Identifying commonalities between cell lines and tumors at the single cell level using Sobolev Alignment of deep generative models

Soufiane Mourragui, Joseph C. Siefert, Marcel Reinders, Marco Loog, Lodewyk F.a. WesselsbioRxiv 2022 · DOI

TL;DR

This paper introduces Sobolev Alignment (SA), a novel computational framework that uses deep generative models and kernel methods to identify shared biological processes between cancer cell lines and patient tumors at the single-cell level. This approach overcomes limitations of existing methods, enabling better understanding of cancer biology and more translatable preclinical models for drug development.

What Problem Does This Paper Address?

Cancer research heavily relies on preclinical models like cell lines to study disease mechanisms and test therapies. However, these models often differ significantly from actual patient tumors due to factors like the lack of a tumor microenvironment, leading to poor translation of findings to the clinic and high attrition rates in drug development. Existing computational methods struggle to effectively align single-cell RNA sequencing (scRNA-seq) data from cell lines and tumors, obscuring shared biological processes. This paper aims to bridge this gap by developing a method to robustly compare and integrate these disparate data sources.

Methods

The core of the proposed method, Sobolev Alignment (SA), involves several key steps:

  1. Deep Generative Modeling (VAE): Two independent Variational Auto-Encoders (VAEs), specifically scVI, are trained on cell line and tumor datasets separately. These VAEs learn to represent complex, non-linear gene expression patterns as a set of "latent factors," effectively capturing biological variability while accounting for scRNA-seq specific issues like dropout and library size.
  2. Kernel Approximation: The mapping from gene expression to latent factors learned by the VAEs is approximated using Falkon, a large-scale kernel ridge regression tool. This step leverages kernel machines for their interpretability and ability to handle large datasets.
  3. Latent Factor Comparison (Sobolev Principal Vectors - SPVs): The approximated latent factors from cell lines and tumors are compared using cosine similarity. This similarity matrix is then decomposed to derive Sobolev Principal Vectors (SPVs), which represent pairs of biological processes (one from cell lines, one from tumors) ranked by their similarity.
  4. Interpretation: The SPVs are interpreted using a Taylor expansion of the Gaussian kernel, allowing for the calculation of gene and interaction term contributions to each SPV. Gene Set Enrichment Analysis (GSEA) is then applied to these contributions to identify conserved biological pathways.

Datasets

  • Kinker dataset: A panel of 53,512 single cells from 198 cell lines, including 11,105 single cells from 32 NSCLC cell lines.
  • Kim dataset: A panel of 208,506 cells from 58 NSCLC cancer patients, including 36,467 epithelial cells.
  • Synthetic datasets: Three synthetically generated datasets with varying levels of overlap between source (cell line) and target (tumor) signals were used to validate the method's capacity to discriminate shared from specific biological processes.
  • McFarland dataset: A multiplexed perturbation screen on 33 NSCLC cell lines exposed to 19 anti-cancer drugs, used to study drug modes of action.

Key Insights and Ideas

  • Cell lines and tumors are not just batch effects: Standard batch correction methods (Seurat, Harmony, LIGER) fail to adequately integrate cell line and tumor data, highlighting fundamental biological differences beyond technical artifacts.
  • Sobolev Alignment captures shared biology: SA successfully identifies conserved biological processes (e.g., mitotic and immune-related pathways) between cell lines and tumors, even when traditional methods fail.
  • SPVs as biomarkers: The derived Sobolev Principal Vectors (SPVs) can serve as biomarkers, relating clinical samples to cell line models.
  • Drug mode of action deciphered: SA can be applied to drug perturbation data to recapitulate known drug mechanisms and investigate the action of uncharacterized drugs.

Key Results

  • SA demonstrated improved co-clustering of cell lines and tumors when combined with Mutual Nearest Neighbor (MNN) correction, outperforming standard batch correction methods.
  • The method successfully predicted EGFR signaling levels and cell cycle states in tumor cells based on their neighborhood in the SA+MNN space, validating the biological relevance of the alignment.
  • Analysis of NSCLC cell lines and tumors revealed conservation of key mitotic pathways (e.g., M phase) and immune-related pathways (e.g., interferon signaling, adaptive immune system).
  • When applied to drug perturbation data, SA recapitulated known modes of action for Idasanutlin, Everolimus, and Trametinib, and provided insights into the mechanism of action for the unknown drug BRD3379, suggesting it acts similarly to BRD4 inhibitors.

Limitations and Future Work

  • Interpretation of VAE latent factors: While the kernel approximation helps, interpreting the direct contributions of genes to VAE latent factors can still be challenging.
  • Kernel Ridge Regression limitations: The authors note that kernel ridge regression might not be ideal for approximating functions with large localized variations, suggesting future work could explore alternative regression methods.
  • Generalizability: The authors suggest that the SA framework is versatile and could be adapted for other molecular features beyond gene expression, such as chromatin accessibility or protein levels, and for more complex models like organoids or patient-derived xenografts.
  • Improving interpretability: Further theoretical work on kernel methods, particularly for Laplacian kernels, could enhance the interpretability scheme.

Summary updated

Abstract

Abstract Preclinical models are essential to cancer research, however, key biological differences with patient tumors result in reduced translatability to the clinic and high attrition rates in drug development. Variability among and between patients, preclinical models, and individual cells obscures commonalities which could otherwise be exploited therapeutically. To discover the shared biological processes between cell line models and clinical tumors we developed Sobolev Alignment , a computational framework which uses deep generative models to capture non-linear processes in single-cell RNA sequencing data and kernel methods to align and interpret these processes. We show that our approach faithfully captures shared processes on a set of three synthetic datasets. Exploiting two large panels of untreated non-small cell lung cancer cell lines and patients, we identify the similarities between cell lines and tumors and show the conservation of key mitotic and immune-related pathways. Employing our approach on a large in-vitro perturbation screen, we show that processes captured by our method faithfully recapitulate the known modes of action of clinically approved drugs and allow investigation into the mode of action of an uncharacterized drug.
BibTeX
@misc{mourragui2022identifying,
  author = {Soufiane Mourragui and Joseph C. Siefert and Marcel Reinders and Marco Loog and Lodewyk F.A. Wessels},
  title = {{Identifying commonalities between cell lines and tumors at the single cell level using Sobolev Alignment of deep generative models}},
  year = {2022},
  doi = {10.1101/2022.03.08.483431},
  howpublished = {bioRxiv},
}
Figure