Computational and Functional Analyses of HER2 Mutations Reveal Allosteric Activation Mechanisms and Altered Pharmacologic Effects Open Access

HER2 belongs to the ERBB family of receptor tyrosine kinases (RTK) that undergo dimerization to trigger downstream signaling pathways involved in cell growth, proliferation, differentiation, and survival (1–4). HER2, like the other ERBB family members, contains an extracellular domain (ECD), a transmembrane domain (TMD), and an intracellular region consisting of a juxtamembrane domain (JMD), a kinase domain (KD), and a C-terminal tail domain (CTD). The ECD is comprised of four subdomains (I–IV): two leucine-rich subdomains, L1 (I) and L2 (III), that typically participate in ligand binding (except for HER2), and two cysteine-rich subdomains, CR1 (II) and CR2 (IV), that contain multiple intramolecular disulfide bonds that line the ECD dimer interfaces (1, 5, 6). For ligand-binding ERBB family members, in absence of ligand, the ECD adopts an autoinhibited tethered (closed) conformation, whereas upon ligand binding, the ECD adopts an untethered (open) conformation, leading to homodimerization or heterodimerization (1, 3). Unlike other members of the ERBB family, HER2 ECD constitutively adopts an untethered, active conformation independently of ligand interactions (4). HER2 is the preferred partner to form heterodimer with other ligand-binding ERBB family members, although HER2 homodimers have been reported in HER2-overexpressing cells (1–3). The formation of ERBB dimers is mediated by the CR1 and CR2 of the ECD, as well as the rearrangement of the TMD and JMD to support the asymmetric dimerization of the KD (7).

Overexpression and amplifications of wild-type (WT) HER2 are known to play an oncogenic role in a subset (∼20%) of breast and gastric/gastroesophageal junction cancers (8). These findings led to the approval of anti-HER2 therapies for the treatment of HER2⁺ breast cancer, including mAbs (trastuzumab, pertuzumab, and margetuximab), antibody–drug conjugate (ADC; ado-trastuzumab emtansine and trastuzumab deruxtecan), and small molecules (lapatinib, neratinib, and tucatinib; refs. 9–16). Trastuzumab and ADC trastuzumab deruxtecan are currently approved for the treatment of HER2⁺ gastric or gastroesophageal junction adenocarcinoma (12, 15). More recently, oncogenic mutations in the HER2 gene have been identified in patients with breast, lung, colorectal, gastric/gastroesophageal, bladder cancers, and other cancers (8). Mutations in HER2, including single-nucleotide variants (SNV; point mutations), insertions, deletions, and fusions, are rare events and most commonly occur at allosteric sites including the ECD, the TMD, the JMD, and the KD. According to a comprehensive genomic profiling using next-generation sequencing conducted on a collection of 272,548 unique patient solid tumor samples (the Foundation Medicine, Inc. data: Apr 28, 2020—MI20200428 FMI), the most common HER2 mutations include S310F/Y (0.5%, in the ECD), the exon 20 insertion A775-G776insYVMA (0.2%, in the KD), R678Q (0.2%, in the JMD), L755S (0.2%, in the KD), and V842I (0.1%, in the KD).

As HER2 is subject to over 2,000 unique variants, we sought to define the full spectrum of oncogenic HER2 mutations expressed by solid tumors. Here we describe the screening process that led to the determination of a panel of 37 allosteric HER2 driver mutations, comprising 15 previously characterized and 22 newly identified oncogenic mutations, and further analyses aimed at elucidating the mechanism of oncogenicity associated with these mutations, as well as at understanding the pharmacology of available anti-HER2 agents in these mutants.

Ba/F3 cells were purchased from DSMZ (Braunschweig, Germany) and maintained in RPMI1640 medium (Life Technologies) supplemented with 10% FBS, 1% l-glutamine, and 10 ng/mL mouse IL3 (R&D Systems). Lack of endogenous murine erbB family member protein expression in mammalian hematopoietic cells (17) was confirmed in parental Ba/F3 cells (Supplementary Fig. S1). Ba/F3 transformed cells were maintained in the medium with 15-µg/mL blasticidin S HCl (Life Technologies). U87MG cells were purchased from ATCC and maintained in MEM (Life Technologies) supplemented with 10% FBS and 1% l-glutamine. The cell lines were periodically checked for Mycoplasma using MycoAlert Plus Mycoplasma Detection Kit (Lonza).

