Cell Panel Profiling Reveals Conserved Therapeutic Clusters and Differentiates the Mechanism of Action of Different PI3K/mTOR, Aurora Kinase and EZH2 Inhibitors

This article is featured in Highlights of This Issue, p. 2823

In cell panel profiling, dose–response curves are determined of a compound in a panel of cell line proliferation assays, and this is an important tool to study the mechanism of action and selectivity of novel cancer therapies, in addition to find drug response biomarkers before initiating expensive and time-consuming experiments in animal models or trials in patients (1, 2).

Cell panel profiling has a long history as a tool in cancer drug discovery. Over the period between 1989 and 2014, tens of thousands of compounds have been profiled in the 60 cell line panel of the National Cancer Institute (NCI-60), which resulted in the discovery of bortezomib and eribulin (3, 4). In parallel, a 39-cell line panel was set up by the Japanese Foundation for Cancer Research (JFCR-39; refs. 5, 6).

With the advent of genomics, attention increasingly focused to coupling cell line biology to compound response. High-throughput cell line profiling showed that cell panels could identify patient subpopulations, for instance BRAF-mutant cell lines are in particular sensitive to BRAF inhibitors and EGFR-mutant cell lines are relatively sensitive to EGFR inhibitors (7, 8). As a result, even larger and more fully genetically characterized cell panels were developed, encompassing up to a thousand cell lines (9–11). The Genomics of Drug Sensitivity in Cancer (GDSC; ref. 9) and the Cancer Cell Line Encyclopedia (CCLE; ref. 10) analyzed 24 drugs in 1,036 cell lines and 138 drugs in 727 cell lines, respectively. Other studies include the Cancer Therapeutics Response Portal study of 481 drugs in 860 cell lines (CTRP; refs. 11, 12) and that of GlaxoSmithKline, of 19 compounds in 311 cell lines (GSK; refs. 8,13). These studies discovered, for instance, that PARP inhibitors were active in cells with the EWS–FLI1 translocation, which has led to clinical evaluation of PARP inhibitors in Ewing sarcoma (9, 14).

Despite the successes, the use of drug screening in large cell line panels has come under debate recently. As the shutdown of the NCI-60 panel was announced, many important drug candidates and selective tool inhibitors remain unscreened, making it harder to find the best tool inhibitors to probe new biological pathways (1, 15, 16). Second, a comparison of the CCLE and GDSC panels prompted a discussion about data reproducibility (17–21). It was pointed out that the rank correlation between IC₅₀s on the same drug in the same cell lines could vary between 0.1 and 0.6, with a median of 0.28, which is poor (17). One cause could be that the GDSC maintains universal dose ranges and extrapolates IC₅₀s if they fall outside the standard range. For example, for the potent compounds bortezomib and docetaxel, about 40% of the profiling IC₅₀s consist of extrapolated values. However, also when the area-under-the-curve (AUC) activity measures are used (as e.g. in refs. 10, 11) which are not extrapolated, rank correlations remained poor (17). A tell-tale sign is that, after data clustering by the GDSC and CTRP, the highly related taxanes docetaxel and paclitaxel fell into two different therapeutic subsets (9, 11). This is improbable and inconsistent with earlier profilings (4, 6).

In defense of the large screens, it has been argued that using rank correlation as measure for data consistency is not appropriate, and a Pearson correlation on log IC₅₀s should be used (20). In addition, the data serve their purpose of identifying genomic markers (20, 21). Care should be taken therefore not to dismiss these valuable data sets. Unfortunately, and in contrast to the NCI-60 and JFCR-39 screens, none of the large screens provide internal replicate data, which is essential to get insight into the limits of the reproducibility of cell line screening.

