Using artificial intelligence for error correction in single cell analyses

The increased sensitivity of the technique, however, also means increased susceptibility to the batch effect. "The batch effect describes fluctuations between measurements that can occur, for example, if the temperature of the device deviates even slightly or the processing time of the cells changes," Maren Büttner explains. Although several models exist for the correction of these deviations, those methods are highly dependent on the actual magnitude of the effect. "We therefore developed a user-friendly, robust and sensitive measure called kBET that quantifies differences between experiments and therefore facilitates the comparison of different correction results," Büttner says.

Besides the batch effect, a phenomenon known as dropout events poses a major challenge in single-cell sequencing. "Let's say we sequence a cell and observe that a particular gene in the cell does not emit any signal at all," explains Dr. Dr. Fabian Theis, ICB Director and professor of Mathematical Modeling of Biological Systems at the TUM. "The underlying cause of this can be biological or technical in nature: either the gene is not being read by the sequencer because it is simply not expressed, or it was not detected for technical reasons," he explains.

To recognize these cases, bioinformaticians Gökcen Eraslan and Lukas Simon from Theis's group used a large number of sequences of many single cells and developed what is known as a deep learning algorithm, i.e. artificial intelligence which simulates learning processes that occur in humans (neural networks).*

Drawing on a new probabilistic model and comparing the original and reconstructed data, the algorithm determines whether the absence of a gene signal is due to a biological or technical failure. "This model even allows cell type-specific corrections to be determined without two different cell types becoming artificially similar," Fabian Theis says. "As one of the first deep learning methods in the field of single-cell genomics, the algorithm has the added benefit that it scales up well to handle data sets containing millions of cells."

But there is one thing the method is not? and this is important to emphasize: "We're not developing software to smooth out results. Our chief goal is to identify and correct errors," Fabian Theis explains. "We're able to share these data, which are as accurate as possible, with our colleagues worldwide and compare our results with theirs," -- for example when the Helmholtz researchers contribute their algorithms and analyses to the Human Cell Atlas, because reliability and comparability of the data are of paramount importance.