Genetic Heterogeneity and Tissue-specific Patterns of Tumors with Multiple PIK3CA Mutations

PIK3CA is mutated in a large number and wide variety of cancer histologies, and is one of the most frequently mutated oncogenes across cancer (1). PIK3CA encodes for p110α, the catalytic subunit of the PI3K lipid kinase complex, and mutated PI3K drives cancer cell signaling and tumor growth (2). PI3K has been a longstanding drug target in oncology, with numerous past clinical trials investigating dual PI3K/mTOR inhibitors, pan-PI3K inhibitors, and p110α-specific inhibitors across many different cancer types bearing wildtype and mutant PIK3CA genotypic backgrounds (3–5). In patients with PIK3CA-mutant tumors, although most of these trials led to improvement in progression-free survival (PFS), many of these drugs did not move forward in the drug development process due to significant on-target toxicities including hyperglycemia and rash (6–9). These studies led to the investigation of p110α-specific inhibitors in combination with the estrogen receptor (ER) degrader fulvestrant in patients with PIK3CA-mutant ER-positive (ER⁺) metastatic breast cancer. Alpelisib with fulvestrant improves PFS in this patient population with well-tolerated adverse events (10–12), and this led to FDA approval of this combination therapy.

Given the side effects of PI3K inhibitors, which can be dose limiting even in patients with PIK3CA-mutant tumors, efforts have been made to define genomic predictors of increased sensitivity to PI3K inhibition. We and others have reported a frequent phenomenon of double- and multiple PIK3CA (multi-PIK3CA) mutations in 10%–15% of PIK3CA-mutant cancers (13, 14), which typically appear in cis on the same allele (13). Double-PIK3CA mutations increase PI3K pathway signaling in vitro and in vivo (13, 14) and are associated with increased sensitivity to PI3K inhibitors in vitro and for patients with breast cancer (13). Together, these studies suggest that multi-PIK3CA mutations represent a bona fide oncogene. This work has led to the testing of PI3Kα inhibitors in early-stage clinical trials in patients with varied tumor types that bear multi-PIK3CA mutations (9).

While initial efforts to genomically characterize tumors with multi-PIK3CA mutations have been focused on breast cancer, their frequency, mutational positions, co-occurrence with other gene alterations, and associations with mutational signatures are poorly understood across other cancer histologies. Moreover, it is not currently known how specific combinations of multi-PIK3CA mutations in patients correlate with clinical tumor response. Here, we comprehensively characterize multi-PIK3CA mutant tumors in a genomic dataset from a large pan-cancer cohort and further examine response patterns in patients with breast cancer on the largest Phase III clinical trial testing a PI3K inhibitor to date.

Comprehensive genomic profiling of formalin-fixed, paraffin-embedded (FFPE) tissue sections of 352,392 samples across 66 tumor types was performed using FoundationOne®, FoundationOne®CDx, or FoundationOne®Heme in a Clinical Laboratory Improvement Amendments–certified, College of American Pathologists–accredited laboratory (Foundation Medicine Inc.). Our study cohort comprised a subset of 17,303 patient samples profiled using an older version of the FoundationOne® assay (2012–2014), that were consented for research at the time of our study; these samples were included in some previous studies (14, 15). Hybrid capture was carried out on at least 324 cancer-related genes and select introns from 34 genes frequently rearranged in cancer. For the breast cancer cohort, HER2 status was determined from the HER2 (ERBB2) amplification status based on the FoundationOne® assay and ER status was derived from accompanying pathology reports, where available. Approval for this study, including a waiver of informed consent and Health Insurance Portability and Accountability Act (HIPAA) waiver of authorization, was obtained from the Western Institutional Review Board (protocol no. 20152817). The Institutional Review Board granted a waiver of informed consent under 45 CFR § 46.116 based on review and determination that this research meets the following requirements: (i) the research involves no more than minimal risk to the subjects; (ii) the research could not practicably be carried out without the requested waiver; (iii) the waiver will not adversely affect the rights and welfare of the subjects.

