MFmap: A semi-supervised generative model matching cell lines to tumours and cancer subtypes

Abstract

Translating in vitro results from experiments with cancer cell lines to clinical applications requires the selection of appropriate cell line models. Here we present MFmap (model fidelity map), a machine learning model to simultaneously predict the cancer subtype of a cell line and its similarity to an individual tumour sample. The MFmap is a semi-supervised generative model, which compresses high dimensional gene expression, copy number variation and mutation data into cancer subtype informed low dimensional latent representations. The accuracy (test set F1 score >90%) of the MFmap subtype prediction is validated in ten different cancer datasets. We use breast cancer and glioblastoma cohorts as examples to show how subtype specific drug sensitivity can be translated to individual tumour samples. The low dimensional latent representations extracted by MFmap explain known and novel subtype specific features and enable the analysis of cell-state transformations between different subtypes. From a methodological perspective, we report that MFmap is a semi-supervised method which simultaneously achieves good generative and predictive performance and thus opens opportunities in other areas of computational biology.

Computational estimation of quality and clinical relevance of cancer cell lines

“…contrastive PCA followed by MNN or quantile normalisation plus ComBat), with only the second case showing the two datasets properly mixed, while maintaining tissue type separations. A different approach is applied in (Ronen et al , 2019 ; Zhang & Kschischo, 2021 ), based on a variational autoencoder (VAE) that identifies, in an unsupervised manner, non‐linear latent factors from the initial feature space common to both CCLs and tumours.…”

“…The scoring objective is instead one of the most pursued across the examined methods (Fig 4B and Table 2 ). This goal is usually achieved through the use of a correlation (Spearman's or Pearson's) or similar metric (Kendall or Jaccard index, Euclidean distance or cosine coefficient) (Domcke et al , 2013 ; Chen et al , 2015 ; Sun & Liu, 2015 ; Vincent et al , 2015 ; Jiang et al , 2016 ; Luebker et al , 2017 ; Sinha et al , 2017 , 2021 ; Vincent & Postovit, 2017 ; Liu et al , 2019a ; Ronen et al , 2019 ; Batchu et al , 2020 ; Fang et al , 2021 ; Zhang & Kschischo, 2021 ), sometimes applied to a new “corrected” feature space (Warren et al , 2021 ). This similarity score is usually computed first sample‐wise, then for each CCL averaged across tumours from a given tumour type/subtype (usually matching that in the CCL annotation).…”

“…Generally, pursuing the selection objective helps also identifying tumour types and subtypes lacking adequate model representation, providing guidelines for new in vitro model development. Moreover, it boosts pharmacogenomic associations whose significance is diluted when accounting for low‐quality models (Najgebauer et al , 2020 ; Salvadores et al , 2020 ; Peng et al , 2021 ; Zhang & Kschischo, 2021 ). For instance, (Salvadores et al , 2020 ) identified drug sensitivity markers across cancer types using only “golden sets” of CCLs strongly resembling their cancer type of origin based on transcriptomic and epigenomic profiles.…”

“…For instance, Celligner (Warren et al , 2021 ) examines the intra‐cluster variability in a corrected gene expression space that integrates CCLs and tumours, finding this reflective of known tumour subclassifications for breast, kidney, leukaemia and skin cancer. Conversely, supervised methods (Sinha et al , 2017 ; Yu et al , 2019 ; Salvadores et al , 2020 ; Peng et al , 2021 ; Zhang & Kschischo, 2021 ) use prior known subtype labels of patient tumours to build a classifier and then predict the subtype of CCLs, similarly to the strategy adopted in the pan‐cancer approaches described in the previous section. Unsupervised strategies based on a correlation metric ranked CCLs based on their average similarity to tumour subtypes (Sinha et al , 2021 ; Fang et al , 2021 ; Liu et al , 2019b ; Vincent & Postovit, 2017 ).…”

“…This strategy prevents capturing interactions between omics and a proper representation of underlying biological mechanisms. A more refined strategy, intermediate integration , is developed in Maui (Ronen et al , 2019 ) and MFmap (Zhang & Kschischo, 2021 ). Briefly, Maui considers GE, CNA, and Mut omics and builds a multimodal stacked variational autoencoder (VAE) to represent tumour samples and CCLs in a low dimension latent space, with the latent factors regarded as higher‐order genomic features.…”