Predicting ligand-dependent tumors from multi-dimensional signaling features

Abstract

Targeted therapies have shown significant patient benefit in about 5-10% of solid tumors that are addicted to a single oncogene. Here, we explore the idea of ligand addiction as a driver of tumor growth. High ligand levels in tumors have been shown to be associated with impaired patient survival, but targeted therapies have not yet shown great benefit in unselected patient populations. Using an approach of applying Bagged Decision Trees (BDT) to high-dimensional signaling features derived from a computational model, we can predict ligand dependent proliferation across a set of 58 cell lines. This mechanistic, multi-pathway model that features receptor heterodimerization, was trained on seven cancer cell lines and can predict signaling across two independent cell lines by adjusting only the receptor expression levels for each cell line. Interestingly, for patient samples the predicted tumor growth response correlates with high growth factor expression in the tumor microenvironment, which argues for a co-evolution of both factors in vivo.

Fig. 1

Proliferation screen across 58 cell lines. a Ligand/Receptor and antagonistic antibodies used in the in vitro proliferation screen. b Results of the proliferation screen across 58 cell lines. Dots mark a significant increase in ligand induced proliferation or decrease in the presence of ligand plus antibody. The ligand effect is normalized to the medium control, whereas the antibody plus ligand effect is relative to ligand alone. The two cell lines marked with an arrow, as well as five additional cell lines that were not included in the proliferation screen, were used to train the computational model to signaling data. c Correlation pattern of ligand and antibody effects across all cell lines. d Linear correlation of receptor expression to ligand induced proliferation. The proliferation in response to ligand (y-axis) is displayed as log10-fold change with respect to day 0. The receptor surface levels (x-axis) are absolute measurements of receptors/cell by qFACS on a log10-scale

Fig. 2

Structure of computational signaling model. a The receptors EGFR, HER2, ErbB3, Met, and IGF-1R can form several homo and heterodimers after ligand binding. b In the model, receptors are synthesized and either dimerize spontaneously or bind a ligand to form homo- and hetero-dimers, which results in trans-phosphorylation of the receptors. Activated receptors signal downstream and are prone for internalization, which leads to either degradation or dephosphorylation by a phosphatase followed by recycling to the cell surface. Downstream, the MAPK and PI3K cascade activate S6K1 and ultimately converge in the phosphorylation of S6. The MAPK and PI3K signaling pathways are interconnected via multiple crosstalk mechanisms

Fig. 3

Importance of receptor surface levels for model response, shown for a selection of calibration cell lines. a Cell line dependent signaling features: Model response to EGF stimulation of two different cell lines resulting in sustained or transient receptor phosphorylation in the BxPc-3 and IGROV-1 cells. Their respective receptor surface levels are shown on the left. The model fits are represented by the colored lines with respective uncertainties (67% confidence intervals) as shades. Data points are shown as dots in the same color. b Model fits for the cell line ACHN stimulated with HGF, EGF and the combination. c Model response to co-stimulation of EGF plus HRG in comparison to the stimulation with EGF, HRG or IGF-1 alone in H322M cells

Fig. 4

Strategies for predicting ligand-induced phenotypic response. Based on the receptor expression of individual cancer cell lines, either a univariate or multivariate approach can be used to predict the phenotypic response to ligand stimulation. a Univariate approaches relate the respective receptor expression to the observed ligand induced proliferation for each of the four ligands separately. b–c Multivariate approaches such as bagged decision trees (BDTs) relate high-dimensional feature sets to the observed phenotype. b In this case the feature set consists of the five receptor surface levels as well as information about the respective ligand stimulation and mutation status. c The calibrated and validated signaling model allows to simulate the expected signaling dynamics for each individual cell line based on its receptor expression and ligands present. Based on the mechanistic knowledge that the signaling model incorporates, it can expand the initial five-dimensional feature set to a 12-dimensional feature set. This expanded feature set, together with information about mutation status is now connected to the observed growth responses by a bagged decision tree

Fig. 5

Prediction of ligand-induced proliferation using BDTs. a Ratio of true predictions after BDT training with simulated signaling features or receptor expression only, compared to random predictions in the presence of EGF, HRG, IGF or HGF. b For 500 random splits of training and testing cell lines, the BDT outcome is compared to random growth assessment as histogram and cumulative density function, showing the significant improvement due to mechanistic modeling. c Data of in-vitro cell viability screen showing proliferation response (green) or no significant response (red) in different 2D representations of the feature space

Fig. 6

Predicting ligand dependent tumors from the TCGA data set. a Predicted percentage of tumors that would response to ligand exposure. The predictions were obtained by using the receptor RNA expression measured in breast, colorectal, lung, and ovarian cancers as inputs to our model. b The measured RNA expression of the ligands in predicted responders (red) vs. non-responders (green). The mean expression (black horizontal lines) and statistical significance of differences is indicated. The receptor mRNA expression is measured in transcripts per million and is displayed on a log-2 scale