Identification of Predictive ERBB Mutations by Leveraging Publicly Available Cell Line Databases

Although cancer tissues harbor thousands of mutations (1, 2), only few have clinical value as predictive biomarkers guiding therapeutic decision making (3). This imbalance may be explained by several factors. A simple explanation is that the majority of all somatic mutations are nonfunctional passengers resulting from the general genetic instability of the cancer genomes (4). On the other hand, statistical data about the so-called long tail of the frequency distribution of somatically mutated genes across cancer samples indicate that 85% of all hotspots are found in less than 5% of tumors of all cancer types (5, 6). These latter observations imply that a significant number of potentially clinically relevant predictive mutations still remain to be identified. This conclusion also involves that a major obstacle for identifying new predictive mutations has been their rare occurrence and relative lack of structurally defined hotspots altering specific amino acid residues. In addition, technical obstacles have limited the use of functional in vitro analyses to test thousands of mutations in relevant cellular and molecular contexts (7).

The ERBB family of RTKs contains four homologous receptors: EGFR/ERBB1, ERBB2/HER2, ERBB3/HER3, and ERBB4/HER4. Upon ligand binding, the ERBB receptors form either homo- or heterodimers, which results in autophosphorylation and activation of downstream signaling (8–10). In cancer tissues, the activation may become enhanced or ligand-independent by somatic oncogenic mechanisms including single missense mutations, small deletions and insertions, truncation of longer segments of the receptor ectodomain, or amplification (11–13). These events may induce oncogene addiction and sensitivity of the cancer cells to compounds blocking ERBB signaling.

A number of ERBB targeting compounds have been approved for clinical use. The best characterized examples include the anti-ERBB2 antibody trastuzumab that is used to treat early and metastatic breast cancer with ERBB2 amplification, and the EGFR-selective TKIs erlotinib and gefitinib that have been approved for patients with non–small cell lung cancer (NSCLC) harboring the missense mutations or small deletions in the kinase domain of EGFR (14). More recently, the repertoire of ERBB TKIs has been expanded by osimertinib that is designed to also block signaling by cells harboring an EGFR T790M mutation conferring resistance to the first-generation EGFR TKIs (15), as well as the pan-ERBB TKIs afatinib, neratinib and dacomitinib, that block the signaling of all kinase-competent ERBB receptors (16–18). Activating somatic mutations have also been reported in ERBB2, ERBB3, and ERBB4 in different cancer types (3, 19–24). However, the predictive significance of these remains to be fully elucidated.

Several publicly available databases include sequencing and drug sensitivity data from thousands of cancer tissue samples and cell lines (25–27). These data allow systematic analysis of the relationship between genetic alterations and drug sensitivity in relevant cellular contexts. We hypothesized that mutations that confer sensitivity to a targeted drug, such as an ERBB-targeted TKI, can be identified as outliers in the drug sensitivity data by comparing cells lines harboring wild-type and mutant variants of the specific target gene for the drug.

By combining the drug sensitivity data from Cancer Cell Line Encyclopedia (CCLE; ref. 25), Genomics of Drug Sensitivity in Cancer (GDSC; ref. 28), and Cancer Therapeutics Response Portal (CTRP; ref. 29), we were able to assess the predictive potential of 412 ERBB mutations in 296 cancer cell lines to 10 different ERBB-targeting TKI compounds. By normalizing the data with the known sensitivity of cells harboring the classical predictive EGFR mutations to EGFR-selective TKIs, putative novel predictive ERBB mutations were identified. These findings demonstrate the potential of the approach in identifying predictive biomarkers for cancer therapy with oncogene-targeted drugs.

Human cancer cell lines with available mutation data were searched from COSMIC (Catalogue of Somatic Mutations in Cancer; https://cancer.sanger.ac.uk/cosmic; v71) and CCLE (https://portals.broadinstitute.org/ccle/home; 24-Feb-2015 data) databases. CNV data were searched from cBioPortal (https://www.cbioportal.org/; based on CCLE/Broad 2019 information). For cell line nomenclature, names used by CTRP (https://ocg.cancer.gov/programs/ctd2/data-portal; v2) were primarily used, and when not available from CTRP, adopted from GDSC (http://www.cancerrxgene.org; v17.3), as indicated in Supplementary Table S1.

ERBB TKI drug sensitivity data for cell lines with known mutation status for any of the four ERBB genes were downloaded from the websites of CCLE, GDSC, and CTRP. Data were processed with RStudio (R; RRID:SCR_000432; ref. 30) using code available at Code Ocean (https://codeocean.com/capsule/7916720). To normalize AUC data, relative AUC (rAUC) values were established for each cancer cell line/ERBB TKI/database combination. To this end positive and negative reference values were generated for each TKI and database as described in Results. To normalize EC₅₀ data, modeled EC₅₀ (mEC50) values were determined after log-logistic modeling (code available at Code Ocean; https://codeocean.com/capsule/7916720) with Akaike Information Criterion. Potential predictive mutations were identified from the dataset by selecting cell lines with a rAUC value of ≥1.0 and a mEC50 value of ≤1.0 μmol/L for a given drug. All the processed data are included in Supplementary Table S2.

