The toolkit, called mSigSDK, was developed by Aaron Ge and colleagues at the National Cancer Institute’s Division of Cancer Epidemiology and Genetics, working with collaborators at the University of Maryland School of Medicine and Brazil’s National Laboratory of Scientific Computing. Rather than building yet another signature-fitting algorithm, the team took a different approach: mSigSDK is an orchestration layer, a JavaScript-based software development kit that loads into a tested desktop browser through a single dynamic import, with no installation whatsoever. Once loaded, it can run analyses entirely on the user’s own device, meaning that sensitive genomic data — the mutation spectra and mutation annotation format rows that describe a patient’s tumor — never have to leave the computer they sit on.

At the heart of the problem mSigSDK addresses is a quiet but consequential inconsistency in the field. Several mature packages for mutational signature analysis exist, most prominently written in R and Python, and each has its own conventions for input data, its own fitting algorithms, and its own quirks. The choice of fitting tool can meaningfully affect the results a researcher obtains, which makes triangulating across multiple tools highly desirable. Yet doing so has been operationally expensive: a scientist would need to install and maintain several environments, convert data between formats, and manually reconcile outputs. Worse, no standard machine-readable format existed to capture the parameters of an analysis, the evidence supporting a review, and the provenance of the results — the kind of documentation that makes a computational finding reproducible and trustworthy.

mSigSDK tackles this by wrapping four established analysis tools — SigProfilerAssignment, MuSiCal, deconstructSigs, and sigminer — behind a uniform adapter layer. Through that layer, the kit orchestrates the tools so they can be run side by side on the same data without the user touching their individual installation requirements or input formats. The SDK also adds its own browser-native non-negative least squares fitting, a standard mathematical approach for estimating how much each known reference signature contributed to a tumor’s observed mutation pattern, as well as non-negative matrix factorization-based exploratory extraction, which allows researchers to discover novel signatures rather than only testing against a fixed catalogue such as the COSMIC reference set.

The engineering achievement here is substantial. Running R and Python code inside a browser is possible thanks to technologies like WebR and Pyodide, which compile those language runtimes to WebAssembly, but mSigSDK goes further by making the whole experience seamless. A zero-install demonstration reached a fully rendered local report in roughly 1.82 seconds. The kit emits configurable, rule-based review evidence — essentially automated documentation of why an analysis reached its conclusions — and serializes portable, schema-validated reports with built-in visualization and export. That means an analysis performed on one machine can be packaged as a machine-readable file, checked against a formal schema, and shared or archived with its full provenance intact.

Privacy is a central design consideration. Because analysis runs locally, mutation spectra and mutation annotation format rows stay on the user’s device by default. The authors are transparent about one caveat: optional reference-context helpers may transmit genomic coordinates to public endpoints unless a strict-local mode is enabled. This kind of explicit, switchable data-flow control is increasingly important as genomics moves into clinical settings where patient data cannot simply be uploaded to a remote server. A browser-native tool that keeps raw data on-device while still offering state-of-the-art analysis could lower the barrier for hospitals and research groups working under strict data governance rules.

The validation work behind the paper is unusually thorough. The team tested each of the four adapters against the corresponding package running locally on 38 single-base substitution spectra from lung adenocarcinoma samples in the Pan-Cancer Analysis of Whole Genomes project. Three of the four adapters reproduced their package’s local execution exactly, and the fourth matched to floating-point precision. The kit’s data converter matched SigProfilerMatrixGenerator version 1.3.6 exactly for single-base substitution spectra in both 96-channel and 1,536-channel formats, for double-base substitution spectra in the 78-channel format, and for the 83-channel small insertion-deletion classification, provided that repeat and microhomology annotations derived from SigProfilerMatrixGenerator were supplied to mSigSDK. In other words, running these tools in the browser yields the same numbers as running them on a desktop installation.

But the benchmarks also revealed something the field should pay attention to. Across 2,700 synthetic spectra generated at three noise levels, the mean cosine similarity of estimated exposures exceeded 0.96 for all tools — meaning the overall mixture of signatures was recovered well. Yet the mean per-spectrum F1 score for identifying which signatures were genuinely active ranged from 0.532 to 0.924 at 10 percent noise, a striking spread. On real PCAWG data, all tools reconstructed the observed spectra well, with mean cosine similarities between 0.982 and 0.994, but they disagreed sharply on how many signatures were active per sample, calling anywhere from 4.13 to 17.39 active signatures. This confirms that the choice of fitting tool matters, and it makes the case for multi-tool triangulation all the more compelling.

Speed is another highlight. Warm-start refitting — re-estimating signature exposures when a small change is made to the input — took between 18.2 and 73.7 milliseconds on Windows and 22.2 to 44.6 milliseconds on macOS for 120 samples. For the heavier scenario of 300 samples against 40 reference signatures, refitting completed in 1.80 to 6.81 seconds on Windows and 2.12 to 5.89 seconds on macOS. Those are interactive timescales, fast enough for a researcher to adjust parameters and see updated results in real time, something that batch-oriented command-line workflows have never offered.

The broader significance of mSigSDK may lie less in any single benchmark than in what it signals about the future of scientific software. By combining multi-tool orchestration, exploratory signature extraction, rule-based review evidence, and schema-validated reporting in a zero-install browser environment, the kit lowers the cost of doing rigorous, reproducible, and privacy-preserving mutational signature analysis to nearly zero. Developed with support from the National Cancer Institute’s Intramural Research Program and released as open access, it invites a much wider community — from large consortia to individual labs in resource-limited settings — to interrogate the mutational histories written into cancer genomes. If the fingerprints in our DNA are a diary, mSigSDK hands more readers than ever the pen-light needed to read it.

Subject of Research: Browser-native computational analysis of cancer mutational signatures

Article Title: mSigSDK: browser-native computation of mutational signatures

Article References: Ge, A., Zhang, T., Martins, Y. C., Landi, M. T., Park, B., Chen, K., Balasubramanian, J., & Almeida, J. S. (2026). mSigSDK: browser-native computation of mutational signatures. BMC Bioinformatics. https://doi.org/10.1186/s12859-026-06681-z

Image Credits: AI Generated

DOI: 10.1186/s12859-026-06681-z

Keywords: mutational signatures, cancer genomics, JavaScript, browser-native computation, bioinformatics software, SigProfilerAssignment, MuSiCal, deconstructSigs, sigminer, non-negative least squares, PCAWG, reproducibility