Here, we present a data set of inhibition profiles of 122 compounds on a panel of 44 or 66 cell line proliferation assays (Oncolines). Of these, 61 compounds are registered drugs, 42 are investigational drugs, while 19 are in discovery phase or are tool compounds. Compared with other profiles, this data set contains 38 inhibitors for which no data have been reported in the literature, including many inhibitors with novel mechanisms of actions, such as EZH2 inhibitors, BET inhibitors, isoform-selective PI3K inhibitors and Aurora kinase inhibitors (Table 1; Supplementary Tables S1 and S2). Compared with the larger panels, the panel shows higher data quality, as evidenced by the correlations of duplicate profiles of 16 chemically and mechanistically different inhibitors, over a period of three years. Clustering analyses of all library profiles revealed separate clusters of different therapeutic modalities, i.e., taxanes, platins, topoisomerase and EGFR inhibitors, which is highly consistent with earlier results (4, 7, 11) and new clusters comprising compounds sharing novel biochemical targets. This compound response database therefore serves as a valuable reference set that can be used to benchmark other cancer cell line panels and provides an up to date and unbiased view of therapeutic clusters available in antiproliferative therapies.

Cell lines were obtained from the American Type Culture Collection (ATCC) from 2011 to 2014 (Supplemental Table S1B) and cultured in ATCC-recommended media. All experiments were carried out within nine passages of the original vials from ATCC who authenticated all cell lines by short tandem repeat analysis. Cell proliferation assays were carried out as described (22) using ATPlite 1step (Perkin Elmer). Exposure time was 72 hours for all compounds, except for epigenetic compounds, for which exposure time was 120 hours (Supplementary Table S2). Percentage growth was calculated, relative to the growth of unexposed cells, and relative IC₅₀s were fitted using a four-parameter logistics curve. From these and seeding cell density data, absolute GI₅₀s were calculated (3, 22). For AUC calculations, all percent-effect data within the test range were summed. If ranges were adapted, AUC data were not calculated.

In the course of three years, 122 inhibitors were tested on 44 cell lines, and after a panel extension, on 66 cell lines (Fig. 1; Supplementary Table S1). Compounds were dissolved in DMSO generally 48 hours before testing. Cisplatin and carboplatin were dissolved on the day of testing. An overview of all IC₅₀ data is presented in Supplementary Table S3A.

When compounds were inactive, their IC₅₀ was set to 31,600 nmol/L. Finally, ¹⁰log IC₅₀s (in nmol/L) were used for all subsequent analyses. When compounds had been profiled twice, the data of the largest panel was chosen, and, if equal, the earliest data set.

To compare our profiles with each other and with literature, Pearson correlations between ¹⁰log IC₅₀s were used (ρ; Figs. 1 and 2). All further calculations were performed in R (23). Drug-sensitivity scores were calculated in the package DSS (24). Hierarchical clustering (Fig. 3) used the Ward method and 1 − ρ as distance measure. The resulting tree was validated with multiscale bootstrap resampling (R package pvclust; ref. 25). A cutoff for the minimally significant Pearson correlation between profiles (ρ_min = 0.29; P < 0.05) was obtained by converting the Pearson correlation to a t statistic (26). Then t was translated into P values with the Student t distribution function. Affinity propagation clustering (APC) was performed using the package apcluster, setting a fixed number of 43 clusters (Table 2; Supplementary Table S4A; ref. 27). Network trees were generated using the Fruchterman–Reingold algorithm, in the package igraph (Fig. 4; ref. 28). Principal component analysis was performed using the function princomp (Fig. 4; ref. 23).

The inhibitory activity of compounds on Aurora A and C (Carna Biosciences, Inc.) was determined with LANCE Ultra TR-FRET assays (PerkinElmer) at K_M,ATP and using ULight-labeled PLK substrate peptide (PerkinElmer, cat. no. TRF-0110; ref. 29). IC50s from 9-point dose–response curves were converted to K_D using the Cheng–Prusoff equation (Fig. 4D).

