Theseus is a relatively brute force approach to tackling the PCR chimera problem in scRNA-seq by filtering molecules that, for a given cell barcode and UMI pair, have low read abundance. We call this TPT normalization and is defined as follows. For a given pair of cell barcode and UMI, u, and a particular transcript i:

TPT_i,u = (reads_i,u)/sum(reads_j,u)

Where the denominator is summed over all transcripts j sharing the same u.

The user can define a threshold to filter molecules with low TPT and obtain a new expression matrix.

Yorimasa, is a slighlty more sophisticated option, in which the user specifies labels for cells and a set of genes that should be relatively unique for that population. A random forest classifier is trained on the dataset using a diverse set of features. A filter is applied using the out-of-bag probability estimates to remove chimeric molecules.

The features used may include:

• Reads per molecule
• TPT
• Relative abundance of the cell barcode
• Relative abundance of the transcript
• Transcript GC content
• Transcript length

This folder contains iPython notebooks that process Drop-seq output (examples are included in the data subfolder along with precomputed gene level features) with Yorimasa and Theseus and a python script that processes 10X output run using cellranger 1.2+ with Theseus. The names Bellerophon and Perseus were taken already in the context of PCR chimeras.

Usage for 10x Theseus script:

The h5 file is typically the molecule_info file

python theseus_10x.py  FULLPATH2H5.h5 OUTPUTPATH TPT_THRESHOLD NUMBER_OF_CELLS

Python Dependencies: pandas, numpy, matplotlib, seaborn, scipy, sklearn, statsmodels