NEW YORK – BostonGene last week published details of a machine learning algorithm that distinguishes tumor microenvironment (TME)-associated transcriptomic elements from bulk RNA-seq data.

The algorithm, called Kassandra, is provided with the company's Tumor Portrait Test and was developed to predict the percentage of blood and tissue cells comprising the TME. The decision tree machine learning deconvolution algorithm was published in the journal Cancer Cell, and could find use in applications such as tracking immunotherapy response and to increase the predictive potential of established biomarkers.

Although the TME comprises a small fraction of tumor and bulk RNA-seq reads, its cellular subsets, such as natural killer cells, can significantly impact therapeutic response and clinical outcomes.

"With high accuracy, Kassandra can identify up to 51 unique cell types, including minor subpopulations, and has been optimized for FFPE samples, enabling utilization in clinical settings," said Nathan Fowler, BostonGene's chief medical officer.

Waltham, Massachusetts-based BostonGene trained Kassandra on millions of artificial transcriptomes made to recapitulate tissue-like and blood-like bulk RNA profiles and imitate tissue heterogeneity, derived from 9,404 tissue and blood RNA-seq profiles from various immune and stromal cell populations, including malignant cells from 24 cancer types.

After training, BostonGene researchers evaluated Kassandra's performance on RNA-seq datasets from tumor samples found on the TCGA database and on donor-derived peripheral blood mononucleocyte and peripheral blood lymphocyte samples, as well as on several single-cell RNA-seq datasets.

Importantly, Kassandra-reconstructed TMEs correlated better to their corresponding tumor types than other deconvolution algorithms such as CIBERSORT, ABIS, and quanTIseq, and correlated well with PD-L1 status and immunotherapy response in several cancer validation cohorts.

Kassandra identified 18 TME subpopulations from tissue-based samples and 41 subpopulations from blood, including several low-abundance cell types.

Gordon Mills, director of precision oncology for the OHSU Knight Cancer Institute, complimented BostonGene's study, saying via email that the large number of samples used for training and the comparisons to other algorithms represented a "major positive" for Kassandra.

"The ability to accurately define cell types from deconvolution of RNA-seq data would represent a major advance in the field," he said. "The Kassandra decision tree algorithm was designed and trained on a very large set of RNA-seq data including data from purified populations that was not available in the development of prior algorithms."

He also praised the authors for having "carefully considered and attempted to mitigate challenges with previous algorithms such as bias, noise, and batch effects."

Mills cautioned, however, that Kassandra was not without its limitations, noting that it would not be an effective tool for understanding signaling processes that drive cell type frequency, cell type localization, or certain differences between cell populations, such as markers used to call particular cell lineages.

"For example," he said, "it cannot distinguish between tumor-invading, peritumor, or distant immune cells. As a major challenge, it does not provide any information on spatial organization, cell-cell interactions, and cell communities, all of which are important for understanding the tumor ecosystem."

Fowler agreed that Kassandra does not assess spatial heterogeneity or temporal dynamics from a single biopsy but said that it can provide tissue TME composition from multi-regional biopsies and longitudinal biopsies, suggesting some insights into spatial heterogeneity and architecture as well as dynamics, respectively.

"Further," he added, "we have a separate multiplex immunofluorescence platform to derive spatial architecture and spatial heterogeneity."

BostonGene presented this platform earlier this year at the annual American Society of Clinical Oncology conference.

The company has entered into multiple collaborations aimed at using its technology to derive clinically actionable information from patients' TME.

Last month, for example, it entered an agreement with Memorial Sloan Kettering to provide molecular laboratory and computational analysis services for several cancer precision medicine research initiatives. Last year, the company entered into similar agreements with Tokyo-based NEC, the MD Anderson Cancer Center, and others.

BostonGene makes Kassandra available to all physicians upon ordering a Tumor Portrait Test.

"Providing cell deconvolution to physicians underscores BostonGene's goal of offering a comprehensive and holistic portrait of each unique patient tumor to the physician for informed decision-making," Fowler said.

Privacy Policy.  Terms & Conditions. Copyright © 2022 GenomeWeb, a business unit of Crain Communications.  All Rights Reserved.