Predicting patient response with models trained on cell lines and patient-derived xenografts by nonlinear transfer learning

Abstract

Preclinical models have been the workhorse of cancer research, producing massive amounts of drug response data. Unfortunately, translating response biomarkers derived from these datasets to human tumors has proven to be particularly challenging. To address this challenge, we developed TRANSACT, a computational framework that builds a consensus space to capture biological processes common to preclinical models and human tumors and exploits this space to construct drug response predictors that robustly transfer from preclinical models to human tumors. TRANSACT performs favorably compared to four competing approaches, including two deep learning approaches, on a set of 23 drug prediction challenges on The Cancer Genome Atlas and 226 metastatic tumors from the Hartwig Medical Foundation. We demonstrate that response predictions deliver a robust performance for a number of therapies of high clinical importance: platinum-based chemotherapies, gemcitabine, and paclitaxel. In contrast to other approaches, we demonstrate the interpretability of the TRANSACT predictors by correctly identifying known biomarkers of targeted therapies, and we propose potential mechanisms that mediate the resistance to two chemotherapeutic agents.

Meta-analysis of preclinical pharmacogenomic studies to discover robust and translatable biomarkers of drug response

“…Given the many differences between immortalized cancer cell lines and in situ tumours, it is expected that the expression of genes and their relationship with drug sensitivity phenotypes will not be conserved between the two contexts, even for the same cancer types. Published methods exist to “align” cell line and tumour samples [65], or alternatively to find a transformation of gene expression measurements that captures common variation between cell line and tumour expression patterns [66, 67]. However, none of these methods can identify which particular genes are conserved in their expression patterns, or give a probability to predictive single-gene marker translating to the clinical context.…”

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

“…We previously introduced two computational approaches to compare cell line and tumor gene expression profiles in an unsupervised manner: PRECISE 11 and TRANSACT 13 , respectively based on PCA and kernel PCA. These two dimensionality reduction methods, however, do not account for the specific properties of scRNA-seq data, such as zero-inflation or over-dispersion, which limit their applicability to single-cell data.…”

“…Several computational studies have already attempted to characterize the molecular differences between cell lines and patients. The first category of approaches consist of designing machine learning tools to capture the common information relevant for transferring biomarkers of drug response [9][10][11][12][13][14] . A second category of approaches compares the genomic landscapes of cell lines and tumors in an unsupervised fashion [15][16][17] .…”

“…Using a panel of lung cancer tumors and a panel of cell lines, we first show that the differences between cell lines and tumors cannot be modeled as a classical batch effect. As previous methods comparing cell lines to tumors were not designed for scRNA-seq data 11,13 , we developed a novel computational approach, Sobolev Alignment (SA), which mobilizes recent advances in large-scale kernel-based machine learning [22][23][24] and deep-learning-based probabilistic modelling of scRNA-seq profiles [25][26][27] . We show that the application of SA to three simulated single cell datasets accurately captures the (constructed) shared biology between datasets.…”

Computational estimation of quality and clinical relevance of cancer cell lines

“…For example, Geelher et al (Geeleher et al , 2014 ) used batch correction (ComBat) on gene and miRNA expression data between in vitro model and patient tumour to build a predictive model of drug sensitivity on CCLs and further validated the predicted drug response in primary tumours with respect to known clinical trial results. More recent methodologies such as PRECISE (Mourragui et al , 2019 ) and TRANSACT (Mourragui et al , 2021 ) first learn a shared feature subspace (linear and non‐linear, respectively) and then use it to build a predictive model for drug response. Sharifi‐Noghabi et al ( 2020 ) proposed an adversarial inductive transfer learning method that focuses on discrepancies in both gene expression (input) and drug response (output), adapting both aspects in the two different domains.…”