Processing of the sequence data and identification of different classes of genomic alterations was carried out, as described previously (16). The mutations identified were short variants (base substitutions and indels), copy-number alterations, and rearrangement events. Tumor mutational burden (TMB) was calculated as the number of somatic base substitutions or indels per megabase (Mb) of the coding region target territory, and was determined on 0.8–1.2 Mb, in a manner described previously (17). Samples with a TMB of at least 10 mutations/Mb were classified as TMB-high, while the remaining samples were classified as TMB-low, where available. Microsatellite instability status was also determined through the analysis of homopolymer repeat loci, as described previously (17). Multi-PIK3CA specimens were defined as those with at least two known or likely pathogenic short variants in the PIK3CA gene.

PIK3CA mutation pairs within 500 nucleotides of each other were assessed for cis/trans orientation. Read pairs were extracted from alignment data and were limited to pairs where both reads had a mapping quality score greater than 25. To avoid confounding factors from potential sequencing error, cis/trans calling was performed for variant pairs with more than 10 read pairs that supported either the mutant or wildtype sequence at both variants. cis support was defined as mutation pairs with reads supporting both mutant sequences. trans support was defined as mutations with read support of one mutant sequence and one wildtype sequence. Mutation pairs were considered as cis-oriented when more than four read pairs supported the cis status.

Where AF represents the observed allelic fraction, mc denotes the estimated mutant copies, and wc denotes the wildtype copies, as estimated from somatic-germline-zygosity (SZG) method to distinguish somatic and germline origin of genomic alterations (18). Second, the highest identified VTF across all variants for a sample was used as an approximate for “tumor fraction” for the sample (TF). Third, the clonality of each identified alteration was determined by the ratio of the VTF to the estimated TF for the sample. Mutations present at a ratio of at least 50% were denoted as clonal while the remaining were annotated as subclonal mutations. Clonality estimation was only performed in a subset of individuals that satisfied quality filters from the SGZ output: quality control (QC) pass status and no copy number model switch.

Mutational signature calling was performed as described previously (19). All single base substitutions were included in the analysis except for known oncogenic driver alterations and predicted germline alterations. Samples suitable for the identification of mutational signatures were determined by their overall TMB distribution with samples exhibiting at least 20 assessable alterations being characterized into the following six major mutational signatures: alkylating; apolipoprotein B mRNA-editing enzyme, catalytic polypeptide (APOBEC); mismatch repair (MMR); DNA polymerase epsilon, catalytic subunit (POLE); tobacco; and ultraviolet (UV). A sample was deemed to have a dominant signature if a mutational class harbored a score of 0.4 or greater. Assessed samples without an identified dominant mutational signature were annotated as “None.”

Samples exhibiting multi-PIK3CA mutations were further assessed at both the mutation and codon level to identify specific patterns of enrichment based on tumor histology. Mutations were also tested for enrichment as a second hit versus a primary hit using a Fisher exact test for the univariate comparisons of proportions. In addition, the dataset was also interrogated for the patterns of mutual exclusivity and co-occurrence of multi-PIK3CA mutations with other cancer-associated genes, including a specific assessment of genes in the PI3K pathway based on tumor histology. A Fisher exact test was utilized to identify significantly co-occurring (OR > 1) as well as mutually exclusive (OR < 1) patterns using a P-value threshold of 0.05. A FDR-based correction for multiple testing was applied for this analysis.

NIH-3T3 cells expressing PIK3CA mutants E726K/H1047R and H1047R were serum starved as described previously (13), in five biological replicates per cell line. Samples were lysed, denatured, reduced, alkylated, and digested with trypsin. Serine and threonine phosphopeptides were enriched using titanium dioxide resin, and samples were labeled with isobaric mass tags (TMT 10-plex, Thermo Fisher Scientific). All labeled peptides were mixed, fractionated, and analyzed by MS-MS. Reporter ion intensities were quantified (MaxQuant) to determine the relative amounts of phosphorylated peptides of cells bearing E726K/H1047R compared with H1047R.