Human HER2 cDNA (RC212583; Origene) was subcloned into pMXs-IRES-Blasticidin (Cell Biolabs). HER2 mutants were generated by site-directed mutagenesis. Retroviral expression vector retrovirus was produced by transient transfection of HEK 293T cells with the retroviral ERBB expression vector pMXs-IRES-Blasticidin, pCMV-Gag-Pol vector, and pCMV-VSV-G-Envelope vector (Cell Biolabs). Briefly, HEK 293T cells were plated (2.5 × 10⁵ per plate) in 6-well plates (Corning) and incubated overnight. The next day, retroviral plasmids (1 µg of ERBB, 0.32 µg of pCMV-Gag-Pol, and 0.16 µg pCMV-VSV-G) were mixed in 150 µL of Opti-MEM (Life Technologies). The mixture was incubated at room temperature for 5 minutes and then added to Opti-MEM containing transfection reagent Lipofectamine (Invitrogen) and incubated for 20 minutes. Mixture was then added dropwise to HEK 293T cells. The next day, the medium was replaced with fresh culture medium, and retrovirus was harvested at 24 hours.

Ba/F3 cells (4.0 × 10⁴ cells) were infected with 20 µL of viral supernatant supplemented with 8-µg/mL polybrene by centrifuging for 30 minutes at 1,000 rpm. Cells were placed in a 37°C incubator overnight. Then, 100 µL cells and supernatant were added to 900 µL of RPMI1640 medium (Life Technologies) supplemented with 10% FBS, 1% l-glutamine, and 15 µg/mL blasticidin S HCl (Life Technologies). Cells were placed in a 37°C incubator and monitored for outgrowth over a 2-week period.

U87MG cells (1.0 × 10⁶ cells) were infected with 2 mL of viral supernatant supplemented with 8-µg/mL polybrene in T150 flasks. Cells were placed in a 37°C incubator overnight. After 24 hours, cells were given a media change and were placed in a 37°C incubator overnight. Cells were then selected in presence of 25 µg/mL blasticidin S HCl (Life Technologies) and monitored for outgrowth.

Ba/F3 cell lines were resuspended at 1.1 × 10⁵ cells/mL in RPMI containing 10% heat inactivated FBS and 1% l-glutamine and dispensed in triplicate (1.5 × 10⁴ cells/well) into 96-well plates. To determine the effect of drug on cell proliferation, cells were incubated for 3 days in the presence of vehicle control or test drug at varying concentrations. Inhibition of cell growth was determined by luminescent quantification of intracellular ATP content using CellTiter-Glo (Promega). Dose–response curves were plotted by comparing cell numbers on day 0 versus 72 hours posttreatment. The number of viable cells was determined and normalized to vehicle-treated controls. Inhibition of proliferation, relative to vehicle-treated controls was expressed as a fraction of 1 and graphed using GraphPad Prism 9 software (GraphPad Software). EC₅₀ values were determined with the same application.

Cell extracts were prepared by detergent lysis RIPA (Sigma) containing 10-mmol/L iodoacetamide (G-Biosciences), protease inhibitor (Sigma), and phosphatase inhibitors (Sigma) cocktails. The soluble protein concentration was determined by micro-BCA assay (Pierce Biotechnology). Protein immunodetection was performed by electrophoretic transfer of SDS-PAGE separated proteins to nitrocellulose, incubation with antibody, and chemiluminescent second step detection. Nitrocellulose membranes were blocked with 5% nonfat dry milk in TBS and incubated overnight with primary antibody in 5% BSA. The following primary antibody from Cell Signaling Technology were used at 1:1,000 dilution: phospho-HER2 and total HER2. β-Actin antibody, used as a control for protein loading, was purchased from Sigma. Horseradish peroxidase–conjugated secondary antibodies (Cell Signaling Technology) were incubated in nonfat dry milk at 1:5,000 dilution for 1 hour. SuperSignal chemiluminescent reagent (Pierce Biotechnology) was used according to the manufacturer's directions, and blots were imaged using the Alpha Innotech image analyzer and AlphaEaseFC software (Alpha Innotech).

