Mapping the perturbome network of cellular perturbations

Abstract

Drug combinations provide effective treatments for diverse diseases, but also represent a major cause of adverse reactions. Currently there is no systematic understanding of how the complex cellular perturbations induced by different drugs influence each other. Here, we introduce a mathematical framework for classifying any interaction between perturbations with high-dimensional effects into 12 interaction types. We apply our framework to a large-scale imaging screen of cell morphology changes induced by diverse drugs and their combination, resulting in a perturbome network of 242 drugs and 1832 interactions. Our analysis of the chemical and biological features of the drugs reveals distinct molecular fingerprints for each interaction type. We find a direct link between drug similarities on the cell morphology level and the distance of their respective protein targets within the cellular interactome of molecular interactions. The interactome distance is also predictive for different types of drug interactions.

Fig. 1

The perturbation landscape of chemical compounds in the human interactome. a Interactome consisting of 16,376 proteins and 309,355 physical interactions, with 1096 unique targets of the 267 drugs used in this study highlighted in yellow. b Exemplary subgraph for one drug and summary of collected annotations. c–e Drug library overview: covered classes of protein targets (c), affected pathways (d), and therapeutic classes (e). f Drug targets form localized modules in the interactome (Glass′ ∆ < −1 indicates strong localization). g–j Degree of localization correlates with similarity of biological processes (g), molecular function (h), cellular component (i), and known disease associations (j) of target proteins. k Distribution of interactome overlaps across all drug pairs (S_AB<0 indicates overlapping modules, S_AB<0 separate ones). l–o Interactome overlap is associated with drug similarity in terms of biological processes (l), molecular function (m), cellular component (n), and known disease associations (o). p Drugs belonging to the same therapeutic class are characterized by overlapping interactome modules. The bars in g–j, l–p indicate the mean over all measurements, error bars show the 95% confidence interval. q Representation of interactome-based drug relationships in a 3D perturbation space. Each sphere represents a compound, the diameter is proportional to its interactome localization, the overlap between two spheres is proportional to their module overlap S_AB. Five broad therapeutic disease classes are indicated in color. Modules of drugs that are used to treat similar diseases are co-located. We also observe overlaps between modules of drugs that are used to treat a particular disease and modules of drugs in which the respective disease may occur as a side effect

Fig. 2

Mathematical framework for a complete description of pairwise perturbation interactions. a Single readout measurements (number of features f = 1, e.g. cell viability) can only distinguish three interaction types between perturbations: positive, negative (e.g., synergy/antagonism), or non-interaction. b High-dimensional readouts (f > 2, e.g. cell morphology), enable the identification of two directed interaction types (positive and negative), describing whether the effect of a perturbation was increased or decreased, as well as one undirected type, describing emergent features that are not present in any of the individual perturbations. c Each perturbation (i.e., drug treatment) is associated with a specific cell phenotype. When two perturbations are combined, we identify the superposition of their individual effects as non-interaction, and any deviation as interaction, respectively. d Perturbations can be represented as vectors in the high-dimensional feature space, pointing from the unperturbed to the perturbed state. Two perturbations span a 2D plane S, which also contains the expected non-interaction (NI) phenotype. The NI point divides S into four quadrants that can be identified with different interaction types. Any measured phenotype of two combined perturbations can be decomposed into two components within S, which determine the directed interactions, and a third component perpendicular to S, indicating the emergence of entirely new features. e Summary of all 18 possible interaction types that may occur between two perturbations

Fig. 3

Imaging screen for identifying cell morphology changes induced by chemical perturbations and their combinations. a Overview of the imaging pipeline. We treated an epithelial MCF-10A cell line with 267 chemical compounds and all 35,511 pairwise combinations. Cells were plated in a 384-well plate, stained for nucleus, cytoskeleton, and mitochondria and imaged using automated high-throughput microscopy at 20× resolution. b Extraction of morphological features: images with technical artifacts are identified and removed using a machine learning approach. Cells are segmented and features are extracted using the software CellProfiler. We remove all features that are too noisy, not robust across replicates, show too little variation or are redundant, resulting in a 78-dimensional feature vector for each treatment. c The 78-dimensional morphological space projected onto its first two principal components. Each dot represents the cell morphology of a specific treatment (DMSO control, single drug, or drug combination). d Example images of drugs resulting in a strong morphological phenotype. e Evaluation of the morphological similarity for replicates of the same drug, drugs with the same mechanism of action (MOA) and drugs of the same therapeutic ATC class (*** denotes P value < 0.001, Mann–Whitney U test). f Morphological similarity versus interactome distance of the respective drug targets. The closer the targets of two drugs are on the interactome, the more similar are the morphological changes they induce in our cell line. Bars in e, f indicate the mean over all measurements; error bars show the 95% confidence interval

Fig. 4

The perturbome drug perturbation network. a The perturbome combines all 1832 identified interactions between 242 chemical compounds into a single network. Compounds resulting in strong morphological changes are colored in orange. Negative interactions are most frequent (red, 44%), followed by emergent (blue, 31%) and positive interactions (green, 21%). b Number of times that each of the 12 pairwise interaction types were observed. Most interactions are uni-directional, and bi-directional cross-talk is rare. c The degree-ordered adjacency matrix uncovers a pronounced core–periphery structure within the perturbome network. d The core consists mainly of negative interactions between drugs with strong morphotypes. e The majority of all observed interactions occur between the core and the periphery, often representing the modulation of a drug with a strong effect by a drug with a weak effect. f Interactions among drugs in the periphery are mainly emergent. g The degree distributions show the frequency of the number of neighbors per compound. There is a marked difference between the number of incoming and outgoing interactions. h Comparison of the number of observed interactions with randomized networks obtained from drug label randomization. Negative (positive) z-scores indicate that the observed number is smaller (larger) than expected by chance

Fig. 5

Linking cellular perturbation interactions with molecular and pathophysiological drug characteristics. a Individual drug characteristics do not correlate with the number of drug interactions. b Performance of a random forest machine learning classifier for predicting drug interactions from pairwise drug characteristics. c Relative importance of different feature classes for the predictions in b. d Distribution of interactome distances between the targets of two drugs that do not interact (gray) show any interaction (orange) and a positive (green), negative (red), or emergent (blue) interaction. The bootstrapping analyses confirms the significant differences among the respective means. The middle line in the boxplot displays the median, the box indicates the first and third quartile, whiskers the 1.5 interquartile range (IQR) (** and *** denote P values <0.01 and <0.001, Mann–Whitney U test). Outlier values are not displayed. e Interactome relationship between interacting perturbations: interacting drugs are generally characterized by overlapping modules. The extent of the overlap is predictive for the interaction type, from small overlap indicating emergent interactions to moderate and strong overlap indicating positive and negative interactions, respectively. f Summary of the relationships between the interactions observed on the cellular level and molecular or pathophysiological drug characteristics. Significant relationships with an interaction type are indicated by colored triangles that point up/down for enrichment/depletion compared to non-interacting drugs. g Interaction fingerprints highlighting how a given interaction type differs in its specific molecular and pathophysiological characteristics compared to the other interaction types. Numbers refer to the characteristics listed in f. Differences are quantified using Cohen’s D, d_C, as a measure of effect size. The black central represents no difference (d_C = 0) in the respective interaction type relative to all interactions; gray lines indicate changes in increments of one unit of d_C