This study assessed circulating tumor DNA (ctDNA) samples from patients enrolled in the SANDPIPER trial (clinicaltrials.gov: NCT02340221) that examined the clinical efficacy and safety of taselisib plus fulvestrant versus placebo plus fulvestrant in patients with ER⁺/HER2⁻ locally advanced or metastatic breast cancer. Dosing, frequency, and time on study metrics have been described previously (6). Briefly, between April 9, 2015 and September 4, 2017, 631 patients were randomized to either the taselisib (n = 417) or placebo arm (n = 214). Median time on study was 10.8 months (range, 1.2–31.7 months) in the placebo arm and 11.2 months (range, 0–30.3 months) in the taselisib arm. Patients received 500 mg intramuscular fulvestrant (cycle 1, days 1 and 15; day 1 of each subsequent 28-day cycle) plus either taselisib or placebo until progressive disease or unacceptable toxicity. Additional details regarding dose interruptions and reductions are described by Dent and colleagues (6).

Genomic alterations in ctDNA collected at baseline were evaluated in 508 patients using the FoundationOne® Liquid assay, which targets 70 cancer-related genes, including PIK3CA. A total of 339 samples harbored one or more pathogenic PIK3CA single-nucleotide variants, of which 66 harbored two or more detectable PIK3CA mutations and were classified as multi-PIK3CA mutant tumors, and of which 273 samples harbored one detectable PIK3CA mutation and were classified as single-PIK3CA mutant tumors. Among the 66 patients with multi-PIK3CA mutant tumors, comparison of the objective response rates for different allelic combinations was performed using a stratified Cochran–Mantel–Haenszel test. Among the 339 patients with PIK3CA-mutant tumors, Kaplan–Meier analysis of PFS, from the clinical cut-off date for the primary analysis (October 15, 2017; ref. 6), were performed with the log-rank test using a Cox proportional hazards regression model to obtain HRs and 95% confidence intervals. To adjust for multiple comparisons, the Benjamini–Hochberg correction was used, wherein statistical significance was defined at a Benjamini–Hochberg–adjusted P-value (q-value) threshold of 0.05.

Statistics, computation, and plotting were carried out using Python 2.7 (Python Software Foundation) and R 3.6.1 (R Foundation for Statistical Computing).

All relevant data are provided within the article and its accompanying Supplementary Data. Because of HIPAA requirements, we are not consented to share individualized patient genomic data, which contains potentially identifying or sensitive patient information. Foundation Medicine is committed to collaborative data analysis, and we have well-established, and widely utilized mechanisms by which investigators can query our core genomic database of >600,000 deidentified sequenced cancers to obtain aggregated datasets. More information and mechanisms for data access can be obtained by contacting the corresponding authors or the Foundation Medicine Data Governance Council at [email protected].

We investigated a pan-cancer cohort comprised of 352,392 samples across 66 tumor types to comprehensively characterize tissue-specific patterns of tumors with multi-PIK3CA mutations and their potential clinical implications.

PIK3CA mutations were identified in 12.1% (42,556/352,392) of samples in the pan-cancer cohort, of which 10.7% (4,552/42,556) exhibited two or more PIK3CA mutations in the same sample (multi-PIK3CA). The clinical and demographic features for single- and multi-PIK3CA mutant tumors were generally similar with no significant differences in stage at presentation, age, and site of biopsy (Supplementary Table S1). A large fraction of multi-PIK3CA mutant tumors exhibited a high TMB, with 36% multi-PIK3CA mutant tumors exhibiting a TMB-high status (≥10 mutations/Mb; Supplementary Table S1). To account for the potential influence of high TMB on the presence of these multiple mutations, we excluded samples that had a definitive TMB-high status from our initial investigations.