The mutation-allostery-pharmacology (MAP) algorithm utilizes a well-curated set of genomic and proteomic features to assess the potential oncogenicity of SNVs, based on a prediction model trained on an internally generated validated mutation dataset from a group of known oncogenes. This prediction algorithm determines the potential impact of mutations causing deleterious effects on the structure and function of proteins, taking into account both gain-of-function and loss-of-function mutations. The reported MAP score, a binary classifier, ranges between 0 and 1, and is a proxy for probability of being oncogenic. This approach results in a relatively high number of false-positive predictions on driver mutations, as the MAP score oncogenicity threshold of 0.3 was chosen for higher sensitivity, sacrificing accuracy.

The atomistic model of the HER2 ECD was prepared on the basis of the HER2 crystal structure (PDB ID: 6J71) using Prime from the Schrödinger suite (18). The S310F mutant was built based on the WT model with a single substitution of residue Ser 310 to Phe. In an effort to enhance any structural changes associated with the HER2 residue 310, simulations of both WT and S310F models were conducted with and without a disulfide bond between Cys299 and Cys311 (referred to as oxidized and reduced states, respectively). We generated four distinct models with visual molecular dynamics (VMD; ref. 19): HER2 WT-oxidized, HER2 WT-reduced, HER2 S310F-oxidized, and HER2 S310F-reduced molecular dynamics (MD) simulations were performed with the NAMD 2.13 software package (20). The CHARMM36 parameter set was used for the protein (21, 22) and the TIP3P model was used for the water molecules (23). A cut-off distance of 10 Å was used for the van der Waals and short-range electrostatic interactions. The long-range electrostatic interactions were computed with the particle-mesh Ewald summation method (24). The SHAKE algorithm was applied to all hydrogen-containing bonds (25). With all other atoms fixed, the waters were energy minimized for 20,000 steps using the conjugate gradient algorithm and subsequently all systems were minimized for 100,000 steps with the conjugate gradient algorithm. After heating the system gently with the canonical ensemble NVT (constant number, volume, temperature) run for 200 picoseconds (ps) at 300 K, the systems were further minimized (100,000 steps), followed by equilibration first in the NVT and then in the isothermal-isobaric ensemble NPT for 200 ps each. Finally, unbiased MD simulations were carried out, with the atomic coordinates of the systems saved every 10 ps. For each of four simulations (HER2 WT-oxidized, HER2 WT-reduced, HER2 S310F-oxidized, and HER2 S310F-reduced), three independent MD simulations (three replicas) were run. All structures and trajectories analyses were performed with VMD (19).

Perturbation response scanning (PRS) analysis was utilized to map the dynamic effect of allosteric mutations on the structure of HER2 (26, 27). PRS simulates a perturbation at a residue and computes the propagation of the response across the protein structure. The procedure is repeated for all residues, resulting in a response matrix P. The averages over the rows and columns of the normalized matrix P reveal the network of sensors and effectors, and allows to visualize the potential dynamic effect of an allosteric mutation. PRS was computed using ProDy on the HER2 crystal structure (PDB ID: 6J71; ref. 28).

The project GENIE (Genomics Evidence Neoplasia Information Exchange) data analyzed in this study are available from the Synapse platform (http://www.synapse.org/genie). Restrictions apply to the availability of the FMI cancer-associated HER2 mutation data, which are available from the authors upon reasonable request with the permission of the Foundation Medicine, Inc.

Comprehensive collection of cancer-associated mutations has revealed that HER2 contains 820 and 2616 unique SNVs based on the project GENIE data and the Foundation Medicine, Inc. data, respectively. These SNVs are distributed throughout the entire amino acid sequence of HER2, encompassing the ECD, the TMD, the JMD, the KD, and the CTD (Fig. 1A). The 2,616 SNVs detected in the Foundation Medicine dataset consist of 2,132 missense mutations, 242 nonsense mutations, 128 indel mutations, and 114 splice site mutations. Of these 2,616 SNVs, a small subset corresponded to previously described driver mutations (Supplementary Table S1; ref. 29). Mutation association data analysis revealed that the vast majority of HER2 driver mutations are observed independently of gene amplification. Among 10,874 instances of HER2-mutant tumors within the Foundation Medicine dataset, only 10% (1,094 cases) presented with copy-number amplification. The majority of HER2 SNVs (2,586/2,616) consists of rarely observed mutations (<0.01%), with only 30 SNVs ranging from intermediate (0.01%–0.1%) prevalence to high (>0.1%) prevalence (Fig. 1B). As there is growing recognition that even rarely occurring mutations can promote potent oncogenic activity (30, 31), we sought to identify the full spectrum of rare and recurrent oncogenic HER2 mutations.

