https://science.bostongene.com/kassandra/
Table of Contents
Kassandra is a robust and accurate cell deconvolution tool developed for the analysis of healthy tissue and tumor biopsies. Based on RNA-seq NGS data of a biological sample, Kassandra predicts cellular composition, including stromal and immune elements, by analyzing gene expression. This will lead to an improved understanding of the tumor microenvironment, which is a critical factor in cancer pathogenesis, clinical outcome, and therapeutic resistance.
Kassandra is a decision tree machine learning algorithm trained on a collection of over thousands of RNA profiles from various sorted cell types. Performance was validated on over 4,000 H&E tissue slides and more than 1,000 samples comprising normal and tumor tissues by comparison with cytometric, immunohistochemical or single-cell RNA sequencing measurements of the same tissue.
"Model Training.ipynb” provides limited examples of the training blood deconvolution model. The model is relatively small and can be trained on an average laptop. To reproduce the full publication model use "configs/full_blood_model.yaml" config. It requires 40-60 gb of memory for generation of training data.
The example dataset consists of 4 tables in the 'data' folder, two with annotation and two with the expression of cancer cell lines and sorted cell samples. To train the full model, download data from https://science.bostongene.com/kassandra/downloads.
Kassandra deconvolution is trained and can be used on gene TPM expression obtained by the specific procedure described in the methods section of the manuscript. The number of transcripts has to be removed from expression before it is aggregated to gene expression and renormalized to TPM. Tumor and blood models described in the manuscript use slightly different procedures. Both procedures are demonstrated in "Model Training.ipynb" notebook.
Inputs in the form of TPM produced by other tools or procedures different from described above are not recommended.
Precise reconstruction of the tumor microenvironment using bulk RNA-seq and a unique machine learning-based algorithm trained on artificial transcriptomes
Aleksandr Zaitsev1, Maksim Chelushkin1, Daniiar Dyikanov1, Ilya Cheremushkin1, Boris Shpak1, Krystle Nomie1, Vladimir Zyrin1, Katerina Nuzhdina1, Yaroslav Lozinsky1, Anastasia Zotova1 , Sandrine Degryse1, Nikita Kotlov1, Arthur Baisangurov1, Vladimir Shatsky1, Daria Afenteva1, Susan Raju Paul2, Diane L. Davies3, Patrick M. Reeves2, Michael Lanuti3, Michael F. Goldberg1, Cagdas Tazearslan1, Madison Chasse1, Iris Wang1, Mary Abdou1, Sharon M. Aslanian1, Samuel Andrewes1, James J. Hsieh4, Akshaya Ramachandran4, Yang Lyu4, Ilia Galkin1, Viktor Svekolkin1, Leandro Cerchietti5, Mark C. Poznansky2, Ravshan Ataullakhanov1, Nathan Fowler1,6*, Alexander Bagaev1*
4Molecular Oncology, Division of Oncology, Department of Medicine, Washington University, St. Louis, MO, USA
BY UTILIZING THE CODE, YOU ARE CONSENTING TO BE AND AGREE TO BE BOUND BY ALL OF THE TERMS OF THIS LIMITED LICENSE, SEE "LICENSE.txt"