For surface plasmon resonance experiments, biotin-labelled kinases (Carna) were immobilized on streptavidin-coated chips (GE Healthcare; Cat. no. BR100531) on a Biacore T200 (GE Healthcare) to level of 4000 response units (RU) using Biacore buffer (50 mmol/L Tris pH 7.5, 0.05% (v/v) Tween-20, 150 mmol/L NaCl and 5 mmol/L MgCl₂), except PI3Kδ, which was immobilized to a level of 8,000 RU. Remaining streptavidin was blocked with biocytin. The kinetic constants of the compounds were determined by single-cycle kinetic experiments as described previously (29). Data reported are the geometric averages of two independent experiments (Fig. 4D and E).

Mutation and gene copy number data of our cell lines were based on COSMIC v.75 of the GDSC (9) and were additionally validated by DNA sequence analysis of a number of important cancer genes (Supplementary Table S1C). For the analysis in Fig. 4F–H, cell lines were classified as having a “wild-type” or a “mutated” genotype, and the logIC₅₀s in both sets were submitted to a t test, followed by a Benjamini–Hochberg correction for multiple testing (22, 30).

NCI-60 data were downloaded from https://dtp.cancer.gov/databases_tools/. All other data came from supplementary data from indicated references. IC₅₀s were transformed to units of ¹⁰log (nmol/L), in parallel with our data. In case of NCI-60 and JFCR-39, data of multiple replicates were averaged (Fig. 2).

The high-throughput Oncolines cell panel has been running over the past three years with 44, and later 66 cell lines according to well-defined protocols (22). Biological variation was monitored by the average proliferation rates of each cell line, showing overall stability of growth rates over the course of three years (Fig. 1A).

Variation in IC₅₀s was probed with the reference doxorubicin, which was chosen because it gives a potent, full sigmoidal dose–response curve in nearly all cell lines (Fig. 1B). The variation in ¹⁰logIC₅₀s of six replicate profilings of doxorubicin was comparable with that of the NCI-60 panel, and similar to reported values in standardized high-throughput assays (Supplementary Fig. S1A; ref. 31).

Next, Pearson correlations (ρ) between replicate compound signatures were investigated (20). Any Pearson correlation above 0.6 is considered “fair” and above 0.8 is considered “near-perfect” (17). The doxorubicin replicates show ρ = 0.99 after 6 days and 0.81 after 643 days (Fig. 1C; Supplementary Table S3B). The replicate profiling of a total of 15 diverse (candidate) drugs, separated by 392 to 854 days, shows correlations between 0.58 and 0.93 (average 0.8; Fig. 1C).

Because it has been claimed that IC₅₀ is not an optimal response metric (24, 32), we investigated if using other metrics would improve correlations. We calculated, from the same raw data, three other response measures: AUC (11, 22), DSS (24), and GI₅₀ (3). AUC is based on percent-growth signals and does not use data-fitting. DSS is an area-based method that uses fitting. GI₅₀ is the concentration of 50% cell growth inhibition and accounts for the starting cell density. For AUC and DSS, lower correlations are found (averages 0.76 and 0.75, Fig. 1C; Supplementary Table S3C). For AUC, this is consistent with literature (20), probably because an AUC value is more sensitive to variations in maximum and minimum signals in the assay. The fitting applied in DSS does not ameliorate this, maybe because DSS was implemented without manual curve curation. However, GI₅₀-based correlations were higher than their IC₅₀ equivalents (average 0.88 compared with 0.80), demonstrating that correcting for cell line growth rates improves reproducibility (32).

Next, we investigated if other factors also contribute to IC₅₀ reproducibility. Plotting the standard deviations of replicate log IC₅₀s shows that they depend on cell line growth rate, and on time between experimental replicates, but not on the standard deviation of cell growth rate in each cell line (Supplementary Fig. S1B–S1D). As this suggested a non-biological source of error, we studied the influence of data interpretation, by repeating curve fitting of all replicate profiles, as data interpretation rules tended to shift over the years of the experiments (Fig. 1C; see Supplementary Table S3B for data and Supplementary Fig. S1E–S1G for final rule set and some illustrative case studies). This resulted in significant improvements. After refitting, only 2 out of 15 replicates have a correlation below 0.8, and none have a correlation below 0.7. The average is 0.88, similar to the GI₅₀-based correlations. This shows that it is feasible to have “near-perfect” (17, 20) correlations between profiles in a cell line panel, and that to increase reproducibility, more effort should be put into uniform and unambiguous computational determination of response parameters (refs. 24, 32; Fig. 1C).