To facilitate this discovery of oncogenic mutations beyond the previously identified HER2 driver mutations using a scalable approach, we developed a computational prediction algorithm, called MAP, based on numerous genomic and proteomic features deemed critical for oncogenicity prediction. This initial computational approach was applied to the fully annotated set of 820 mutations within the project GENIE dataset. The in silico MAP analysis produced a list of 222 mutations with predicted oncogenicity, consisting of 20 already known driver mutations and 202 potentially new driver mutations (Fig. 1A). Next, we evaluated the oncogenicity of HER2 mutations in the context of cell proliferation, as the expression of HER2 oncogenic mutants, but not the HER2 WT, has been shown to transform mouse hematopoietic Ba/F3 cells to IL3 independence. Lack of endogenous murine erbB family member protein expression in Ba/F3 cells (Supplementary Fig. S1; ref. 17) provided a suitable model for assessing transformation by overexpressing an oncogene, albeit precluding HER2/erbB heterodimerization, and assessing the ability for the cells to proliferate in the absence of IL3. We selected half of the predicted 222 mutations for experimental validation in the Ba/F3-retrovirus proliferation assay, including mutations located in various domains (48 in the ECD/TMD region and 63 in the JMD/KD/CTD region) and mutations ranging from rare to high prevalence, and determined 37 mutations to be mitogenic/oncogenic and 74 mutations to be neutral (Fig. 1A; Supplementary Table S2). Additional tumor sample data, including ethnicity, age, sex, race, cancer type, and a list of genes with at least one mutation, reported along with these HER2 mutations, are provided in Supplementary Table S3. The analysis of variant allelic fraction (VAF), representing the ratio of each HER2 SNV to WT of these experimentally validated mutations shows that VAF values from different tumor samples vary greatly and do not appear to correspond with the oncogenicity of HER2 mutations (Supplementary Fig. S2). This extensive collection of annotated HER2 mutations includes 15 previously described and 22 newly identified driver mutations. Seven oncogenic driver mutations are expressed with high prevalence (≥0.1%), nine driver mutations with intermediate prevalence (0.01%–0.1%), and 21 driver mutations are considered as rare/private mutations (<0.01%; Fig. 1B). These oncogenic/driver mutations are distributed throughout the entire amino acid sequence of HER2, nearly all occurring outside of the ATP-binding site. They are often found aggregated together within several functionally critical regions of HER2, forming allosteric hotspots or mutational clusters defined by their impact on protein structure and dynamics. Numerous mutations found in the N-terminal half of HER2 form a cluster in the ECD (Fig. 1A). This cluster is represented by the most prevalent HER2 mutation S310F that is allosterically located in the extracellular CR1 domain involved in dimer activation mechanisms (1, 7, 32). A cluster of prevalent intracellular KD mutations, L755S, exon 20 insertions (A775-G776insYVMA), D769Y, V777L, V842I, T862A, and L869R, are found proximal or within several key structural motifs, such the α-C helix, DFG motif, and A-loop (33, 34). Another prevalent mutation, R678Q, is located in the JMD and it is involved in the rearrangement of TMD/JMD dimerization upon activation (1). We confirmed the oncogenicity of these nine mutations in the context of Ba/F3 cell proliferation assays (Fig. 1A), and these findings are consistent with previous findings (1, 34–42). In addition, our validation approach determined the oncogenic driver status of a TMD mutation G660D (1) with medium prevalence (Fig. 1A) and the extracellular G292R/C mutations (39) with low prevalence in CR1 domain (Fig. 1A; Table 1). These findings support the previous assessment of these mutations as likely oncogenic (1, 39, 43), and the presence of allosteric hotspots formed by low-prevalence oncogenic mutations clustering together with highly prevalent driver mutations in discrete domains.