Ba/F3 (DSMZ; RRID CVCL_0161) and Phoenix Ampho cells (a gift from Dr. Garry Nolan; RRID CVCL_H716) were cultured in RPMI1640 (Lonza) supplemented with 10% FBS (Biowest), ultraglutamine 1 (Lonza), 50 U/mL penicillin, and 50 U/mL streptomycin (Gibco). For maintenance of the Ba/F3 cells, the culture medium was further supplemented with IL3 in the form of 5% conditioned WEHI cell medium. COS-7 (RRID CVCL_0224) and NIH-3T3 (ATCC; RRID CVCL_0594) cells were cultured in DMEM (Lonza) supplemented with 10% FCS, 50 U/mL penicillin, and 50 U/mL streptomycin. Cells were routinely tested for Mycoplasma infection using MycoAlert (Lonza).

pBABE-puro-gateway empty vector was a gift from Matthew Meyerson (Addgene plasmid No. 51070; http://n2t.net/addgene:51070; RRID:Addgene_51070). pBABE-puro-gateway-EGFR has been described previously (7). pBABE-puro-gateway-ERBB2 was a gift from Matthew Meyerson (Addgene plasmid No. 40978; http://n2t.net/addgene:40978; RRID:Addgene_40978). pBABE-puro-gateway-ERBB3, and pBABE-puro-gateway-ERBB4JM-aCYT-2 plasmids were constructed by Gateway cloning (Addgene; ref. 7) using pcDNA3.1ERBB3-HA (31) and pcDNA3.1ERBB4JM-aCYT-2 (32) as the sources for inserts. Point mutations to ERBB inserts were introduced by site-directed mutagenesis using oligonucleotide primers listed in Supplementary Table S3. All constructs were verified by sequencing.

pBABE-based ERBB expression vectors were transfected into Phoenix Ampho cells using Fugene6 transfection reagent (Promega) for the production of infective retroviruses. Ba/F3 and NIH-3T3 cells were infected with the retroviral supernatants, as described previously (7). Cells with integrated retroviral inserts were selected with 2 μg/mL puromycin (Gibco) for 48 hours. Selection pressure was subsequently maintained by the presence of 1 μg/mL puromycin in the culture medium. To generate Ba/F3 cells with simultaneous expression of both ERBB2 and ERBB3, stable puromycin-selected cells expressing ERBB3 were transduced with supernatants from Pheonix Ampho cells expressing pBABE-puro-gateway-ERBB2.

COS-7 cells plated on 6-well plates were transfected with the pBABE-puro-gateway-EGFR, pBABE-puro-gateway-ERBB2 and/or pBABE-puro-gateway-ERBB3 constructs or the empty pBABE-puro-gateway vector using Fugene6. A total of 2 μg of plasmid DNA diluted in Opti-MEM (Gibco) was used per transfection. Cells were used for Western blot analyses 48 hours after transfection.

Ba/F3 cells washed twice with PBS were seeded at 100,000 cells/mL in RPMI1640 + 10% FCS supplemented with or without IL3 (5% WEHI cell-conditioned medium), or 10 ng/mL EGF (R&D Systems) or 20 ng/mL NRG-1 (R&D Systems). Quadruplicate 100-μL aliquots (corresponding to 10,000 cells at the initial seeding density) of the cultures were collected daily into 96-well plate wells for analysis of cell viability with the MTT assay (CellTiter 96 nonradioactive cell proliferation assay; Promega).

Quadruplicate aliquots of Ba/F3 cells (cultured in the presence or absence of IL3 and/or ERBB ligand as indicated) were seeded at 5,000 cells/96-well plate wells in a volume of 75 μL/well and incubated overnight. The following day ERBB TKIs erlotinib, lapatinib, afatinib, or neratinib (all from Santa Cruz Biotechnology) were applied in a volume of 75 μL/well to reach a final concentration ranging from 0 to 10 μmol/L. Cells were incubated in the presence of the drugs for another 72 hours, after which the number of viable cells was analyzed with the MTT-assays. Ba/F3 cells infected with an empty vector cells growing in the presence of IL3 served as a negative control.

For short-term stimulation of the cells with ERBB ligands, NIH-3T3 and COS-7 cells were starved overnight in DMEM + 0% FCS, and stimulated for 10 minutes with 0 or 50 ng/mL of EGF or NRG-1, prior to suspension in the lysis buffer. Cell were lysed and analyzed by Western blotting, as described previously (33), using the following primary antibodies: anti-EGFR (Santa Cruz Biotechnology, Cat No. sc-03, discontinued), anti-ERBB2 (Thermo Fisher Scientific, Cat No. MA-5–14057, RRID:AB_10977723), anti-ERBB3 (Cell Signaling Technology, Cat No. 4754, RRID:AB_10691324), anti-ERBB4 (E200, Abcam, Cat No. ab32375, RRID:AB_731579), anti-phospho-EGFR Tyr1086 (Cell Signaling Technology, Cat No. 2220, RRID:AB_823485), anti-phospho-ERBB2 Tyr1221/1222 (Cell Signaling Technology, Cat No. 2243, RRID:AB_490899), anti-phospho-ERBB3 Tyr1289 (Cell Signaling Technology, Cat No. 4791, RRID:AB_2099709), anti-phospho-ERBB4 Tyr1284 (Cell Signaling Technology, Cat No. 4757, RRID:AB_2099987), anti-AKT (Cell Signaling Technology, Cat No. 2920, RRID:AB_1147620), anti-phospho-AKT (Cell Signaling Technology, Cat No. 4060, RRID:AB_2315049), anti-ERK (Cell Signaling Technology, Cat No. 9102, RRID:AB_330744), and anti-phospho-ERK (Cell Signaling Technology, Cat No. 9101, RRID:AB_331646). Loading was controlled using β-actin (Sigma Aldrich, Cat No. A5441, RRID:AB_476744) or β-tubulin (Sigma-Aldrich, Cat No. T7816, RRID:AB_261770) antibodies.

