Mapping the perturbome network of cellular perturbations

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.

( A to D ) Drug modules with higher interactome localization are characterized by higher similarity of their respective targets in terms of GO term annotations of (A) cellular component, (B) biological process, (C) molecular function, as well as in terms of (D) annotated diseases. Drug module localization is quantified by the z -score of the largest connected component size of the respective drug targets relative to random expectation.
( E ) Comparison of the interactome overlap between two drug modules and the side effects they are associated with. Drugs that have at least one common side effect are characterized by more overlapping interactome modules compared to drugs that do not share any side effects. ( F ) Interactome overlap versus number of shared side effects across all drug pairs. Bars in A-F indicate the mean over all measurements, error bars show the 95% confidence interval. ( A ) Calculation of the significance cutoff for individual wells that was used to analyse morphological and biological similarity. We used the mean vector length plus two standard deviations (red line) of DMSO treated wells (grey). ( B ) Similarity between two wells treated with the same drug (blue) and two randomly picked drugs (grey). ( C ) Similarity between two drugs annotated with the same mechanism of action (blue) compared to two randomly picked drugs (grey). ( D ) Similarity between two drugs with same ATC indications compared to two randomly picked drugs. ( E , F ) Similarity between two drugs binned by their interactome-based distance d AB . Bars in B-F indicate the mean over all measurements, error bars show the 95% confidence interval.

Investigating cell population heterogeneity
Individual cells may respond differently and to varying degrees to a drug induced perturbation. As a consequence, population wide averages of a given morphological feature reflect both the heterogeneity in terms of the phenotypic effect of a drug and in terms of the affected proportion of cells. To quantify the relative contribution of these two effects, we examined the top ten drugs with the largest morphological impact more closely.
Supplementary Figure 15A Fig. 15D). The only exceptions were nuclei intensity features, whose distributions exhibited a mild bimodality ( Supplementary Fig. 15E). The position of the two maxima suggests that they represent cells in different phases of the cell cycle, i.e. containing 2C and 4C DNA amount, respectively. We next tested whether any of the drug treated populations showed feature distributions with different shapes compared to the respective distributions of the DMSO controls. To this end, we centered and rescaled all distributions by removing the mean and scaling to unit variance before computing the Kolomogorov-Smirnov test between the respective DMSO and drug treatment distributions ( Supplementary   Fig. 11F). The results summarized in Supplementary Fig. 15G show that for the vast majority of drug/feature pairs there are no significant differences between the shape of the DMSO and the shape of the drug treatment distributions. Only a few granularity features showed larger differences, which correspond to an increase in the number of cells with a zero value.
This increase is likely the result of the size of the virtual probe that is used to calculate the respective features and does not reflect the emergence of a separate cell subpopulation.
Taken together, we found no evidence for the widespread existence of distinct cell populations. We conclude that the observed morphological heterogeneity among cells of the same treatment is largely driven by effects at the cellular level, instead of reflecting an incomplete penetrance of the perturbation at the cell population level, i.e. a drug affecting only a fraction of all cells.
Comparing feature vector similarity and visual similarity Figure

KEGG
The Kyoto Encyclopedia of Genes and Genomes (KEGG) is a resource for understanding high-level functions and utilities of biological systems, such as the cell, the organism and the ecosystem, from molecular-level information, especially large-scale molecular datasets generated by genome sequencing and other high-throughput experimental technologies 13  contains information on marketed medicines and their recorded adverse drug reactions 16 .
The information is extracted from public documents and package inserts. The available information include side effect frequency, drug and side effect classifications, as well as links to further information, for example drug-target relations. We downloaded the database Nov.
2018 and used PubChem identifier to map our CLOUD drugs to the database.

Offsides
The Offsides database is a resource for off-label effects that are not not listed by the FDA 17 .
The version used in this study was obtained from the Offsides Webpage

Morphological effects across drug concentrations
The effects of drugs generally depend on their dosage. Most commonly, sigmoidal functions are used to model dose-response curves that interpolate between no effect observed at very these results indicate that an increase in drug concentration generally leads to a more pronounced morphological phenotype, which is characterized by a vector of increased length pointing in the same direction within the morphological space.
We next tested whether these findings can also be recapitulated through the high-dimensional interaction framework introduced. To this end, we performed a computational "sham experiment" as illustrated in Supplementary Fig. 10F:  Supplementary Fig. 10H) and no significant differences were observed between different concentrations and replicates or between different concentrations and random combination vectors. This indicates that the observed γ levels reflect technical variance and biological heterogeneity, rather than representing truly emergent morphological features. As described in the main methods part, we apply conservative thresholds to account for such random variations in order to identify only robust interactions and avoid false positives. In summary, we conclude that our general framework for identifying and classifying perturbation interactions using high-dimensional morphological readouts can also be applied across different drug concentrations, when suitable controls and conservative thresholds are used that take technical and biological variance into account.