Tumor-specific analysis of validated oncogenic HER2 mutations reveals a unique tumor tropic distribution of HER2 oncogenic mutations among different cancer types. For example, a cluster of KD mutations at residues L755, V777, and V842 constitutes the predominant mutations in breast and colorectal cancers, while observed at much lower prevalence in lung cancers (Fig. 2). Similarly, the exon 20 insertions and R678Q mutations are the prevalent HER2 mutations in lung and colorectal cancers, respectively, while representing minor fractions in other cancer types (Fig. 2). In contrast, S310F/Y mutations are most prevalent in bladder cancer and consistently represent a sizable fraction (∼10%–15%) in other cancer types, reflecting their agnostic oncogenicity (Fig. 2; ref. 44). Considering that rare driver mutations collectively make up a considerable fraction (10%–19%) of these tumors (Fig. 2; Supplementary Fig. S3), the identification of new driver mutations by employing the MAP technology has provided new opportunities to cluster together rare and prevalent oncogenic mutations to further elaborate on mutational hotspots irrespective of prevalence and tumor type.

We mapped 37 driver and 74 neutral mutations on a full-length HER2 dimer model to examine the molecular attributes of oncogenicity (Fig. 3A; Table 1). Among the 111 mutations tested, 16 of 48 ECD/TMD mutations and 21 of 63 JMD/KD/CTD mutations were determined to be driver mutations. Comparison of proteomic features of driver and neutral mutations revealed several key attributes associated with oncogenic mutations in the ECD and KD (Fig. 3B). As expected, driver mutations have higher tendency to occur where highly conserved and functionally important amino acid residues reside. Structurally, driver mutations are associated with relatively stable motifs (medium dynamics), yet more likely to be associated with conformational changes than neutral mutations. Furthermore, the ECD driver mutations, such as S310F, are uniquely found near intramolecular disulfide bonds, whereas the KD driver mutations, such as L755S, are in close proximity to the ATP-binding site and the asymmetrical KD dimer interface (Fig. 3B). Several less frequently occurring driver mutations, such as mutations at residues G292, G660, and L869, are spatially proximal to prominent driver mutations to form the ECD cluster around S310 (Fig. 3C), the TMD/JMD cluster including R678 (Fig. 3D) and the KD cluster with L755, V777, and V842 (Fig. 3E).

Many HER2 oncogenic mutations affect the extracellular region, and are situated near sites involved in dimerization. Covalent dimerization has been one mechanism described to promote oncogenicity among the ERBB family of RTKs, including oncogenic EGFR ECD mutations expressed in glioblastoma (45) and a select group of HER2 ECD mutations (36). We sought to determine whether this might be a general mechanism for oncogenicity for HER2 ECD mutations. Among 14 ECD driver mutations evaluated, mutations at nine positions result in covalent homodimerization, as detected by resolving proteins under nonreducing conditions (Fig. 3F). This includes the S310F mutation, the most commonly occurring HER2 mutation found in human cancers (Fig. 1; ref. 8), and is the first time this mechanism has been ascribed to this mutation. By analyzing monomer-dimer fractions of phosphorylated HER2, we found that while the covalent dimer constitutes a minor fraction of total HER2 receptors, nearly all phosphorylated receptors, indicating the active form of the receptor, were present in the covalent dimer fraction. Previous analyses that focused on monomer-dimer fractions of total HER2, and not phosphorylated HER2, may have hindered the detection of covalent dimer as the active form of the mutant. Indeed, Greulich and colleagues (36) failed to measure the active form of the commonly occurring HER2 S310F mutation. While covalent dimerization of HER2 has been previously proposed to facilitate oncogenic activation of ERBB proteins (36, 45, 46), the frequent occurrence of ECD driver mutations proximal to disulfide bonds suggests covalent homodimerization as a common mechanism of activation among HER2 ECD mutations. Interestingly, earlier studies by Siegel and colleagues conducted in mouse models expressing high levels of HER2 WT described latent development of mouse mammary tumors being not coincident with the expression of high levels of HER2 (47), but with the subsequent acquisition of HER2 mutations, such as short truncations occurring in the ECD CR2 that produced covalent dimers (46, 47).