NIH-3T3 cells were starved overnight in serum-free medium and stimulated with 0 or 50 ng/mL EGF for 10 minutes. The cells were lysed in a volume of 400 μL. Lysates were precleared by an incubation with 40 μL of Protein G Sepharose 4 Fast Flow beads (GE Healthcare) for 1 hour at 4°C. Precleared lysates containing 400 to 600 μg of total protein were incubated with 1/100 dilution of anti-CBL (Cell Signaling Technology, Cat No. 2747, RRID: AB_2275284) overnight at 4°C. Forty microliters of beads were applied to the solution and incubation was continued for another 1 hour at 4°C. Beads were then washed four times with lysis buffer and bound antibodies and coprecipitating material eluted by adding SDS-PAGE loading buffer and heating at 100°C for 5 minutes. EGFR coprecipitating with anti-CBL was analyzed with Western blotting.

The 2.25 Å resolution X-ray structure of the ERBB2 asymmetric homodimer was obtained from the Protein Data Bank (34) (PDB code 3PP0; ref. 35). To construct the ERBB2-EGFR (activator-receiver, from here onwards) and ERBB2-ERBB4 heterodimers, the homodimers were generated for the 2.8 Å resolution X-ray structure of EGFR (PDB code 2GS2; ref. 36) and 2.5 Å resolution structure of ERBB4 (PDB code 3BCE; ref. 37) using symmetry operations and the program Chimera (38). Next, the ERBB2 activator kinase domain was superimposed with the activator kinase domain of the EGFR homodimer and the ERBB4 homodimer using Chimera (38), followed by substitution with the ERBB2 activator kinase to produce coordinates for models of the ERBB2-EGFR and ERBB2-ERBB4 heterodimers. The ERBB2 E936K mutant forms of ERBB2-ERBB2, ERBB2-EGFR, and ERBB2-ERBB4 dimers were modeled by mutating E936 to K936 in Chimera (38).

The wild-type and E936K ERBB2-EGFR structures were prepared for molecular dynamics simulations (MDS) using the protein preparation wizard in Maestro (39), in which hydrogen atoms were added, protonation states for ionizable groups were determined and the proteins were energy minimized.

The resulting wild-type and mutant ERBB2-EGFR structures were used as starting structures for all-atom MDS using the Amber program (version 18; ref. 40) and ff14SB protein force field (41). Initially, proteins were solvated with TIP3P water molecules (42) in an octahedral box, with 10 Å between solute atoms and the box surface. The systems were neutralized by adding Na⁺ ions, with additional Na⁺/Cl⁻ ions adjusting the salt concentration to 0.15 mol/L.

The wild-type and mutant systems were first minimized, heated, and equilibrated (see ref. 43); production simulation was then performed in an isothermal (300 K)–isobaric (1 bar) ensemble for 350 ns. The resulting trajectories were analyzed using Cpptraj (44) and VMD (45). Hydrogen bonds were defined as a bond distance ≤ 3.5 Å and a bond angle ≥ 135°. Root-mean-square deviations (RMSD, backbone atoms) were computed with respect to the initial frame structure. Pairwise interaction energies between residues and the free energy of binding between the ERBB2 and EGFR kinase monomers in the wild-type and E936K heterodimers were carried out using the MMPBSA.py module (46) available in Amber.

Data were collected from the three databases: CCLE, GDSC, and CTRP. Collectively, these databases include information about 1,460 different human cancer cell lines (Fig. 1A; Supplementary Table S1) with sequence information available from 1,356 unique cell lines, and drug sensitivity data for 24 (CCLE), 265 (GDSC), and 481 (CTRP) different pharmacologic compounds.

Altogether 296 unique cell lines harboring one or more ERBB alteration were identified, and 997 cell lines were determined as ERBB wild-type (Fig. 1A). The total number of unique coding sequence alterations in the ERBB genes was 412 (348 missense, 25 nonsense, 32 frameshift, and 7 deletion, insertion or splice variant alterations), with 90 of the 296 ERBB-mutant cell lines (30.4%) containing more than one ERBB alteration. In addition to ERBB coding sequence mutations, CCLE data were used to annotate cell lines with copy number alterations for any of the four ERBB genes. The coding sequence variations for each ERBB mutant cell lines, as well as copy number information for each ERBB wild-type and mutant cell line, are listed in Supplementary Table S4.

