- Clustering analysis of binarized drug response data reveals anti-cancer agents with similar patterns of response across breast cancer cell lines
The drug response matrix used in this study derives from published cell line viability data from 84 breast cancer cell lines tested in triplicate against nine concentrations of 90 therapeutic compounds, including conventional cytotoxic agents as well as targeted agents such as hormone and kinase inhibitors, often with overlapping activity (7). The concentration required to inhibit growth by 50% (GI50) was used as the response measure for each compound.
As a first step in exploring the similarities and differences between cancer therapeutics, in terms of their patterns of sensitivity/resistance observed across breast cancer cell lines, we created a binarized response dataset using previously established GI50 dichotomization thresholds for each compound (7) and encoded treatment response as a binary variable (sensitive=1 or resistant=0). (See Methods for additional details.)
Unsupervised hierarchical clustering of this binary drug response data matrix (Figure 1, Supplemental Figure S1), examined along the columns, illustrates that cell lines show similar patterns of response based on related transcriptional subtype (e.g. most of luminal and ERBB2 amplified luminal cell lines cluster together). Claudin-low and certain basal subtypes appear most responsive to cell cycle and proteasome inhibitors, whereas luminal subtypes tend to be more sensitive to HDAC and PI3K pathway inhibitors.
Examined along the rows, the heatmap in Figure 1 reveals, as expected, that many compounds with similar and related modes of action tend to cluster together. For instance, six out of seven PI3K pathway inhibitors cluster strongly together, and closely to a cluster containing all five of the HDAC inhibitors. As well, platinum drugs cisplatin, oxaliplatin, and carboplatin cluster together, as do microtubule/spindle inhibitors paclitaxel, ispinesib, and ixebepilone. Thus, breast cancer cells with intrinsic sensitivity to a drug in the PI3K, HDAC inhibitor, DNA damaging platinum, or microtubule inhibitor class, also tend to be sensitive to other agents in that same class.
There are notable exceptions in the compound clusters. Interestingly, BEZ235 a dual ATP-competitive PI3K and mTOR inhibitor for p110α/γ/δ/β and mTOR (p70S6K) clustered away from the PI3K pathway inhibitors group and closer to a more diverse cluster containing MAPK/ERK pathway inhibitors and antimetabolites which are pathways strongly interconnected with mTOR signaling. Additionally, Nutlin-3a, a MDM-2 mediated stabilizer of tumor suppressor p53, clustered with a group on PI3K pathway inhibitors. Akt signaling is known to engage in control of MDM-2 mediated p53 levels and therefore the agent’s clustering highlights the similarities in underlying mechanism of action. These data suggest out-of-class treatment choices for cancers that have responded to an agent in the dominant class defining the cluster.
- Knowledge of intrinsic/prior response could significantly improve response prediction for over half of the drug panel
To more quantitatively address the question of which pairs or groups of agents have statistically significantly associated response profiles, we applied Fisher’s exact test to all pairs of agents. If two agents are significantly associated, knowledge of (intrinsic/prior) sensitivity or resistance to one of the agents could in principle be used to improve the prediction of sensitivity to the other drug on the panel.
After adjusting p-values for multiple testing, 57/90 drugs are significantly associated to at least one other agent (Fig 2A). The number of associated drug response-prediction partners per drug ranged from 0 to 11, with the most highly connected nodes being erlotinib, tykerb:IGF1R, and an AKT1-2 inhibtor from Sigma, likely reflecting the enrichment for PI3K pathway inhibitors on the panel (Fig. 1).
We identified 33 compounds that did not exhibit conditional sensitivity; predominantly comprised of non-specific antimetabolite drugs and MAPK/ERK inhibitors. The antimetabolite compounds that fall into this category were either highly toxic to most cell lines irrespective of sensitivity profile (e.g., methotrexate), or relatively non-toxic (e.g., ibandronate). Interestingly, when investigating the complete set of compounds belonging to both drug classes, the therapeutics that did have a partner significantly associated with response – only had a single one, e.g. MEK1/2 inhibitors AZD6244 and GSK1120212 have only one significantly associated drug partner.
- Nearly all significant associations predict sensitivity to different agents in the same or a similar class
The heatmap in Figure 2B shows the odds ratios (log(OR)) of responding to one drug given knowledge of (intrinsic/prior) response to another drug, for all drugs in the panel with at least one significant predictive pair (BH p<0.05).
As described in Methods, we use this OR matrix as a model for history-dependent response data, in the absence of a dataset derived from cells (or patients) treated sequentially with multiple treatments. This model assumes (1) past response to a drug, whether sensitive or resistant, is based on intrinsic rather than acquired characteristics and (2) a cell line/patient’s intrinsic sensitivity/resistance profile remains unchanged despite the passage of time and treatment, at the time of the next decision point.
