Conceived and designed the experiments: TRB GSF JC. Analyzed the data: TRB UR. Contributed reagents/materials/analysis tools: TRB FPB. Wrote the paper: TRB JC.
The authors have declared that no competing interests exist.
Prediction of drug action in human cells is a major challenge in biomedical research. Additionally, there is strong interest in finding new applications for approved drugs and identifying potential side effects. We present a computational strategy to predict mechanisms, risks and potential new domains of drug treatment on the basis of target profiles acquired through chemical proteomics. Functional protein-protein interaction networks that share one biological function are constructed and their crosstalk with the drug is scored regarding function disruption. We apply this procedure to the target profile of the second-generation BCR-ABL inhibitor bafetinib which is in development for the treatment of imatinib-resistant chronic myeloid leukemia. Beside the well known effect on apoptosis, we propose potential treatment of lung cancer and IGF1R expressing blast crisis.
Protein interaction data are accumulating rapidly and, although imperfect and incomplete, they provide a valuable global description of the complex interplay of proteins in a human cell. In parallel, modern proteomics technologies make it possible to measure in an unbiased manner the protein targets of a drug. Such data reveal multiple targets in a view that contrasts with a previously prevalent paradigm that drugs had single – or a very limited number of – targets. In this context of newly available systems level data and more precise and complete information about drug interactions, it is natural to try to determine the global perturbation exerted by a drug on a human cell to identify potential side effects and additional indications. We present a computational method that aims at making such predictions and apply it to bafetinib, a recently developed leukemia drug. We show that meaningful predictions of additional applications to other cancers or resistant cases and likely side effects are obtained that are not straightforward to determine with existing algorithms. Our method has a strong potential to be applicable to other drugs.
Biomedical research is changing towards a systems pharmacology view of drug action
(A) The drug (blue) bound to a matrix (grey) retains the drug targets (green) and secondary binders (orange). Most unspecific proteins (red) are washed of but some could stick to the matrix. The retained proteins are analyzed with MS/MS. (B) If the soluble drug is supplemented, it blocks the binding pocket of its target yielding a reduced amount of pulled-down proteins that are specific drug binders. Only sticky unspecific proteins and weak drug targets are retained.
Alternatively, precise modeling of perturbations which change the protein interaction network has the potential to predict new drug targets and to provide a detailed mechanism of action simultaneously
We applied our algorithm to the bafetinib (NS-187, INNO-406) target profile. Bafetinib is a small molecule tyrosine kinase inhibitor in development for chronic myeloid leukemia (CML)
Our computational approach to predict the impact of bafetinib on a functional network is based on the human protein-protein interaction network, on the annotation of its nodes and on a drug target profile associated with an affinity measure.
The network is constructed from protein-protein interactions found in the public interaction databases HPRD, MINT, Intact, DIP and BioGRID
The human network of all known protein-protein interactions is associated with its biological processes of gene ontology (GO) derived from UniProtKB and Entrez Gene
The recently published drug target profile of the kinase inhibitor bafetinib measured in the cell line K562 is used
Bafetinib can impact the uniform functional sub-networks in two ways via its targets (
The drug inhibits directly a node of the uniform functional sub-network.
The drug target interacts with the uniform functional sub-network at its periphery. To consider also complexes and cascades, drug targets which are linked via other drug targets to the functional sub-network are included.
This network includes the uniform functional sub-network (triangular nodes) which shares one biological function and all drug targets (grey nodes) interacting with it. Bafetinib (Bafe) can impact nodes (red border) in the uniform functional sub-networks in two ways: Either the drug inhibits directly a node in the uniform functional sub-network (1) or it modulates the function through peripherally interacting drug targets (2).