The three databases covered data for 10 different ERBB-selective TKIs (Fig. 1A). The 10 TKIs included the EGFR selective TKIs erlotinib, gefitinib, PD153035 and pelitinib, the EGFR selective TKIs WZ4002 and WZ8040 with activity also against the resistance mutation T790M, the dual EGFR/ERBB2-selective TKI lapatinib, and the pan-ERBB inhibitors afatinib, canertinib, and neratinib.

AUC was selected as a measure of drug sensitivity as it was available in all three databases, and as it serves as a comprehensive readout covering a range of drug concentrations. To normalize the drug sensitivity data from all three databases, a relative AUC (rAUC) value was determined for each cell line/drug/database combination. This was accomplished by determining a database-specific AUC reference level for positive and negative drug responses. The reference value for a positive response was determined as the mean response (as AUC) of all cells harboring the known EGFR driver mutations, L858R or exon 19 deletions, to the EGFR-selective TKIs erlotinib, gefitinib, and PD153035 in the given database. The reference value for a negative response was determined as the mean response (as AUC) of all the ERBB wild-type cell lines to all the ERBB TKIs analyzed in the given database. The rAUC value for each cell line/drug/database combination was then generated (Fig. 1B) using the formula:

Consequently, rAUC values over 1 indicated a response greater than that associated with the known EGFR driver mutations to EGFR TKIs, rAUC values between 0 and 1 indicated a response smaller than that associated with the EGFR driver mutations to EGFR TKIs but greater than that associated with ERBB wild-type cells to any ERBB TKI, and rAUC values smaller than 0 indicated a response smaller than that associated with ERBB wild-type cells to any ERBB TKI.

EC₅₀ was used as another variable for analysis of drug response to control for high rAUC values associating with micromolar EC₅₀s that confer a risk for off-target effects or may not be clinically relevant (Fig. 1B). To obtain comparable EC₅₀ values, log-logistic modeling was applied to the drug response data, and EC₅₀ values were determined from the resulting fitted dose–response curves. These modeled EC₅₀ (mEC50) values were calculated for each compound that demonstrated an absolute effect of at least 10%. The resulting normalized dataset contained 4,685 unique database, cell line, gene, mutation, and drug combinations (Supplementary Table S2).

Potential predictive mutations were identified from the dataset as those associated with a rAUC of ≥ 1.0 and a mEC50 value of ≤ 1.0 μmol/L for a given drug (Fig. 1C; data are also available in an interactive HTML format in Supplementary Fig. S1). These cutoffs resulted in identification of 26 EGFR, 14 ERBB2, 14 ERBB3, and 22 different ERBB4 mutations in a total of 43 different cell lines (Fig. 2A and B; Supplementary Fig. S2; Supplementary Table S4). Sixty-two of the 76 mutations were missense, three nonsense mutations, eight frameshift alterations, and three deletions or insertions.

To address the association of ERBB amplification with ERBB TKI sensitivity, cell lines with amplification of any of the four ERBB genes but wild-type for ERBB coding sequence mutations, were plotted against the rAUC and mEC50 values as indicators of TKI sensitivity similar to Fig. 1C (Supplementary Fig. S3; data are also available in an interactive HTML format in Supplementary Fig. S4). The analysis indicated that few cancer cell lines harboring ERBB3 or ERBB4 amplification were associated with enhanced sensitivity to the drugs. Moreover, TKIs with selectivity toward EGFR were mostly ineffective regardless of the ERBB amplification status. However, a number of cell lines with EGFR or ERBB2 amplification (n = 14 for EGFR; n = 18 for ERBB2) were sensitive to the pan-ERBB TKI afatinib. These cell lines were mostly from contexts of known sensitivity to ERBB-targeted therapies, including ERBB2-amplified breast, gastric and lung cancer, as well as head and neck and esophageal cancer with amplifications in EGFR and/or ERBB2 (Supplementary Fig. S3).

Of all 114 cell lines wild-type for ERBB mutations but still demonstrating sensitivity to at least one ERBB TKI with rAUC ≥ 1.0, 45 (39.5%) harbored copy-number alterations in at least one ERBB gene (Supplementary Fig. S5). This suggests that factors other than ERBB copy number variations may explain a significant proportion of the variability in the responsiveness of cancer cell lines to different ERBB-targeting TKIs in the absence of ERBB mutations.

Of the 76 unique mutations associated with increased sensitivity to ERBB TKIs, 14 (18.4%) were defined as “oncogenic” according to cBioPortal (searched June 4, 2020; Supplementary Table S5). Of note, these 14 mutations constituted 53.8% of the 26 “oncogenic” ERBB mutations annotated by cBioPortal, supporting validity of the screen. The remaining 12 oncogenic mutations annotated by cBioPortal did not consistently associate with sensitivity to any ERBB TKI when endogenously expressed in the cancer cell lines (Supplementary Fig. S6). The remaining 62 of the 76 (81.6%) unique ERBB mutations were regarded as potentially novel activating mutations. These included 18 EGFR, 10 ERBB2, 12 ERBB3, and 22 unique ERBB4 alterations (Fig. 2A).