As a next step, we compared the IC₅₀s of our cell panel profiling to those of other studies. The overlap of our 122 compounds with other cell line profiling studies is given in Supplementary Table S1A. Analysis of ¹⁰logIC₅₀s measured on identical compounds in identical cell lines shows the best Pearson correlations with data from NCI-60 (ρ = 0.82) and JFCR-39 (ρ = 0.85), followed by data from the GDSC (ρ = 0.73), CCLE (ρ = 0.75), and GSK studies (ρ = 0.78; Fig. 2A and B; Supplementary Fig. S1H). Data from CTRP, which consist only of AUC data, were not analyzed because comparison of AUC metrics requires use of identical dose ranges, which was not the case. Given the fact that all platforms differ in read-out technology, incubation time, cell line passage, cell densities and plate formats, this shows that cell line profiling can be reproducible across platforms (20, 21).

As five cell lines (A549, ACHN, BT-549, NCI-H460, and OVCAR-3) were profiled in all these studies, we next compared the potency across platforms for eight highly active compounds that were profiled in multiple panels (Fig. 2C). For the Oncolines data, mitomycin-C and methotrexate deviate from values measured in at least three other panels (Fig. 2C). Data refitting improved concordance for mitomycin-C, but not for methotrexate (Fig. 2C). Therefore, methotrexate was reprofiled (Supplementary Table S3B). The correlation of the new data with the earlier replicate is high (ρ = 0.87). Thus, it appears that we measure a significantly more potent activity (factor of 15) of methotrexate than that presented in the literature (Fig. 2C). One reason might be that our solutions were always freshly made and that methotrexate can easily degrade when light-exposed (33, 34).

Next, we analyzed the data of 122 anticancer agents targeting all important oncogenic signaling pathways. A total of 38 of these compounds were not part of any profiling study before (Table 1 and Supplementary Table S1A). The ¹⁰log IC₅₀ values were submitted to unsupervised hierarchical clustering (Fig. 3). When duplicate profiles are included, nearly all replicates are neighbors, showing that the data reproducibility is of such quality that we can pick up “identical” compounds (Supplementary Fig. S2A). From the tree without replicates (Fig. 3A), we wanted to isolate the minimum amount of relevant clusters. Therefore, we applied a cutoff at the correlation level that is minimally significant (ρ = 0.29; P < 0.05; see Methods and Supplementary Table S3D).

The cutoff leads to a total of 44 clusters, of which 35 contain at least two compounds, and maximally eight (Table 2). In total, 26 clusters contain two or more compounds with known and similar biochemical targeting, such as EGFR, topoisomerase, or MEK inhibitors (Fig. 3; Table 2). That compounds are grouped according to mechanism, not potency, is illustrated by the MEK inhibitors selumetinib, PD-0325901, and trametinib that have an average IC₅₀ on the full panel of, respectively, 5.3 μmol/L, 0.9 μmol/L, and 0.49 μmol/L, but which still belong to the same cluster (Fig. 3; Table 2). We label the 26 mechanistically defined clusters as ‘highly validated clusters’. This approach to validation conceptually resembles the use of biochemical target profiles to pinpoint highly validated clusters (11). Our highly validated clusters include for instance a joint group of paclitaxel and docetaxel (Table 2), in contrast with GDSC and CTRP data, and in accordance with NCI-60 and JFCR-39 data (4, 6, 9 11).

Because clusters might depend on the clustering method used, we next evaluated affinity propagation clustering (APC; refs. 9, 27). This results in 46 clusters, of which 42 are highly similar to the hierarchical clustering (Table 2; Supplementary Table S4A). Also, with this method docetaxel and paclitaxel cluster together. However, cisplatin and carboplatin do not, which is in contrast to the hierarchical clustering and the literature (Fig. 3; refs. 4, 6).

