DrugComb Documentation

DrugComb is an integrated data portal aimed at collecting, analysing and distributing the results of drug combination screens on cancer cell lines. Its main purpose is harmonisation of drug combination research results. Due to various experimental methodologies used in drug combination studies under the umbrella of personalised anticancer research, it is often problematic to integrate compound screening results from different studies. Various definitions of drug synergy used by researchers exacerbate the issue. As a result, in addition to scientifically challenging task of explaining and predicting effective drug combinations, researchers in the field end up having to invest considerable effort into data preparation and cleaning.

DrugComb data portal provides free access to experimental data for researchers and layperson alike with the purpose of being a unified source of information about drug combination screens. In addition to being a repository containing data from the existing drug combination studies, it allows raw data upload, semi-automated data quality control, calculation and visualisation of drug combination synergy values, dose-response heatmaps, single drug dose-response curves, box plots for CSS, and S score. Third party APIs provide access to STITCH database in order to access chemical-protein association networks between compounds; to PubChem in order to generate molecular structural formulae of compounds; and to ChEMBL for ligand-based target predictions using compounds as input. Future plans include prediction of novel, previously untested drug combinations using machine learning and network-centric tools using drug names and cancer types as input parameters.

Drug synergy is a dimensionless measure of drug-drug interaction. It assesses the degree of deviation of an observed response from the null hypothesis, that is an expected effect of non-interaction. Different mathematical models have different definitions of what synergy is, which results in varying synergy values for the same drug and cell line combination. In DrugComb data portal synergy is calculated using four different reference models:

• ● Bliss model [Bliss, 1939] assumes a stochastic process in which two drugs elicit their effects independently, and the expected combination effect can be calculated based on the probability of independent events as:

  \begin{equation} \label{eq:1} y_{BLISS}=y_1+y_2-y_1*y_2 \\ \end{equation} where $$y_{1,2} \in [0,1]$$ are single drug effects measured as fractional cell death or cell growth

• ● Highest Single Agent (HSA) [Berenbaum, 1989] states that the expected combination effect equals to the higher effect of individual drugs: $$yHSA = max(y1, y2).$$
• ● Loewe additivity model [Loewe, 1953] defines the expected effect yLOEW E as if a drug was combined with itself. Unlike the HSA and the Bliss independence models giving a point estimate using different assumptions, the Loewe additivity model considers the dose-response curves of individual drugs. The expected effect yLOEWE must satisfy:

  \begin{equation} \label{eq:3} \frac {x_1}{\chi_{LOEWE}^1} + \frac{x_2}{\chi_{LOEWE}^2} = 1 \end{equation} where $$x_{1,2}$$ are drug doses and $$\chi_{LOEWE}^1, \chi_{LOEWE}^2$$ are the doses of drug 1 and 2 alone that produce y_LOEWE. Using 4-parameter log-logistic (4PL) curves to describe dose-response curves the following parametric form of equation 3 is derived: where $$E_{min},E_{max} \in [0,1]$$ are minimal and maximal effects of the drug, $$m_{1,2}$$ are the doses of drugs that produce the midpoint effect of $$E_{min} + E_{max}$$, also known as relative $$EC_{50}$$ or $$IC_{50}$$, and $$\lambda_{1,2}(\lambda > 0)$$are the shape parameters indicating the sigmoidicity or slope of dose-response curves. A numerical nonlin- ear solver (such as index.htmlfor R or optimize.nonlin.html for Python) can be then used to determine yLOEWE for (x1, x2).

