Summary
A University of Michigan team has introduced AmbientMapper, a computational method that estimates ambient chromatin contamination in combinatorial single-cell experiments. In a bioRxiv preprint, the authors report a 0.19% wrong-genome rate on a 26-genome reference panel and recovery of low-depth nuclei that standard pipelines discarded.
Researchers at the University of Michigan have introduced a computational workflow for detecting and removing ambient chromatin contamination from combinatorial single-cell sequencing data. The method, described in a bioRxiv preprint posted on September 16, 2026, is designed to distinguish genuine nucleus-specific signal from fragments that enter a library from other material during the experimental workflow.
The study introduces two tools: scifi-demux, for preprocessing scifi-ATAC libraries, and AmbientMapper, a generative model that analyses sequencing reads against multiple references. The authors report that the approach achieved a wrong-genome rate of 0.19% on a panel containing 26 genomes and recovered low-depth nuclei that standard analysis pipelines had discarded.
Combinatorial indexing and droplet-based microfluidics have made single-cell experiments much larger by assigning cells or nuclei combinations of molecular barcodes. In scifi-ATAC, the resulting data describe regions of chromatin that are accessible in individual nuclei. The scale improves the ability to study diverse cell populations, but it also creates more opportunities for background material to be mixed into the signal assigned to each barcode.
How AmbientMapper works
Ambient chromatin can distort chromatin profiles, particularly when a nucleus has relatively few sequencing reads or belongs to a minority population. Standard cell-calling procedures rely heavily on read-count thresholds, which can make low-depth nuclei difficult to distinguish from background. Existing decontamination approaches generally work on aggregated count matrices, after individual reads have already been summarised.
AmbientMapper instead works at the read level. It maps reads competitively against multiple reference genomes or other references, allowing the model to assess which source best explains the observed sequence. It learns the ambient profile from empty and low-complexity barcodes, which provide examples of background signal within the library.
The model then separates nuclei from background and distinguishes singlets—barcodes representing one nucleus—from doublets, in which material from two nuclei is combined. These decisions use the Bayesian Information Criterion, a statistical model-selection measure that balances how well a model fits the data against its complexity.
The companion scifi-demux tool is intended to prepare scifi-ATAC libraries for this analysis. Together, the tools address both preprocessing and the assignment of reads to likely sources.
Tests of accuracy and data recovery
The authors evaluated the approach using three types of evidence: interspecies ground-truth experiments, published multi-genotype libraries and simulated read-level synthetic barcodes in which every contaminating read could be traced to its origin. Interspecies experiments provide a direct way to test assignment because reads from different genomes have known sources.
The authors report that AmbientMapper's calls remained robust when parameters were varied and were stable across the tested experimental designs. On the 26-genome reference panel, the reported wrong-genome rate was 0.19%.
The study also examined the effect of removing contaminating reads. According to the authors, AmbientMapper recovered low-depth nuclei that standard pipelines had rejected and restored biological structure that had been obscured by contamination. This is important for experiments in which rare or low-signal populations may be biologically meaningful but are especially vulnerable to read-count filtering.
The work is a computational methods study of single-cell sequencing data, rather than a clinical study or a new treatment. It is available as a bioRxiv preprint and therefore presents its findings before peer review. The reported 0.19% figure applies to the 26-genome reference-panel evaluation described by the authors; it is not a general error rate for every single-cell workflow. The preprint's abstract also does not quantify the number of nuclei rescued or the size of the improvement in biological-structure recovery.