To further compare the robustness of clustering across platforms, we calculated hierarchical clustering trees and correlation matrices for the 51 compounds that our data set has in common with the NCI-60 panel. Despite the differences in the platforms, clear parallels can be seen between groupings of compounds (Fig. 2D). In the 51-compound NCI-60 data, 11 “highly validated” clusters appear, i.e., that contain at least two compounds with similar biochemical mechanism. These are all represented in the clustering tree of Fig. 3 (see also Supplementary Table S4B). These clusters comprise EGFR, BRAF/MEK, mTOR, and multikinase inhibitors, as well as topoisomerase inhibitors, platins, taxanes, purine analogues and histone deacetylase (HDAC) inhibitors (Table 2). The fact that these clusters can be found in two different platforms with different read-out technology and comprising different cell lines provides good validation that they are truly distinct mechanistic approaches of cancer therapy.

The hierarchical clustering groups compounds with similar biochemical targets, but also reveals some surprises. For instance ibrutinib, an FDA-registered irreversible BTK inhibitor, clusters with the irreversibe EGFR inhibitor afatinib, and other reversible EGFR and HER2 inhibitors (Fig. 3). To further investigate this, we profiled ibrutinib on a panel of biochemical kinase assays (ref. 35; Supplementary Table S3E), which showed that aside from BTK, ibrutinib is also a potent EGFR and HER2 inhibitor, as described before (36). Because of these activities, ibrutinib behaves as an EGFR inhibitor in the cell panel.

The registered ALK/MET kinase inhibitor crizotinib clusters with danusertib, a pan-Aurora inhibitor (Fig. 3). This may be related to the high Aurora A activity of crizotinib (37) and the tyrosine kinase activity of danusertib (38), making a cluster with dual Aurora/Tyrosine kinase inhibitors.

Another paradoxical cluster is that of the ALK inhibitor ceritinib and the IGF1R inhibitors GSK-1838705A and NVP-ADW-742. Ceritinib could be expected to cluster with other ALK inhibitors such as crizotinib. However, recently it was stressed that ceritinib also has IGF1R activity (39). Therefore, this cluster seems to be based on common IGF1R inhibition. In the CTRP panel, the ALK inhibitor NVP-TAE684 was also observed to cluster with IGF1R inhibitors (11). As combined ALK/IGF1R inhibition showed synergistic effects, it was suggested that cluster proximity can predict synergy (11).

Some compounds with common biochemical mechanism are dispersed over various clusters. For instance, all alkylating agents tested, such as, e.g., dacarbazine, melphalan, and temozolomide cluster in different groups. Another group that is spread across clusters are the CDK inhibitors (PHA-793887, dinaciclib, and palbociclib).

In the hierarchical clustering tree, many clusterings are seen of compounds that were profiled for the first time. Examples are the clusters of the two TTK inhibitors NMS-P715 and MPI-0479605, and the cluster of the two CHK1 inhibitors CHIR-124 and SCH-900776 (Fig. 3). This confirms that inhibition of these cell-cycle targets generates distinct cellular profiles (29).

A substantial set of epigenetic modulators was tested, including HDAC, BET, DOT1L, and EZH2 inhibitors. The HDAC inhibitors (entinostat, vorinostat, and panobinostat) form one cluster with the PARP inhibitor olaparib. This is based on their inhibition of among others hematopoietic cell lines, consistent with HDAC inhibitors being most frequently approved for use in hematological cancers (40). Furthermore, as HDAC and PARP inhibitors work synergistically in many settings (41), their proximity supports that clustering can be used to identify synergistic pairs (11).