In the heatmap red represents a positive association (OR>1), or similar pattern of response between agents. If an association is positive, (intrinsic/prior) sensitivity to one drug predicts sensitivity to a second drug; likewise, resistance to the first drug predicts resistance to the second drug. For example, (intrinsic/prior) sensitivity to carboplatin (column) predicts sensitivity to cisplatin (row, red). Overall, the clusters of response-associated agents fall into the broad classes of cytotoxic vs. targeted drugs (Fig 2b). Within the larger cytotoxic cluster are three subclusters: (C1) commonly used DNA damaging agents (e.g, platinums and anthracyclines), (C2) cell cycle kinase inhibitors (e.g., VX-680), and (C3) microtubule poisons including taxanes and ixabepsilone. The targeted agent group contains smaller clusters of response-associated agents featuring (C5) HER2/EGFR inhibitors; (C6) PI3K inhibitors; and (C4) a mixed group of AKT/mTOR/HDAC inhibitors (Fig 2B). Multiscale bootstrap resampling performed to estimate p-values for hierarchical clustering established clusters (C1-C3 and C5-C6) to be highly supported by the data with p-values < 0.05 (Supplemental Figure S4). Cell lines sensitive to one agent in any of these clusters are likely to be sensitive to others in the same cluster. Mixed cluster C4 was not highly supported but there are drug pairs within the cluster, such as rapamycin and vorinostat, with significantly associated responses BH p<0.05 (SI fig S2B).
Associations that might be less apparent in a clustered response heatmap such as Fig 1 are of the ‘cross response type’ variety, wherein (intrinsic/prior) resistance to one drug predicts sensitivity to a different drug or vice versa. This class of relationship is of special interest, because of its potential utility in identifying an effective next treatment for patients who were highly resistant to a prior therapy. However, of the 88 significant drug pair response associations (BH Fisher p<0.05), 85 represent similar profiles of response; and only 3 represent cross-response class associations (Supplemental Figure S2). These resistance-conditioned sensitivity pairs are alkylating antineoplastic agent cisplatin and an HDAC inhibitor LBH589; EGFR inhibitor erlotininb and neddylation inhibitor MLN4924, and anti-helminthic mebendazole and cell cycle inhibitor purvalanol A.
- Recommendation system built on in vitro data of drug response further reveals conditional patterns of drug response in breast cancer cell lines
In order to further refine history dependent response predictions for breast cancer cell lines, as well as create a list of the top recommended compounds, we utilized a recommender systems analysis as detailed in the Methods section. Applying this method to our binary response matrix we obtained a dataset with 200 non-redundant probability of drug response rules with a minimum confidence level of 0.8 and a minimum support of 0.2. Only rules with a one-to-one drug relationship were considered. For each rule, a lift value was calculated to indicate the strength of association between the antecedent drugs (LHS) and the consequent drug recommendations (RHS). This analysis enables us not only to investigate drug associations, but also consider four different types of associations: (1) if sensitive to drug A, then sensitive to drug B (discovered associations n = 92), (2) if sensitive to drug A, then resistant to drug B (n = 25), (3) if resistant to drug A, then also resistant to drug B (n = 70), and (4) if resistant to drug A then sensitive to drug B (n = 13).
The mined rules were used to create a directional network of dependencies between drugs, presumably mediated by relationships between mechanisms of action and the spectrum of therapeutic vulnerabilities in breast cancer. The high complexity of such a representation is better visualized over smaller subsets of the graph; therefore, the graph visualization in Figure 3A highlights the hundred strongest (highest lift) association rules. A complete list of mined rules is further provided in Supplemental Table 2. The arrows in the plot correspond with association type and begin from a label representing an observed antecedent (LHS, intrinsic) drug response (resistance or sensitivity), and point towards the consequent (RHS) drug response. For example, the strongest association according to this analysis is (lift = 2.262) supports an association between sensitivity to HER2/EGFR inhibitor Lapatinib (LHS), and consequent sensitivity to Erlotinib – another EGFR inhibitor (RHS).
The findings made by recommender system around sensitive-to-sensitive relationships, mirror closely observations from previous clustering analyses. All of the previously described drug categories (C1-C6, Figure 2B) can be traced back to node clusters in the recommender system graph (Figure 3A). For example, we can see proximal connectivity between groups of DNA damaging/cell cycle inhibitors like Carboplatin and Cisplatin, as well was Oxaliplatin, Doxorubicin, CPT-11, Epirubicin and Topotecan, mirroring findings of Cluster (C1) in Figure 2B; as well as PI3K pathway inhibitors found in group (C6) (Figure 2B), i.e. GSK2119563, GSK2126458, GSK1059615, Everolimus, Temsirolimus and GSK2141795.
There are a number of new findings in this analysis that are not observed with previous approaches, and they often revolve around associations with opposite response types (“if resistant, then sensitive” and “if sensitive then resistant”). These types of recommendations would be most valuable in clinical settings when next lines of treatment for patients with drug resistance. The recommender system analysis revealed thirteen associations of this specific type, the strongest being the relationships between HER2 inhibitor Erlotinib and proteasome inhibitor MLN4924. In this case, the recommender system suggests Erlotinib for MLN4924 resistant cell lines (lift = 2.07, confidence = 0.857) and vice versa, MLN4924 for lines resistant to Erlotinib (lift = 2.0454, confidence = 0.818). Interestingly, the system also recommends a proteasome inhibitor Bortezomib if intrinsic sensitivity to Lapatinib (HER2 inhibitor) was detected (confidence = 0.875, lift = 2.041) – an association not revealed in prior analysis (Figure 2B Fisher BH p= 0.076).