Interestingly, 17 of 34 (50.0%) of the cell lines with putative novel activating ERBB mutations harbored more than one ERBB mutation, and eight of these cell lines had an ERBB mutation combined with chromosomal gain in an ERBB gene (Fig. 2B). These observations imply that the percentage of functionally relevant novel mutations could be exaggerated in a screen simply relying on in silico analysis, as passenger mutation could be present in the same cell line with an oncogenic mutation.

To address the functional contribution of individual ERBB mutations associated with high rAUC values, 11 ERBB mutations were selected for experimental validation in vitro (Fig. 2). The selected mutations included one mutation in EGFR (Y1069C), two in ERBB2 (L720P and E936K), two in ERBB3 (E928fs*16 and E952Q), and four in ERBB4 (A17V, L780P, G863E, and G936R) that were all (except ERBB4 A17V) identified in a context of multiple ERBB aberrations (Fig. 2A). Two previously established oncogenic mutations, EGFR G719S and ERBB2 S310F, served as positive controls. Sensitivity of the selected mutations and known EGFR driver mutations to individual ERBB TKIs are shown in Supplementary Fig. S7.

To analyze the activity of the receptor variants in a cellular background devoid of significant endogenous ERBB expression, the IL3-dependent Ba/F3 cells (7, 47) were transduced with retroviral vectors encoding the ERBB variants or corresponding wild-type receptors. As expected, the positive controls EGFR G719S and ERBB2 S310F promoted IL3-independent Ba/F3 cell growth in the absence of an activating ligand, unlike the respective wild-type controls (Fig. 3A and B). In addition, EGFR Y1069C and ERBB2 E936K were capable of supporting growth in the absence of both IL3 and an ERBB ligand, indicating transforming potential (Fig. 3A and B). However, none of the ERBB3 variants promoted IL3-independent growth as homodimers (Fig. 3C) or enhanced the transforming potential when overexpressed together with ERBB2 (Supplementary Fig. 8A). Similarly, none of the ERBB4 variants showed enhanced transforming potential. ERBB4 L780P and G863E were actually less potent than wild-type ERBB4 in promoting IL3-independent Ba/F3 cell survival in the presence of the NRG-1 ligand (Fig. 3D).

To address the mechanisms by which the EGFR Y1069C variant promotes growth, the phosphorylation status of EGFR variants was analyzed. Western blot analysis of Ba/F3 cells overexpressing EGFR Y1069C or G719S indicated phosphorylation of both variants at Tyr 1086 to a greater extent when compared with the wild-type EGFR, when the cells were cultured in the absence of IL3 but in the presence of the EGF ligand (Fig. 4A). Moreover, both of the variants were phosphorylated also when the Ba/F3 cells were cultured in the absence of both IL3 and of EGF, a condition under which the cells expressing wild-type EGFR did not survive (Fig. 3A). Consistently, both variants stimulated downstream EGFR signaling (phospho-AKT and phospho-ERK) to a greater extent than the wild-type receptor in the presence of EGF and were able to sustain downstream signaling also in the absence of EGF (Fig. 4A). When the EGFR Y1069C and G719S variants were overexpressed in NIH-3T3 cells, the variants again demonstrated enhanced activity although the effects were less prominent as compared with the Ba/F3 cell background (Fig. 4B).

Tyrosine 1069 (Tyr 1045 if excluding the signal sequence) of EGFR has been shown to serve as a binding site for the ubiquitin ligase c-CBL (48). The binding of c-CBL to EGFR enables the c-CBL–mediated ubiquitination and degradation of EGFR negatively controlling EGFR signaling output (49). Hence, coimmunoprecipitation assays were carried out in NIH-3T3 cells to test whether the Y1069C mutation disrupts the binding of c-CBL to EGFR. As expected, the Y1069C mutation almost completely abolished the interaction between EGFR and c-CBL (Fig. 4C), suggesting that the increased activity of the Y1069C mutant is associated with loss of c-CBL binding.

When overexpressed in Ba/F3 cells, ERBB2 E936K demonstrated enhanced tyrosine phosphorylation when compared with cells expressing wild-type ERBB2 (Fig. 5A). Enhanced phosphorylation was observed in cells maintained in the presence of the ligand NRG-1, as well as in cells emerged as clones independent of both NRG-1 and IL3 for survival and growth, similar to cells expressing the known oncogenic ERBB2 S310F (Fig. 5A). In contrast, expression of the mutant versions of other NRG-1 receptors, ERBB3 or ERBB4, did not induce receptor phosphorylation when compared with respective wild-type controls (Supplementary Fig. S8B and S9A). Consistently with the Ba/F3 transformation assays (Fig. 3D), ERBB4 L780P and ERBB4 G863E actually demonstrated reduced or no kinase activity in phospho-Western blot analyses of Ba/F3 or NIH-3T3 cells (Supplementary Fig. S9A and B). Similarly, ERBB3 E928fs with truncated kinase domain and ERBB3 E952Q, that failed to promote IL3-independent Ba/F3 growth even when expressed together with ERBB2, showed reduced or no transactivation of ERBB2 when expressed in a context of ERBB2/ERBB3 heterodimers (Supplementary Fig. S8).