The BET inhibitors JQ1 and I-BET-762 (also known as BRD4 inhibitors, which is one of the BET-family members) form a separate cluster close to the BCL-2 inhibitors (ABT-737 and navitoclax). All are very potent inhibitors of leukemia and lymphoma lines such as MOLT-4, Jurkat E6-1, CCRF-CEM, and SR, which is in line with the clinical trials for BET and BCL-2 inhibitors, that focus on leukemias and lymphomas (42).

Of the class of EZH2 inhibitors, we profiled five representatives, which form two clusters, the first consisting of GSK-343, EPZ-005687, and UNC-1999. The second consisting of GSK-126 and EPZ-6438 (Fig. 3). Both groups are active in many leukemic cell lines; the second EZH2 cluster also strongly inhibits the osteosarcoma line MG-63. It must be noted that the first group shares a common indazole scaffold, which is not seen in compounds from the second group (Supplementary Table S2), so the clustering might be related to a scaffold-specific off-target activity.

To further verify the EZH2 inhibitor subgroups, we constructed a network tree (Fig. 4A). This tree is an independent analysis from hierarchical clustering because every inhibitor can have more than two neighbours. For clarity, the network contains only inhibitors of which the profiles highly correlate with the EZH2 inhibitors, and correlates thereof (ρ ≥ 0.5, Fig. 4A). This tree confirms that there are two kinds of EZH2 inhibitors.

A total of six Aurora-kinase inhibitors were profiled (tozasertib, AMG-900, GSK-1070916, danusertib, MK-5108, and MLN-8054), which all have been tested in patients (Table 1). The first three form a separate cluster that borders CDK and TTK inhibitors. Danusertib clusters with crizotinib (see above). The other two, MK-5108 and MLN-8054, border BET and BCL2 inhibitors. Follow-up analysis with a network tree shows two groups, one consisting of MK-5108 and MLN-8054, and one of the remaining four inhibitors (Fig. 4B).

To study if subgroups are related to isoform selectivity, we measured Aurora kinase A, B, and C selectivity in enzyme activity assays and binding assays based on surface plasmon resonance (Fig. 4D). Consistent with literature, MK-5108 and MLN-8054 are Aurora A selective inhibitors, whereas tozasertib, AMG-900, GSK-1070916, and danusertib are pan-Aurora inhibitors (Fig. 4D). Thus the two biochemical groups overlap with the groups in the network tree (Fig. 4B) and correlate with Aurora A and pan-Aurora selectivity, respectively.

Studying cell line profiles, it appears that the four inhibitors in the pan-Aurora cluster (Fig. 4D) are more active than the Aurora A inhibitors in diverse lines such as SK-N-AS, LS174T, BxPC-3 and AN3 CA, but also acute lymphoblastic leukemia (ALL) cell lines such as MOLT-4 and CCRF-CEM. This matches the observation that inhibitors of Aurora B (such as pan-Aurora inhibitors) are more effective than selective Aurora A inhibitors in ALL patients (43).

The PI3K/AKT/mTOR pathway contains many important drug targets. Particularly, mTOR inhibitors and the PI3Kδ inhibitor idelalisib have been approved for use as anticancer drugs, while many other compounds are in clinical trials (44, 45). We profiled a total of 12 inhibitors, of which 3 were not characterized in a larger panel before (Supplementary Table S1A). The hierarchical clustering assigns all inhibitors to the upper part of the tree, as expected for compounds active in the same pathway (Fig. 3). The only exception is the allosteric AKT inhibitor MK-2206, which profile seems most related to the platins (Fig. 3).

With regard to mTOR inhibitors, the hierarchical clustering clearly distinguishes the allosteric agents everolimus and temsirolimus (also known as rapamycin analogues or rapalogs) from two ATP-competitive inhibitors of mTOR (dactolisib and AZD-8055).

PI3K inhibitors are separated into two clusters, the first containing alpelisib, buparlisib, pictilisib, and apitolisib, the second containing duvelisib, idelalisib, and TGX-221. These groups are confirmed in a network analysis (Fig. 4C). Interestingly, buparlisib, the most advanced pan-PI3K inhibitor in the clinic (Table 1; ref. 44) shows correlation with a substantial number of profiles of non-PI3K inhibitors (Fig. 4C).

