. Our results show that Swift is able to increase yield by 13.8%, at comparable error rate. Availability and Implementation: Swift is implemented in C++and supported under Linux. It is supplied under an open source license (LGPL3), allowing researchers to build upon the platform. Swift is available from http://swiftng.sourceforge.net. Contact: new@sgenomics.org; nava.whiteford@nanoporetech.com Supplementary information: Supplementary data are available at Bioinformatics online.

The new generation of short-read sequencing technologies requires reliable measures of data quality. Such measures are especially important for variant calling. However, in the particular case of SNP calling, a great number of false-positive SNPs may be obtained. One needs to distinguish putative SNPs from sequencing or other errors. We found that not only the probability of sequencing errors (i.e. the quality value) is important to distinguish an FP-SNP but also the conditional probability of "correcting" this error (the "second best call" probability, conditional on that of the first call). Surprisingly, around 80% of mismatches can be "corrected" with this second call. Another way to reduce the rate of FP-SNPs is to retrieve DNA motifs that seem to be prone to sequencing errors, and to attach a corresponding conditional quality value to these motifs. We have developed several measures to distinguish between sequence errors and candidate SNPs, based on a base call's nucleotide context and its mismatch type. In addition, we suggested a simple method to correct the majority of mismatches, based on conditional probability of their "second" best intensity call. We attach a corresponding second call confidence (quality value) of being corrected to each mismatch.

As was the case in the beginning of the sequencing era, the new generation of short-read sequencing technologies still requires both accuracy of data processing methods and reliable measures of that accuracy. Inspired by the classic of the genre, the Phred method, we generalized those findings in the area of base quality value calibration. We introduce a simple, straightforward statistically established way to measure the performance of a calibrator, and to find an optimal way to assess its reliability. We illustrate the method by assessing the performance of several calibrators/predictors for Illumina, Genome Analyser 2 (GA2) data. The choice of the best predictor is based on optimization of validity, discriminative ability and discrimination power for several candidate predictors. We applied the method on data from one experimental run for genome of the phage varphiX, and found the best predictor out of ten candidates to be 'Purity', a statistics derived from corrected cluster intensities. The source code for the comparison of the predictors is available from the authors by request.

Abstract Next generation DNA sequencing methods have created an unprecedented leap in sequence data generation, thus novel computational tools and statistical models are required to optimize and assess the resulting data. In this report, we explore underlying causes of error for the Illumina Genome Analyzer (IGA) sequencing technology and attempt to quantify their effects using a human bacterial artificial chromosome sequenced to 60,000 fold coverage. Seven potential error predictors are considered: Phred score, read entropy, tile coordinates, local tile density, base position within read, nucleotide call, and lane. With these parameters, logistic regression and log-linear models are constructed and used to show that each of the potential predictors contributes to error (P<1×10 −4 ). With this additional information, we apply the logistic model and achieve a 3% improvement in both the sensitivity and specificity to detect IGA errors. Further, we demonstrate that these modeling approaches can be used as a feedback loop to inform laboratory methods and identify specific machine or run bias.