It is difficult to predict which mode of perturbation has a higher impact. Directly inhibiting a pivotal sub-network member can completely disrupt a function. Nonetheless, biological signaling networks often have multiple alternative routes and protein isoforms to rescue the cell. Drug targets acting at the periphery can modify significantly the function through interaction or modulation of a modification, e.g., phosphorylation. By this mechanism, the inactivation of different branches and isoforms is possible. Furthermore, functional boundaries are often loosely defined and incompletely annotated. We thus treat both perturbation modes equally and therefore the perturbed functional sub-network (
In the protein-protein interaction (edges) network, the Bafetinib profile (grey nodes) perturbs the biological process (triangular nodes) which is pivotal in BCR-ABL dependent CML. The drug affinity (at) is indicated by the node size. Kinases in the target profile have a red label. Proteins of the uniform functional sub-network interacting with inhibited kinases are shown with a red node border. K562 cells contain ABL1 and its fusion protein BCR-ABL which is not found by the algorithm in this sub-network. However, ABL1 pulldown is hidden by BCR-ABL and hence missed as target. Western plots proved ABL1 as a competed target
We define a score
The first feature describes how frequent the annotation is present in the sub-network. Peripheral drug targets don't share the functional annotation (
The second feature puts the drug impact in relation to the sub-network size. Generic biological functions result in very big sub-networks, in which the drug targets play overall no important role anymore. Furthermore, the drug should preferentially perturb a function at several different points. Hence, the proportion of the number
Lastly, the binding affinity
The affinity of bafetinib to its targets is used to score the impact on sub-networks in equation (1). The higher the protein amount in mass spectrometry analysis, the higher the number of different detected peptides covering the protein sequence
An empirical p-value is calculated via randomization of the interactome. First, the interaction partners of each node are randomly selected. It is ensured that the degree of each node remains constant. Second, the annotation is randomly assigned to the nodes, while the total number of each term is preserved. The presented algorithm is applied to 500 random instances of the interactome. The empirical p-value is calculated from the fraction of randomized interactomes containing a sub-network with a score equal or better to the tested score divided by the total number of random instances. The highest score of all the random instances
The presented approach is programmed in the statistical environment R/Bioconductor and available at
Networks are visualized with Cytoscape
We present a novel strategy to analyze the mechanisms of action of bafetinib. The target profile is weighted with respect to its drug affinity and its impact on protein interaction networks is scored. Ten perturbed functional sub-networks are scored higher than any sub-network of the 500 randomized interactomes (
The drug profile interferes with many nodes of the uniform function sub-network (triangular nodes). The drug affinity is indicated by the node size (a large node means high affinity). Kinases in the target profile have a red label. Proteins of the uniform functional sub-network interacting with inhibited kinases are shown with a red node border. EGFR expressing cells are not known to carry the fusion protein BCR-ABL which diminishes the influence of BCR-ABL on the network (dashed lines).
Perturbed functional sub-networks | Nodes | Targets (competed) | Score |
Epidermal growth factor receptor signaling pathway | 57 | 25 (7) | 0.213 |
Insulin receptor signaling pathway | 68 | 26 (7) | 0.198 |
Aging | 96 | 29 (7) | 0.185 |
Regulation of MAP kinase activity | 94 | 22 (7) | 0.162 |
Induction of apoptosis by intracellular signals | 53 | 19 (5) | 0.135 |
Extracellular structure organization and biogenesis | 81 | 21 (6) | 0.135 |
MAPKKK cascade | 188 | 29 (8) | 0.131 |
Heart development | 104 | 21 (7) | 0.127 |
Protein amino acid autophosphorylation | 71 | 21 (5) | 0.126 |
Cell cycle arrest | 73 | 21 (5) | 0.125 |
(Randomization: N = 500; p-value<0.002).
Inactivated apoptosis signaling plays a pivotal role in BCR-ABL dependent CML pathogenesis
The impact of bafetinib on apoptosis in CML is manifested with 5 targeted kinases at the periphery (
Even if we know that the drug has an inhibitory effect on the target kinases, we cannot predict without additional knowledge whether missing phosphorylation has an enhancing or decreasing effect on the biological process. The constitutively active kinase BCR-ABL results in a strong anti-apoptotic phenotype. Inhibition counteracts this behavior
The top ranked perturbed functional sub-network is based on the epidermal growth factor receptor (EGFR) signaling pathway (
EGFR is not expressed in hematopoietic cells (such as K562) but this sub-network strongly suggests that bafetinib has the potential to interfere with EGFR signaling for instance in lung cancer cells. Recently, it was shown through the combination of chemical proteomics, phosphoproteomics and functional genomics that dasatinib, a broad-spectrum kinase inhibitor, leads to apoptosis in lung cancer cells via inhibition of SRC, EGFR, FYN and, notably, LYN
Second highest is the perturbation of insulin receptor signaling pathway. It was suggested that bafetinib, CGP76030 and nilotinib might overcome imatinib resistance in blast crisis patients which feature BCR-ABL gene amplification
A potential side effect of several tyrosine kinase inhibitors, like sunitinib and dasatinib, is an increased risk for cardiotoxicity
We validated the robustness of the algorithm by following the rank of the biological process upon leaving-one-out (supplementary
Furthermore, we investigated the effect of hubs on the sub-network ranks, which might exert an influence on the phenotype upon inhibition. It is not clear whether hubs are, in the context of our analysis, highly important or “general signal diluters”. Therefore, we weighted up and then down the affinity of the targets by log10 of their node degree. Multiply by this factor, i.e. increasing hubs importance, the top 6 sub-networks remain unchanged and “cell cycle arrest” even improved its rank by one. The others sub-networks were substituted by “response to insulin stimulus”, “response to peptide hormone stimulus” and “cellular response to hormone stimulus”, which are in line with insulin receptor signaling. Upon down-weighting by division, the top 4 sub-networks still remained unchanged. We conclude for robustness and reasonable independence of local topology. In other words, the function of the sub-network at hand seems to play a strong role in scoring, which is appropriate.