Although the ERBB2 S310F variant effectively promoted pAKT and pERK signaling, associated with an induction of endogenous ERBB3 in Ba/F3 cells, these downstream signaling responses were modest in cells expressing ERBB2 E936K (Fig. 5A). The E936K variant also promoted ERBB2 autophosphorylation when overexpressed in COS-7 cells but did not demonstrate any additional activity when overexpressed together with ERBB3 in the context of ERBB2-ERBB3 heterodimers (Fig. 5B).

To understand the activation mechanism of the ERBB2 E936K mutation, we performed structural analysis and molecular dynamics simulation (MDS). Glutamate E936, located on the αG helix of the kinase domain, is conserved among ERBB kinases. In the wild-type ERBB2 homodimer structure (PDB code 3PP0; ref. 35), E936 (activator kinase) forms part of the dimer interface, facing residues from the β4-β5 loop and the C-terminal end of the juxtamembrane-B segment of the receiver kinase (Fig. 5C). These residues are conserved in EGFR and ERBB2, but variation in the β4-β5 loop is observed for ERBB3 (Glu, Gly-Ser) and ERBB4 (Glu, Ser-Pro) in comparison with EGFR and ERBB2 (Glu, Thr-Ser; Fig. 5C).

In the wild-type ERBB2-ERBB2 structure and ERBB2-EGFR model, prior to MDS, the side-chain polar atoms of E936 of the ERBB2 activator kinase can form an interdimer hydrogen bond with the backbone nitrogen atom and side-chain hydroxyl group of S792/S784 in the ERBB2/EGFR receiver kinase (Fig. 5D). In contrast, with the E936K mutant, the ϵ-amino group of K936 could form a hydrogen bond with the hydroxyl group of S792/S784 and, a potentially stronger salt bridge with E717/E709 of the ERBB2/EGFR receiver kinase (Fig. 5D), absent in the wild-type dimers. The E936K mutant would strengthen the interaction between the asymmetric dimers, likely prolonging the duration of the activated state. Similarly, ERBB2-ERBB4 heterodimer model shows that E715 and S789 from the ERBB4 receiver kinase can respectively form similar ionic and hydrogen bond interactions with E936K in the mutant activator ERBB2 kinase (Fig. 5E). The effect of the E936K mutation might be negligible in the heterodimer formed between ERBB2 and ERBB3 since the pseudokinase ERBB3 serves predominantly as an activator due to its weak kinase activity (50, 51), which is consistent with our results (Fig. 5B). In contrast, because the EGFR kinase domain strongly prefers to be the receiver kinase domain among the ERBBs (51), followed by ERBB2, the E936K mutation should have a more pronounced effect on the ERBB2-EGFR dimer.

A 350 ns simulation of the wild-type and E936K-mutant ERBB2-EGFR heterodimers recorded that both wild-type E936 (4.2%) and mutant E936K (9.2%) hydrogen bond with S784 (β4-β5 loop) of the receiver EGFR domain. The β4-β5 loop in the mutant receiver kinase appears slightly more stable and less variable (RMSD 0.26 ± 0.06 Å) compared with the wild-type complex (RMSD 0.40 ± 0.18 Å; Fig. 5F) and may reflect the more frequent interactions observed between S784 and E936K: the interaction energy for E936K-E707 in the mutant heterodimer represents a stronger interaction (ΔG = −0.4 ± 0.3 kcal/mol) relative to the wild-type (ΔG = −0.03 ± 0.15 kcal/mol). The free energy of binding also indicates a stronger interaction between the monomers of the mutant complex (ΔG of binding −77.1 ± 8.2 kcal/mol) in contrast to the wild-type complex (−66.6 ± 11.1 kcal/mol; Fig. 5G). The structural analysis suggests that the E936K mutation at the ERBB2 C-lobe would lead to stronger interactions at the dimer interface in the ERBB2-EGFR heterodimer and ERBB2-ERBB2 homodimer, resulting from increased hydrogen bonding and unique ionic interactions between the monomers.

To experimentally test whether ERBB2 E936K mutant could serve as a more potent activator kinase in ERBB2/EGFR heterodimers, ERBB2 wild-type or E936K were transfected together with wild-type EGFR to COS-7 cells and stimulated or not with EGF. Indeed, ERBB2 phosphorylation was increased in cells expressing both wild-type EGFR and ERBB2 E936K as compared with cells solely expressing wild-type receptors (Fig. 5H). Quantification of four replicate experiments showed that the expression of EGFR and ERBB2 E936K heterodimers produced two times higher phospho-ERBB2 levels than the heterodimers of wild-type receptors (P = 0.026; Fig. 5I). These findings indicate that the E936K mutant activated EGFR more effectively than the wild type.

To address the sensitivity of EGFR Y1069C and ERBB2 E936K to clinically used ERBB TKIs, Ba/F3 cells expressing the mutant or wild-type receptors were cultured for 72 hours in the presence of increasing concentrations of erlotinib, lapatinib, afatinib, or neratinib. Cells expressing EGFR G719S and ERBB2 S310F were included in the experiments as internal positive controls. As all the four mutant variants were able to promote IL3-independent Ba/F3 cell growth both in the presence or absence of an activating ligand (Fig. 3A and B), the drug sensitivity analyses were carried out both in the absence and presence of EGF or NRG-1. Cells infected with an empty vector and cultured in the presence of IL3 served as a control reflecting off-target toxicity.