After excluding samples that had a definitive TMB-high status, one or more PIK3CA mutations were identified in 11.2% (33,470/299,319) of samples in the pan-cancer cohort, with the highest prevalence observed in uterus/endometrial (36.0%), breast (32.3%), cervical (28.5%), and colorectal (17.0%) tumors (Fig. 1A). Multi-PIK3CA mutations were rarer, identified in less than 1% of the total cohort and in 8.7% (2,925/33,470) of samples with any PIK3CA alteration (Fig. 1A and B; Supplementary Table S2). Breast, uterine/endometrial, and cervical tumors exhibited the highest prevalence of multi-PIK3CA mutations, nearly 4% in each tumor type, compared with other tumor types. Among breast, uterine/endometrial, and cervical tumors with one or more PIK3CA mutations, multi-PIK3CA mutations were identified in 11.0%–13.5% of tumors (Fig. 1A). In contrast, among colorectal tumors with one or more PIK3CA mutations, multi-PIK3CA alterations were identified in only 7.6% of tumors.

An examination of the histologic subgroups within breast, cervical, and uterine/endometrial cancers revealed varying multi-PIK3CA patterns (Fig. 1C; Supplementary Table S3). Among breast tumors, invasive lobular carcinomas (ILC), predominantly ER⁺, exhibited a much higher prevalence of multi-PIK3CA mutations (8.5%, 133/1,568 samples) than other subgroups (P < 10⁻⁵), such as invasive ductal carcinomas (3.3%) and metaplastic carcinoma (2.0%). In uterine/endometrial tumors, the endometrioid histology showed the highest prevalence of multi-PIK3CA mutations (7.3%; P < 10⁻⁵), while serous and clear-cell subtypes exhibited a lower prevalence (∼2%). The two histologic subgroups of cervical tumors, cervical adenocarcinomas and squamous cell carcinomas, exhibited similar rates of multi-PIK3CA mutations (4.3% and 3.7%, respectively). Of note, multi-PIK3CA mutations were enriched in HER2⁻ breast tumors (P = 0.005), with a prevalence of 5.1% in the ER⁺/HER2⁻ and 2.7% in the ER⁻/HER2⁻ subgroup (Fig. 1D; Supplementary Table S4). Among ER⁺/HER2⁻ with one or more PIK3CA mutations, multi-PIK3CA mutations were found in 13.9% (76/546) of samples. In contrast, HER2⁺ breast tumors, despite having frequent PIK3CA mutations overall (32%, 97/292 samples), exhibited a lower prevalence of multi-PIK3CA mutations (∼1%, 3/292 samples) compared with HER2⁻ tumors (Fig. 1D; Supplementary Table S4).

Among samples with multi-PIK3CA mutations in the pan-cancer cohort, 93% exhibited exactly two PIK3CA mutations (double-PIK3CA). To determine the clonal status of these mutations, we computationally assessed the clonal fraction of the two PIK3CA mutations in each double-PIK3CA mutant case. Among the 1,495 double-PIK3CA mutant cases with available clonality prediction, in 67% (1,002/1,495) of cases, both of the PIK3CA mutations in a sample were determined to be clonal (clonal-clonal pairs), while 21% had one clonal and one subclonal mutation (clonal-subclonal pairs), and the remaining 12% were classified as subclonal for both mutations (Supplementary Table S5). Clonality varied by tumor histology (Fig. 1E; Supplementary Table S5). Breast and uterine/endometrial tumors exhibited elevated rates of clonal-clonal pairs, compared with the overall cohort (76% and 82%, respectively), with fewer than 5% of cases having both subclonal mutations. In contrast, only 65% of colorectal tumors were found to harbor clonal-clonal pairs and 13% were determined to harbor subclonal-subclonal pairs. Similarly, cervical and lung non–small cell carcinomas (NSCLC), although limited by the number of assessable mutations, showed relatively fewer clonal-clonal pairs (49% and 33%, respectively).