Because the catalytic subunit of PI3K exists in several isoforms, we compared the affinity for PI3Kα to δ by surface plasmon resonance. This shows that alpelisib, buparlisib, pictilisib, and apitolisib all have pan-PI3K activity and potently inhibit PI3Kα (Fig. 4E). In contrast, TGX-221, duvelisib, and idelalisib all have relatively high inhibitory activity on PI3Kβ, γ, or δ, and low inhibitory activity on PI3Kα (Fig. 4E). Thus, the two PI3K clusters in Fig. 4C correlate with PI3K isoform selectivity.

To study the biological relevance of the PI3K clusters, we characterized our cell lines for mutations in PTEN, PIK3CA, and PIK3R1, which are three of the most frequently mutated genes in cancer, and components of the PI3K pathway (46). Then we studied which of the inhibitors in our library most selectively inhibited mutant cell lines. For PTEN mutants, these are the rapalogs, followed by duvelisib (Fig. 4F). For PIK3CA-mutant cell lines, these are MK-2206, temsirolimus, and the PI3Kα-selective inhibitor alpelisib (Fig. 4G). For PIK3R1-mutant cell lines, this is the PI3Kγ/δ-selective inhibitor duvelisib (Fig. 4H). Thus different inhibitors of the PI3K pathway selectively inhibit subsets of cell lines dependent on PI3K signalling. Notably, only the association between rapalog sensitivity and PTEN-mutation achieves high significance (Fig. 4F), and indeed this is the only biomarker that has achieved success in the clinic so far (44).

Next, we investigated the mechanistic commonalities that connect the hierarchical clusters. Because the correlations at the root of the hierarchical clustering tree (Fig. 3) are not significant, we studied the variation in the cellular panel profiles by principal component analysis, which summarizes the variation in IC₅₀ response profiles in a few essential vectors. Principal component 1 (PC1) captures mostly differences in potency (Supplementary Fig. S2B). In contrast, PC2 distinguishes between inhibitors of PI3K/AKT/mTOR signalling and RAS/RAF/MEK signalling (Fig. 4I). PC3 distinguishes between EGFR inhibitors and Aurora inhibitors. When the compound contributions to PC2 and PC3 are plotted, it is clear that the four quadrants are each occupied by PI3K/mTOR, EGFR, RAF/MEK, Aurora kinase, and Bcl-2 inhibitors (Fig. 4I). This is consistent with the four major branches in the clustering tree (Fig. 3) and indicates that these four are the major therapeutic vectors in our data.

Here, we present the profiling of 122 compounds, of which many are drugs or drug candidates that were never profiled before (Table 1). Because of the controversies surrounding reproducibility of cell panel profiling, we carefully characterized variation in our assays. The metrics generated can serve as a useful benchmark for future studies (Fig. 2; Supplementary Fig. S1). We showed that, using optimized experimental and data interpretation protocols, an average correlation of 0.8 between internal control profiles is attainable (Fig. 1C), leading to data correlations across platforms of >0.7 (Fig. 2A and B), which is a clear improvement over the literature (17, 20) and which is sufficient to produce useful clusterings.

Correction for cell growth rate significantly improves correlations. In addition to that, data interpretation appears a great source of error, something which is supported by the finding that our data correlate best with NCI-60 and JFCR-39 panels, which also comprise carefully manually fitted data despite having technologically the greatest differences with our panel.

Within 122 Oncolines profiles, we identified 44 clusters with correlations >0.3. Although 44 might seem a high number of clusters, another profiling study, also using a correlation cutoff, found 49 profiles from 130 agents tested (7).