Both EGFR Y1069C and ERBB2 E936K demonstrated sensitivity to ERBB TKIs (Fig. 6). EGFR Y1069C was most sensitive to the irreversible ERBB inhibitors afatinib and neratinib with IC₅₀ values at the low nanomolar range, with the cells cultured in the absence of the ligand generally sensitive to concentrations about one order of magnitude smaller as compared with cells cultured in the presence of EGF. The IC₅₀ for afatinib was significantly smaller in cells expressing EGFR mutants as compared with cells expressing wild-type EGFR both in the absence and presence of EGF, whereas the sensitivity to neratinib was significantly enhanced only in the absence of the ligand (Fig. 6A). Cells expressing ERBB2 E936K were also sensitive at low or even subnanomolar concentrations to both afatinib and neratinib. In contrast to EGFR Y1069C, cells expressing ERBB2 E936K were significantly more sensitive than cells expressing wild-type ERBB2 only for the response to afatinib in the presence of the NRG-1 ligand (Fig. 6B). The ERBB TKI sensitivity of cell expressing ERBB4 A17V or ERBB4 G936R did not significantly differ from the sensitivity of cells expressing wild-type ERBB4 (Supplementary Fig. S9C). Taken together, these observations validate the identification of previously unknown actionable mutations with the approach based on oncogene-targeted database screen.

Here, we leveraged the existing genomic and drug sensitivity data in publicly available cell line databases to identify novel predictive ERBB mutations for ERBB TKI sensitivity across cancers. We followed a hypothesis that cell lines harboring an ERBB mutation conferring sensitivity to ERBB TKIs would stand out as outliers in the drug response data among cell lines harboring passenger ERBB mutations. By combining and normalizing drug response data from CCLE, CTRP, and GDSC, we were able to identify 43 “exceptional responder” cell lines out of 296 cell lines harboring ERBB mutations, with 62 novel, potentially predictive ERBB mutations. Experimentally characterizing 11 of these mutations we discovered EGFR Y1069C and ERBB2 E936K as novel activating ERBB mutations conferring sensitivity to ERBB TKIs. In addition to the colorectal cancer cell line CCK81 (CCLE), EGFR Y1069C variant has previously been reported in a clinical sample representing cholangiocarcinoma (2). The ERBB E936K variant has only been reported from the cell line CTV-1 derived from a patient with leukemia (CCLE).

Our analysis also identified 14 known oncogenic ERBB variants, which are mostly located in recurrent hotspots. In contrast, the novel mutations associated with increased sensitivity to ERBB TKIs are clinically relatively rare and were dispersed throughout the ERBB receptors. Thus, our results strengthen the argument that the “long tail” of nonrecurrent cancer-associated mutations may also be of functional relevance (5). This conclusion is in line with our recent analysis of 7,216 EGFR mutations generated by random mutagenesis that identified transforming potential in 21 EGFR single-nucleotide variants at mostly nonrecurrent positions (7).

A significant proportion of cell lines (50.0%) demonstrating ERBB TKI sensitivity and harboring novel ERBB mutations, also had one or more cooccurring mutations or amplifications in other ERBB genes. This could potentially represent cooperation of oncogenic genomic alterations to promote cancer cell growth, or a single oncogenic alteration accompanied by benign passenger alterations. A recent pan-cancer analysis of multiple mutations provided evidence supporting the cooperation hypothesis. Saito and colleagues reported that mutations of low functional activity may cooperate with other mutations occurring in cis within the same oncogene to promote enhanced oncogenic signaling (52). These observations may also explain how individually rare mutation events, like most of the ERBB mutations analyzed here, participate in a relatively common cooperative phenomenon that drives tumor growth. Of note, ERBB receptors are among the oncogenes most frequently harboring multiple somatic mutations in clinical cancer samples (52).

Our experimental characterization included ERBB mutations in two cell lines where several ERBB mutations cooccurred (CCK81 and CTV-1; Fig. 2B). In these cell lines, only one ERBB mutation was found to be activating and transforming (EGFR Y1069C in CCK81; ERBB2 E936K in CTV-1), suggesting a mechanism where other ERBB mutations are passengers. On the other hand, two ERBB TKI-sensitive ERBB2-amplified breast cancer cell lines both harbored the same concomitant point mutation in ERBB3 (E952Q). This is an intriguing observation considering the importance of ERBB3-mediated signaling for the growth of ERBB2-amplified breast cancer cells (53). Although in our analyses in the Ba/F3 system, the ERBB3 E952Q mutation was not more potent than wild-type ERBB3 in promoting growth or ERBB2 phosphorylation when coexpressed with ERBB2 (Supplementary Fig. S8), the variant has previously been reported to moderately enhance ERBB3 signaling in CHO cells (54). This difference may be due to the presence of different heterodimerization partners or other cell context-dependent signaling characteristics that are sensitive to a mutation at a residue potentially critical for interactions of ERBB3 at the kinase interface (55).

