This reduction in background noise leads to clearer separation between distinct tissue areas. P.S. If you like what you read, I’m currently looking for a new position as a Bioinformatician / Data Scientist in academia or industry — feel free to reach out!

As an example, we applied GTestimate to the 10x Genomics sagittal mouse-brain dataset and observed cleaner normalized expression patterns: the choroid-plexus marker Ttr retained its true signal while unspecific background expression dropped by up to 50 % compared to NormalizeData

The choice of normalization method also has a major impact on spot-wise normalization in Spatial Transcriptomics.

🧠GTestimate for Spatial Transcriptomics In this final spotlight from our GTestimate paper https://t.co/p7YoUL6MAH I want to focus on an alternative application of GTestimate: Spatial Transcriptomics.

This quantified uncertainty could be leveraged for imputation, confidence weighting, or data-quality metrics in future work.

We can estimate for each cell how much of its transcriptome remained unobserved — the missing mass. For a high quality dataset, we found an average missing mass of 30 %, reaching up to 70 % for some cells (Supplementary Fig. S11).

🧩GTestimate: The “Missing Mass” I’ve been highlighting different aspects of our paper https://t.co/p7YoUL6MAH over the last few days. Today I want to point out a unique feature of the Good–Turing estimator: it tells you not only what you see but also how much you miss.

Result: GTestimate resulted in the highest clustering accuracy overall, outperforming NormalizeData at 14 of 15 tested resolutions and SCTransform at 10 of 15. This demonstrates the strong downstream impact of our new normalization method. Check it out: https://t.co/p7YoUL7kqf

We used the Fu et al. (2024) PBMC dataset (Liu dataset) https://t.co/uqm1Xw9SK2 This dataset contains 9266 experimentally annotated cells, we used to compare the unsupervised clustering performance of different Seurat pipelines using NormalizeData, SCTransform, and GTestimate.

📊GTestimate: Clustering Performance During the review process for our paper, Reviewer #1 asked us to benchmark our method on datasets with known cell-type labels.

If this approach sounds interesting, I’d be happy to discuss potential applications or collaborations.

Beyond benchmarking, cta-seq could also be used to characterize rare cell types or specific cell populations, by first identifying these cells in a typical scRNA-seq run and then resequencing the sample while focusing your efforts on the cells of interest for extra depth.

Using these data, we showed that the Good–Turing estimator reduced relative gene-expression estimation error by 17 % on average compared to the traditional Maximum Likelihood approach (see Figure 1 c-d).

Together with Christopher Esk we developed cta-seq, a cell-targeted PCR amplification approach inspired by TAP-seq. This allowed us to sequence 18 selected cells twice at two vastly different depths, with the ultra-deep versions serving as ground-truth gene expression profiles.

While benchmarking our normalization method GTestimate, we quickly realized that simulations weren’t enough — we needed ground-truth gene expression profiles of the same single cells sequenced twice.

🧪 cta-seq: Ultra-deep scRNA-seq for a small set of cells Our new paper was published yesterday in GigaScience! https://t.co/p7YoUL7kqf Today I want to highlight our new cta-seq method for ultra-deep scRNA-seq of a small set of cells.

P.S. If you like what you read, I’m currently looking for a new position as a Bioinformatician / Data Scientist in academia or industry — feel free to message me.

GTestimate is an easy-to-use R package that can serve as a drop-in replacement for Seurat’s NormalizeData() and integrates smoothly into common workflows. Check back tomorrow or follow me for more details over the coming days.

In our paper, we show that replacing ML with the Good–Turing estimator, which accounts for unobserved genes, markedly improves relative gene-expression and cell–cell distance estimation.

scRNA-seq data are notoriously shallow — only ~5 000 molecules per cell are sequenced — yet many normalization methods still rely on the Maximum Likelihood (ML) estimator to calculate relative gene expression per cell, which does not perform well at such shallow sampling depths.