The high prevalence of clonal-clonal double-PIK3CA mutants prompted us to further interrogate the allelic configurations of the mutation pairs, which may result in different functional effects. Mutation pairs in cis (same allele) would lead to a single protein product with two mutations whereas those in trans (opposite alleles) would lead to two independent proteins with separate mutations. However, the assessment of allelic configurations, especially from FFPE samples, is often challenging (13), especially when the mutations being assessed are located outside of a single sequencing read. With these considerations, a subset of 513 double-PIK3CA mutants with mutation pairs within 500 nucleotides of each other were evaluated for their cis/trans orientation, based on the paired read support for each mutation (Supplementary Table S6). This approach generated a well-defined separation for the cis and trans orientation across tumor types (Supplementary Fig. S1). Double-PIK3CA mutants in breast and uterine/endometrial tumors were often found to occur on the same allele; 69% (99/143) and 67% (30/45) were characterized as in cis, respectively, in the two cohorts (Fig. 1F; Supplementary Table S6). In contrast, tumor types such as colorectal, NSCLC, and cervical showed a higher number of trans double-PIK3CA mutations, with a prevalence of 52%, 58%, and 80%, respectively (Fig. 1F; Supplementary Table S6).

We also assessed the interplay between the clonal dynamics and the allelic configurations for double-PIK3CA mutants, where feasible (Fig. 1G). The breakdown of the different tumor types for each clonal group and allelic configuration group is shown in Supplementary Fig. S2. Clonal-clonal pairs were largely identified to be in cis whereas double-PIK3CA mutants with subclonal PIK3CA mutations (clonal-subclonal or subclonal-subclonal) were often determined to be in trans.

To examine the mutational landscape of tumors with multiple PIK3CA mutations, we performed a co-occurrence analysis with alterations in other genes. Overall, in the pan-cancer cohort, multi-PIK3CA mutations were found to be largely mutually exclusive with alterations in many well-known cancer driver genes such as TP53, CDKN2A, CDKN2B, ALK, EGFR, STK11, BRCA1, and BRCA2, among others (Supplementary Fig. S3; Supplementary Table S7). Notably, multi-PIK3CA mutations were also found to be mutually exclusive with gene alterations in the cyclin D-CDK4/6-INK4-Rb pathway (CDK4, CDK6, and RB1). Specifically in breast cancers, alterations in ERBB2, BRCA1, and BRCA2 were not enriched in multi-PIK3CA mutant tumors (Supplementary Fig. S3). Oncogenes CCNE1 and MYC that are often amplified in tumors, were also identified to be mutually exclusive with multi-PIK3CA mutation status, particularly in breast and uterine/endometrial tumors (Supplementary Fig. S3; Supplementary Table S7). Mutations in other genes in the PI3K pathway were largely exclusive of multi-PIK3CA mutation status (Supplementary Fig. S4). Individual tumor types revealed additional tissue-specific patterns (Supplementary Fig. S4). Particularly, in multi-PIK3CA mutant breast cancer, our analysis revealed a statistically significant lack of alterations in the PI3K pathway genes PTEN (OR = 0.3, P < 10⁻⁵), PIK3R1 (OR = 0.2, P < 10⁻⁵), AKT1 (OR = 0.2, P < 10⁻⁵), and AKT2 (OR = 0.4, P = 0.03; Supplementary Figs. S3 and S4; Supplementary Table S7). However, in uterine/endometrial tumors, while alterations in most assessed PI3K pathway genes were not enriched in multi-PIK3CA mutant tumors, alterations in PTEN were frequent (OR = 2.9, P < 10⁻⁵; Supplementary Figs. S3 and S4; Supplementary Table S7). Overall, these findings suggest that multi-PIK3CA mutations identify a biologically unique subset of tumors which frequently lack other driver gene alterations.