Of our 44 clusters, 26 are “highly validated” i.e., contain at least two compounds with similar biochemical mechanisms. This again shows that the molecular target class of inhibitors can predict the in vitro response profile (11, 13). In total, 17 of the highly validated clusters have been observed in earlier profilings (refs. 4, 6, 9, 11, 13; Table 2), demonstrating that cluster identities are independent of cell panel composition and represent a relevant and systematic classification of therapeutic modalities. The other 9 highly validated clusters were never described before, including TTK, multikinase ABL, WEE1/CHK1, and the subclasses of Aurora, PI3K, and EZH2 inhibitors (Table 2). This suggests that many of the novel antiproliferative approaches not only address new biochemical targets but also lead to distinctly new cell line response profiles. It is expected that if additional therapies are profiled, new clusters will appear. To facilitate their identification, we generated an overview of the 72 validated clusters identified so far in the literature (Supplementary Table S4C). It is clear that cancer cell line profiling can be used to test how distinct a new therapy is, and to get an overview of the diversity within cancer therapy.

The finding that biochemical targeting is predictive of cluster identity is supported by our analysis of the cellular profiles and biochemical selectivities of panels of Aurora and PI3K inhibitors, which shows that subclusters can relate to isoform selectivity of these inhibitor classes. Thus, there are cell line populations that are more sensitive to Aurora A or pan-Aurora inhibitors. Others are more sensitive to pan-PI3K, or PI3Kβ-, γ-, or δ- selective inhibitors. It is possible that more classes of PI3K inhibitors will appear if the same inhibitor set is tested on a larger panel, incorporating cell lines representing chronic lymphocytic leukemia, which is one of the target diseases for PI3Kγ/δ inhibitors. It must be noted that isoform selectivity is not relevant for all compounds, as for instance all non-covalent EGFR inhibitors cluster together (Fig. 3), whereas some of them also have additional ERBB2 activity (35–37) and our panel comprises a.o. the ERBB2-overexpressing line AU-565 (10).

Principal component analysis shows that our data set differentiates four major different therapy areas: PI3K/mTOR (related to metabolism), EGFR and MEK (related to growth factor signalling), Aurora (related to cell-cycle signalling), and Bcl2 (related to apoptosis). Intriguingly, this coincides with the antiproliferative subsets of hallmarks of cancer, as defined by Hanahan and Weinberg (Fig. 4I; ref. 47). Thus, these theoretical signalling classes empirically emerge from our unbiased screen, which supports the classification of cancer therapy with the hallmarks concept.

In conclusion, we have addressed controversies concerning the reproducibility of cancer cell line profiling and shown that data can contain meaningful information even if panels run over several years. The most direct proof is that compound profiles cluster according to biochemical mechanism. Clustering not only reveals side activities for known compounds, such as the EGFR activity of ibrutinib, but also shows that isoform-selective compounds are therapeutically distinguishable subclasses, as for PI3K and Aurora inhibitors. Moreover, cell panel profiling can identify compounds with synergistic effects, by combining compounds with similar cellular selectivities (48), or by combining cluster neighbours (11). Thorough cell line profiling is therefore a powerful in vitro tool for categorizing and understanding novel cancer therapeutics.

No potential conflicts of interest were disclosed.

Conception and design: J.C.M. Uitdehaag, G.J.R. Zaman

Development of methodology: J.C.M. Uitdehaag

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): J.A.D.M. de Roos, M.B.W. Prinsen, N. Willemsen-Seegers, J.R.F. de Vetter, A.M. van Doornmalen, M. Sawa

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): J.C.M. Uitdehaag, J.A.D.M. de Roos, M.B.W. Prinsen, N. Willemsen-Seegers, J.R.F. de Vetter, J. Dylus, A.M. van Doornmalen, J. Kooijman, G.J.R. Zaman

Writing, review, and/or revision of the manuscript: J.C.M. Uitdehaag, J.A.D.M. de Roos, J. Dylus, S.J.C. van Gerwen, J. de Man, R.C. Buijsman, G.J.R. Zaman

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): J. Dylus, A.M. van Doornmalen, S.J.C. van Gerwen, J. de Man

Study supervision: G.J.R. Zaman

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.