HER2 S310F/Y mutations induce covalent homodimerization of activated proteins through intermolecular disulfide bonds. This mutation site is directly adjacent to a disulfide bond formed between C299 and C311 within CR1 (Fig. 4A), and is part of a critical interface that engages with the dimer arm of adjacent HER2/ERBB molecule (Fig. 3C; ref. 48). To gain insights into the potential structural changes associated with the substitution of Ser310 with Phe or Tyr, we generated a minimized homology model of HER2 S310F (PDB ID: 6J71) mutant in solvent (Supplementary Fig S4). The model suggests that this amino acid position can accommodate a considerably larger amino acid side chain of either Phe or Tyr without any obvious steric effects on nearby residues. Next, we explored the mutational impact of S310F on the stability of local structural motifs, including a β-hairpin motif formed by residues 299–312 and the C299/C311 disulfide bond. We performed MD equilibrium simulations of the HER2 ECD in the presence and absence of S310F mutation (Fig. 4B). Our initial simulations performed for 100 nanoseconds (ns) with an intact C299–C311 disulfide bond show stable CR1 structures for both HER2 WT and HER2 S310F (Fig. 4B). Minimal deviations of simulated oxidized structures from the starting models are consistent with the stable CR1 structure supported by numerous disulfide bonds and the limitation of simulations time. Subsequent simulations with C299 and C311 as free cysteines, revealed similarly stable WT structure, whereas the associated β-hairpin motif of the S310F mutant became markedly unstable (Fig. 4B), indicated by increased root mean square fluctuations and the decreased stability of secondary structures (Supplementary Fig. S5). These observations suggest that the S310F mutation likely destabilizes the local β-hairpin motif and interferes with the disulfide bond formation between C311 and C299, resulting in unpaired Cys residues at the ECD dimer interface.

We confirmed the direct contributions of C299 and C311 in covalent homodimerization of HER2 S310F by showing that Ser substitutions of these Cys results in reduced dimer formation (Fig. 4C). The residual dimer formation of the HER2 S310F/C311S/C299S mutant under nonreducing conditions suggests the possibility of S310F generating other unpaired Cys residues involved in nearby disulfide bonds to form intermolecular disulfide bonds. To gain insight into potential long-range effects of S310F and other HER2 ECD mutations, we performed PRS analysis of the HER2 ECD (27). Simulated effects of perturbation-induced structural alterations of the HER2 ECD suggest that the S310 site and many residues near a putative ligand-binding site are effective transducers of long-range perturbation affecting the CR2 subdomain, which contains numerous disulfide bonds and forms the secondary ECD dimer interface (Supplementary Fig. S6). It also remains conceivable for the HER2 S310F mutation to facilitate HER2 activation through enhanced noncovalent dimerization, as previously proposed (36). Hence, our observations can be reconciled with previous studies by proposing that the HER2 S310F/Y mutation induces the covalent homodimerization of HER2 through the intermolecular disulfide bond formation facilitated by an intermediary noncovalent interaction.

To establish the general applicability of covalent homodimerization mechanism of HER2, we characterized two other oncogenic HER2 variants, p95-M611 and Δ16, with free cysteines present in the CR2 domain (Fig. 4A). We used U87MG human glioma cells to express HER2 (Fig. 4D; Supplementary Fig. S7). HER2 p95-M611 is generated through aberrant initiation of translation, producing an N-terminally truncated oncogenic form missing a significant portion of the ECD. HER2 Δ16 is generated by aberrant splicing, producing an oncogenic form that excludes exon 16 in the CR2 domain. As the CR2 domain directly participates in the ECD dimer interface, the majority of p95-M611 and Δ16 HER2 variants resulted in activated covalent dimers, involving C623 of HER2 p95-M611 and C626/C630 of HER2 Δ16 in the intermolecular disulfide bond formation (Fig. 4D).

We sought to determine the effect of HER2 oncogenic mutations that promote covalent homodimerization on the activity for the antibody trastuzumab. The trastuzumab binding site (HER2 residues 579–625) is located in the ECD CR2 (Fig. 4A). While trastuzumab is effective at reducing the amount of total HER2 WT and consequently interfering with the formation of phosphorylated dimers, it does not effectively inhibit HER2 dimer phosphorylation for each of the three HER2 mutants examined (HER2 S310F, HER2 p95-M611, and HER2 Δ16; Fig. 4E). Similarly, we observed trastuzumab markedly affected proliferation of BT474 cells expressing HER2 WT (64% inhibition at 1 µg/mL), but not of Ba/F3 cells individually expressing HER2 p95-M611 and HER2 Δ16 (p95-M611: 0% inhibition, Δ16: 11% inhibition; Fig. 4F). These results demonstrate how trastuzumab activity is lost either by alterations that directly affect the trastuzumab-binding site (HER2 p95-M611) or by alterations through allostery (HER2 Δ16). We therefore postulate that it should not be possible to optimally reposition trastuzumab to treat tumors expressing these HER2 oncogenic variants. Indeed, clinical data have shown reduced efficacy for trastuzumab in breast tumors exhibiting high HER2 p95-M611 expression (49–51) or in other tumor types such as endometrioid cancer where HER2 p95-M611 expression is also high (52).