To show the general interest of our method we applied the algorithm to data we published recently analyzing lung cancer (HCC297) treatment with dasatinib
In comparison to our approach, classical GO enrichment analysis (p-value<0.01) of the 33 bafetinib drug targets result in 33 significant biological processes with high redundancy in the GO tree (supplementary
In contrast to GO enrichment analysis, the presented method does not as much rely on accurate annotations. Possible missing annotations of drug targets interacting at the periphery with a functional sub-network have only a minor effect on the score. However, we would like to point out that boundaries of pathways and biological processes are very diffuse. Crosstalk between different signaling cascades and metabolic pathways is essential for a living cell. Integrating protein interactions to peripheral drug targets provides a way out of this dilemma and can catch therefore more relevant processes than GO enrichment. Alternatively, augmenting the drug target profile with their direct interactors, results in a set of 831 proteins. GO enrichment analysis (p-value<0.01) of this set results in 676 biological processes (
Competition experiments in chemical proteomics provide an additional layer of security to the drug target profile. Secondary and unspecific binders are difficult to distinguish from true drug targets. They are often similar in the range of peptide counts and other properties. The competition with a soluble drug and our affinity score helps in identifying biological target proteins. Interestingly, unspecific binders influence the perturbation algorithm only marginally since the proteins are dispersed all-over the interactome and have no affinity to a specific uniform functional sub-network. Furthermore, their binding affinity score is 0. On the contrary, secondary binders of true drug targets increase the crosstalk to the functional sub-network which is attacked by the true target. Hence they can be used advantageously embedding the true targets in a specific context.
In conclusion, we identified successfully known mechanisms in CML as well as potential new applications and possible side-effects. We believe that the proposed computational approach can shed light in mechanisms of other drugs including highly promiscuous compounds and when soluble compound competition data are lacking. Hence, we provide an R package at
The bafetinib targets (grey nodes) disrupt the insulin receptor signaling pathway. The drug profile interferes with many nodes of the uniform function sub-network (triangular nodes). The drug affinity is indicated by the node size (large node equals high affinity). Kinases in the target profile have a red label. Proteins of the uniform functional sub-network interacting with inhibited kinases are shown with a red node border.
(0.47 MB TIF)
The bafetinib targets (grey nodes) disrupt the heart development suggesting putative risk factors. The drug profile interferes with many nodes of the uniform function sub-network (triangular nodes). The drug affinity is indicated by the node size (large node equals high affinity). Kinases in the target profile have a red label. Proteins of the uniform functional sub-network interacting with inhibited kinases are shown with a red node border.
(0.49 MB TIF)
Leave-one-out analysis. The ranks of the first five subnetworks (
(0.11 MB PDF)
Bafetinib (INNO-406) drug target profile.
(0.06 MB DOC)
Significantly hit sub-network by the highly promiscuous drug dasatinib.
(0.12 MB XLS)
Classical gene ontology (GO) enrichment analysis (p-value<0.01). The secondary and unspecific binders have a large influence on GO enrichment analysis.
(0.04 MB DOC)
Classical gene ontology (GO) enrichment analysis on all direct interactors of the drug profile (p-value<0.01).
(0.40 MB XLS)
We thank Eric Haura for datasets and all our CeMM colleagues for help and discussions during this research project.