• ● Zero Interaction Potency (ZIP) [Yadav et al., 2015] calculates the expected effect of two drugs under the assumption that they do not potentiate each other:

CSS - drug combination sensitivity score is derived using relative IC50 values of compounds and the area under their dose-response curves. AAC [Yang et al., 2013] and DSS [Yadav et al., 2014] metrics have been used as an inspiration for the CSS derivation. CSS of a drug combination is calculated such that each of the compounds is used at a fixed concentration (background drug) and another is at varying concentrations (foreground drug) resulting in two CSS values, which are then averaged. Each drug’s dose-response is modelled using 4-parameter log-logistic curve, such that: \begin{equation} y = y_{min} + \frac{y_{max} - y_{min}}{1+10^{\lambda{(log_{10} IC_{50}- x')}}} \end{equation} where $$y_{min},y_{max}$$ are minimal and maximal inhibition and $$x'=\log_{10} x$$ The area under the log-scaled dose-response curve (AUC) is then determined according to where $$[c_1,c_2]$$ is the concentration of the foreground drug tested [Malyutina et al., 2019].

DrugComb data portal provides access to an ever-growing number of drug combinations. At the moment of writing there is 2276 unique drugs, 93 cell lines representing 10 tissues and over 430k unique drug combinations obtained from four different studies. Please see "Studies as data sources" section for more information on the studies used.

Upon initial access user is informed about the data statistics of DrugComb, including the number of drugs, the number of cell lines, the number of tissue types and the number of drug combinations. We considered a drug combination as an experiment where a drug combination has been tested with multiple doses on a certain cell line, resulting in a dose-response matrix. Therefore, if the same drug combination has been tested in multiple cell lines, each of dose-response matrices will be considered as one drug combination. There is a search bar that allows searching for a specific data type (Cell Line, Tissue, Drug name and Study source).

Search is initiated by clicking on the search bar and choosing the initial search category. The following categories are available: Cell line, Tissue, Drug, Study. Category value to include (“=”) or exclude from search (“!=”) should be given sequentially. Different categories could be combined using AND or OR operators. Manual entry of the search string or full text search at the moment of writing is not available.

It is possible to select user-refined categories using the panel on the left side. A histogram gives a visual overview of cell lines and the number of drug combinations per tissue, upon selection this overview gets updated.

Table view

By default 10 drug combinations are shown on the landing page of the Table tab (see below), however, this number can be increased up to 100 drug combinations per page.

It is possible to sort drug combinations in ascending or descending order using values in any of the columns.

Graph view

When a drug combination is selected, it will be directed to the Graph view which contains the graphical results of sensitivity, synergy and annotation, each of which consists of one sub tabs.

Sensitivity view

It gives access to the graphical results of drug combination sensitivity, including the histogram of the selected drug combinations across all the cell lines, the histogram of the selected cell line across all the drug combinations, the drug combination sensitivity (CSS) - synergy (S) bar plot with the selected drug combination highlighted, the full dose-response matrix and the monotherapy dose-response curves.

Synergy view

It gives access to the graphical results of drug combination synergy, including the synergy landscapes over the dose matrix that are determined using four mathematical models (ZIP, BLISS, LOEWE and HSA). The synergy scores are in the unit of % inhibition and color-coded in the range of -30 (green) to +30 (red). Higher synergy scores indicate more synergistic interaction that leads to higher % inhibition of cancer cells. See below

Annotation view

It gives access to annotation of the selected drug combinations in terms of their chemical structures, protein targets and affected biological pathways according to third-party databases including STITCH (http://stitch.embl.de/) , PubChem (https://pubchem.ncbi.nlm.nih.gov/) , and ChEMBL (https://www.ebi.ac.uk/chembl/) STITICH databse is accessed using the following parameters to refine the drug combination network: (edited)

• ● Threshold of significance for the interaction to be displayed (0 to 1000) = 550;
• ● Limit, as max number of nodes to return = 20;
• ● Network flavour = “evidence”, where various colours represent various nature of entities’ interactions;
• ● Species = 9606 for Human, please refer to http://www.uniprot.org/taxonomy for taxa information;

Network generated with the aforementioned parameters for Gefitinib and Memantine is shown below:

Pubchem view provides the compounds' skeletal formula. More information can be retrieved by the highlighted CID hyperlinks:

ChEMBL view provides the predicted drug targets with the probability > 0.3. The value column shows the binding affinity in nM. The target_accession column shows the UniProt ID (https://www.uniprot.org/) for the protein target. More information can be retried from ChEMBL via the highlighted ChEMBL IDs.