Previous studies have shown that erlotinib, an effective tyrosine kinase inhibitor (TKI) against EGFR ATP-site mutations including the E746-A750del mutation, induces covalent dimerization and activation for select allosteric EGFR oncogenic mutants (45). This results in paradoxical stimulation of cells expressing these mutants. We sought to understand whether covalently activated HER2 oncogenes also exhibit altered pharmacology for small molecules binding to the ATP site. We tested the effect of the FDA-approved inhibitor lapatinib on the proliferation of cells driven by HER2 WT (BT474) or the panel of BaF3 transformants expressing HER2 S310F, HER2 p95-M611, or HER2 Δ16 (Fig. 5A; Supplementary Table S4). For each mutant, lapatinib exhibited reduced potency versus HER2 WT. Because reduced activity for select EGFR TKIs against covalently activated EGFR mutants is associated with the propensity of the TKI to induce covalent dimerization, we tested the effect of lapatinib on covalent homodimerization for the HER2 p95-M611 mutant. Here, we similarly observe a dose-dependent increase in covalent homodimerization of HER2 p95-M611 following lapatinib treatment (Fig. 5B and C). These observations suggest that HER2 allosteric mutants with enhanced covalent homodimerization have altered pharmacology that renders existing TKIs ineffective (Fig. 6).

The expanded use of next-generation sequencing testing has revealed hundreds of unique mutations affecting HER2, which are distributed throughout the amino acid sequence, including the ECD, the TMD, and the KD. We have developed a MAP-scoring algorithm based on machine learning that can be used to predict for oncogenic mutations in silico, followed by scaled biology to experimentally validate oncogenicity in isogenic cell panels. We applied this technology to HER2 to reveal a spectrum of 37 driver mutations and 74 neutral mutations. This spectrum of driver mutations consists of 15 previously characterized oncogenic mutations and 22 newly characterized mutations that markedly increased Ba/F3 cell proliferation. All validated mutations, except for T862A, which occurs at the orthosteric ATP-binding site of HER2, are located at allosteric positions throughout the amino acid sequence. Further characterization of validated mutations revealed distinct proteomic fingerprints of oncogenic mutations clustered together within the ECD and KD of HER2. Some mutations are prevalent across tumor types, while other mutations are very rare. Our data demonstrate how even rarely occurring mutations in HER2 can exhibit potent oncogenicity and possess similar molecular attributes as prevalent driver mutations. As the number of unique mutations is expected to rise over time, an effective means to identify oncogenic mutations, such as MAP technology, will be crucial for expanding the use of precision medicine to treat cancers driven by rare oncogenic mutations.

Our study showed that covalent homodimerization is a common mechanism of oncogenicity for HER2 mutations affecting the ECD (Fig. 6). While covalent dimerization of HER2 has been previously proposed to facilitate enhanced activation of a limited number of HER2 mutations (36), we demonstrated that covalent homodimerization mechanisms have a common oncogenic role among several HER2 ECD allosteric mutations. This mechanism parallels our previous characterization of oncogenic EGFR mutations in glioblastoma, where most ECD mutations are activated as covalently linked dimers, including the EGFR variant III forming covalent dimers through EGFR C307 that is structurally and functionally equivalent to HER2 C311 (45, 53). We demonstrate for the first time how the HER2 S310F mutation, the most common HER2 mutation, is activated as a covalently linked dimer. In silico analysis of the HER2 ECD structure suggests the S310F mutation induces both localized and long-range structural disorders to generate free cysteines that engage in the intermolecular disulfide bond formation. This key observation was overlooked in prior studies that failed to address the oligomeric state of the phosphorylated HER2 S310F mutant (36). Our analysis of the phosphorylated pool of HER2 under nonreducing conditions reveals that activated HER2 proteins are largely covalent dimers when the HER2 S310F mutation is present, and this is true for at least six other ECD mutants of HER2. Covalent homodimerization as a mechanism for oncogenic activation pertains not only to HER2 genomic mutations, but also to HER2 oncogenic variants produced by alternative initiation of translation (HER2 p95-M611) and by alternative splicing (HER2 Δ16). There are numerous examples of other oncogenes undergoing activation through mutation-induced covalent dimerization (54).