Another interesting observation of cooccurring genetic events was the unexpectedly high (20.9%; Fig. 2B) frequency of ERBB4 deletions in TKI-sensitive ERBB-mutant cell lines. In addition, two of the four functionally studied ERBB4 variants, L780P and G863E, were virtually kinase dead, and completely failed to promote Ba/F3 cell growth (Fig. 3D and Supplementary Fig. S9). Structurally, these observations are consistent, in that L780 in wild-type ERBB4 is involved in hydrophobic interactions with residues from the αC and αE helices and F862 of the DFG motif, interactions key to stabilizing the active state conformation. The L780P mutation could compromise the stability and drive the transition from the active toward the inactive state of ERBB4. Similarly, the ERBB4 G863E mutation, located at the DFG motif critical for phosphotransfer in the kinase activation loop (56, 57), would cause a disruption of the active state αC helix conformation due to steric clashes and charge–charge repulsion. These observations, together with previous reports of both activating (22, 23) and kinase-impaired (57) ERBB4 mutations contributing to malignant growth, justify future addressing of the functional role of ERBB4 gain and loss of function in the context of multiple ERBB alterations.

Putative limitations of the study include the obvious bias of missing functional ERBB variants as they were either not included in the cell line collections analyzed, or as their effects are masked by other concomitant oncogenic events in the same cell line. However, all 26 ERBB mutations defined as “oncogenic” by cBioPortal were represented by at least one cell line. Moreover, none of the 12 mutations that were listed as “oncogenic” by cBioPortal, but were not defined activating in our analysis, did show significant sensitivity to any of the tested ERBB inhibitors in a dose-response analysis (Supplementary Fig. S6). It is of note, however, that for example, for EGFR L861Q, the small number of cell lines available precluded any conclusions about the activity of the variant. The annotation of the cell lines as amplified for any of the ERBB genes was also based on crude categorization to cells harboring ERBB amplification or not, and did not allow for quantitative analysis of the influence of the gain in gene copy numbers for the observed effect.

Taken together, our findings indicate that activating and potentially predictive ERBB mutations can be found by a systematic screen of mutations outside of the most prominent mutation hotspots. The data further suggests that the ERBB variants with the greatest functional significance frequently cooccur with other ERBB changes.

M.K.A. Koivu reports grants from Finnish Cultural Foundation, Varsinais-Suomi Regional Fund during the conduct of the study, as well as from Turku University Foundation outside the submitted work. D. Chakroborty reports grants from K. Albin Johanssons Foundation, Juhani Aho Foundation for Medical Research, and University of Turku Graduate School, and nonfinancial support from Finnish Cancer Foundation during the conduct of the study. M.Z. Tamirat reports grants from Doctoral Network in Informational and Structural Biology (Graduate School of Åbo Akademi University) and other from CSC IT Center for Science during the conduct of the study. M.S. Johnson reports grants from Sigrid Juselius Foundation, other from NordForsk Nordic POP, and other from CSC IT Center for Science during the conduct of the study. K.J. Kurppa reports grants from Sigrid Juselius Foundation, Finnish Cultural Foundation, Instrumentarium Science Foundation, and K Albin Johansson Foundation during the conduct of the study, as well as other from Puma Biotechnology outside the submitted work. K. Elenius reports grants from Academy of Finland, grants from Cancer Foundation Finland, and grants from Turku University Central Hospital during the conduct of the study, as well as other from Abomics, Novo Nordisk, Orion, Roche, and Puma Biotechnology outside the submitted work.

M.K.A. Koivu: Conceptualization, data curation, software, formal analysis, validation, investigation, visualization, methodology, writing–original draft, writing–review and editing. D. Chakroborty: Software, formal analysis, investigation, visualization, methodology, writing–review and editing. M.Z. Tamirat: Formal analysis, validation, visualization, writing–original draft. M.S. Johnson: Formal analysis, supervision, validation, visualization, writing–original draft. K.J. Kurppa: Conceptualization, investigation, visualization, writing–original draft, writing–review and editing. K. Elenius: Conceptualization, resources, supervision, funding acquisition, visualization, methodology, writing–original draft, writing–review and editing.

We thank CSC IT Center for Science (project 2002814, and computing infrastructure); Dr. Jukka Lehtonen (Biocenter Finland Bioinformatics Network) for scientific IT support. The Structural Bioinformatics Laboratory is part of the university's Drug Discovery and Diagnostics strategic area. This study has been supported by Academy of Finland (274728, 316796; to K. Elenius), the Cancer Foundation of Finland (to K. Elenius), Turku University Central Hospital (to K. Elenius), Doctoral Network in Informational and Structural Biology (Graduate School of Åbo Akademi University; to M.Z. Tamirat), Sigrid Juselius Foundation (to K.J. Kurppa and M.S. Johnson), NordForsk Nordic POP (to M.S. Johnson), K. Albin Johansson Foundation (to K.J. Kurppa and D. Chakroborty), Juhani Aho Foundation for Medical Research (to D. Chakroborty), University of Turku Graduate School (to D. Chakroborty), Instrumentarium Science Foundation (to K.J. Kurppa), Finnish Cultural Foundation (Varsinais-Suomi Regional Fund 85171988 (to M.K.A. Koivu), Central Fund 00200590 (to K.J. Kurppa), and supercomputing from CSC IT Center for Science (to M.S. Johnson; and to M.Z. Tamirat).

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.