We showed that these allosteric HER2 mutations exhibit a unique pharmacology that renders existing anti-HER2 therapies ineffective toward HER2 allosteric mutants (Fig. 6). These HER2 allosteric mutants demonstrate resistance toward trastuzumab, the anti-HER2 antibody that recognizes the ECD of HER2 WT, consistent with clinical data that show reduced effectiveness of trastuzumab in patients expressing the ECD variant HER2 p95-M611 (50). We found that the small-molecule TKI lapatinib displayed reduced activity against allosteric HER2 oncogenes versus HER2 WT. This reduction of activity was associated with the TKI-induced propensity of these mutants to promote HER2 dimerization (Fig. 6). This finding also parallels our observations of paradoxical dimer induction first described for EGFR-mutant oncogenes expressed in glioblastoma (45). The unique pharmacology that characterizes this family of HER2 allosteric oncogenes means that existing HER2 WT directed therapies (including antibodies and small molecules) may not be optimally used to treat cancers expressing HER2 allosteric oncogenes. Because allosteric mutations can introduce global changes in protein conformation that ultimately impact the pharmacology of the ATP-binding site (Fig. 6), screening of HER2 oncogenes as intact proteins in whole-cell assays is necessary for the design of a potent inhibitor against this family of HER2 allosteric oncogenes. Indeed, we have recently deployed our unique screening strategy to design a potent and selective inhibitor of allosteric ERBB mutants, including the allosteric HER2-mutant family addressed in this study (55). Recognition of mutation-induced dimerization as a general mode of activation among RTK oncogenes beyond the ERBB family points to the necessity of novel inhibitor design approaches to target allosteric mutations currently unaddressed by existing RTK-directed therapies.

N. Ishiyama reports a patent for METHODS OF CHARACTERIZING AND CLASSIFYING SOMATIC VARIANTS pending to none and being an employee and shareholder of Black Diamond Therapeutics. A. Salomatov reports a patent for METHODS OF CHARACTERIZING AND CLASSIFYING SOMATIC VARIANTS pending to none. D. Romashko reports a patent for ASET-023 - Methods of Treating Cancers with Alkyne Substituted Quinazoline Derivatives pending. S. Thakur is an employee and shareholder of Black Diamond Therapeutics. A. Mentes is an employee and shareholder of Black Diamond Therapeutics. J.F. Hopkins is an employee of Foundation Medicine Inc. and a stockholder of Roche Holdings AG. G.M. Frampton is an employee of Foundation Medicine Inc. and shareholder of Roche AG. L.A. Albacker reports employment at Foundation Medicine, Inc.. and stockholder of Roche Holdings AG. C. Roberts reports personal fees from Black Diamond Therapeutics, Inc. during the conduct of the study and personal fees from Civetta Therapeutics, Outrun Therapeutics, and Stablix Therapeutics outside the submitted work. E. Buck reports a patent for Methods of Characterizing and Classifying Somatic Variants pending. No disclosures were reported by the other authors.

N. Ishiyama: Conceptualization, data curation, software, investigation, visualization, methodology, writing–original draft, writing–review and editing. M. O'Connor: Resources, investigation, visualization, methodology, writing–original draft. A. Salomatov: Conceptualization, data curation, software, investigation, methodology, writing–original draft. D. Romashko: Resources, investigation, visualization, methodology. S. Thakur: Data curation, investigation, visualization. A. Mentes: Software, investigation, visualization, methodology, writing–original draft. J.F. Hopkins: Data curation, investigation, writing–review and editing. G.M. Frampton: Data curation, investigation. L.A. Albacker: Data curation, investigation, writing–original draft. A. Kohlmann: Software, supervision, investigation, writing–original draft, project administration, writing–review and editing. C. Roberts: Supervision, investigation, project administration, writing–review and editing. E. Buck: Conceptualization, supervision, investigation, writing–original draft, project administration, writing–review and editing.

Professional writing assistance was provided by Emily Cullinan, PhD, and Francesca Balordi, PhD, of The Lockwood Group (Stamford, Connecticut) in accordance with Good Publication Practice (GPP3) guidelines, with funding by Black Diamond Therapeutics, Inc.

The publication costs of this article were defrayed in part by the payment of publication fees. Therefore, and solely to indicate this fact, this article is hereby marked “advertisement” in accordance with 18 USC section 1734.