In contrast, few genes were concurrently altered with multi-PIK3CA mutation status. Among the top co-occurring genes identified in the overall cohort were CDH1, MAP3K1, ESR1, and ARID1A (Supplementary Fig. S3). However, these patterns of co-occurrence are likely driven by tumor types that contribute significantly to the overall cohort of multi-PIK3CA mutants. For example, 43% (1,245/2,925) of the multi-PIK3CA mutant tumors in the overall cohort occurred in breast cancer samples (Supplementary Table S2). A closer inspection of tumor type–specific patterns also supported some possible underlying biases. Among the top concurrently altered genes in the overall cohort, CDH1, MAP3K1, and ESR1 were primarily identified in breast cancer (Supplementary Fig. S3; Supplementary Table S7). Even within breast cancer, we identified an elevated presence of multi-PIK3CA mutants in ILCs and ER⁺/HER2⁻ subgroups (Fig. 1C and D). Therefore, mutations known to be elevated within these subgroups [e.g., ESR1 mutations in endocrine-resistant ER⁺ breast cancer (20–22), CDH1 inactivating mutations in ILC (23)] may further contribute to the observed co-occurrence patterns. Taken together, the elevated presence of multi-PIK3CA mutants identified with these subgroups along with known subtype-specific gene alterations may result in the observed co-occurrence patterns. Beyond breast, other tumor types also showed few co-occurring gene alterations with multi-PIK3CA mutations (Supplementary Fig. S3; Supplementary Table S7).

We also examined differences in co-alteration prevalence between single-PIK3CA and multi-PIK3CA mutated tumors (Supplementary Fig. S5; Supplementary Table S8). Across all tumor types, alterations in TP53 were generally less prevalent in multi-PIK3CA mutant tumors compared with single-PIK3CA mutant tumors. Particularly in breast tumors, activating alterations in ERBB2 and MYC, and loss-of-function alterations in PTEN, were also less prevalent in multi-PIK3CA mutant tumors compared with single-PIK3CA mutant tumors. These findings further suggest that multi-PIK3CA mutant tumors represent a unique population compared with single-PIK3CA mutant tumors.

Previous studies have described the association of PIK3CA alterations with hypermutated tumors, particularly those exhibiting APOBEC-driven mutagenesis (24). We therefore examined the mutational burden and mutational signatures in the full cohort of multi-PIK3CA mutant tumors including both TMB-high and TMB-low groups to elucidate the underlying mutational processes in these tumors. Overall, multi-PIK3CA mutant tumors were more prevalent in the TMB-high cohort compared with the TMB-low cohort (3.1% vs. 1.0%, P < 10⁻⁵). This trend was consistent across TMB-high and -low cohorts across other multi-PIK3CA mutant tumor types as highlighted by uterine/endometrial (17.8% vs. 4.0%), breast (12.5% vs. 4.0%), and cervical tumors (12.9% vs. 3.8%; Fig. 2A; Supplementary Table S9). In addition, this relationship was likewise observed among all PIK3CA-mutated tumors (18% vs. 9%; Supplementary Table S9). Next, in a subset of 43,907 TMB-high samples where we could assess mutational signatures, we examined whether there were mutational signature differences based on the PIK3CA mutation status. We found that APOBEC and DNA MMR signatures were more frequently observed in PIK3CA-mutant tumors relative to samples that lacked a PIK3CA mutation (Fig. 2B; Supplementary Table S10). In contrast, mutational signatures corresponding to tobacco and UV showed a decreasing prevalence from non-PIK3CA mutant to single- and multi-PIK3CA mutant tumors (Fig. 2B). A closer inspection of each mutational signature group revealed that nearly 8% (558/7,071) of TMB-high tumors with an APOBEC signature had multi-PIK3CA mutations, compared with less than 1% (2,919/299,046) observed multi-PIK3CA mutations in the TMB-low cohort (P < 10⁻⁵; Fig. 2C; Supplementary Table S11). Although limited by sample size (n = 332), multi-PIK3CA mutant tumors were identified in 37.1% of tumors with a POLE mutational signature and were present at a similar frequency as the single-PIK3CA mutants (36.5%) in this mutational signature group (Fig. 2C; Supplementary Table S10). Similar to the TMB-low cohort, we also assessed the orientation and clonality of the double-PIK3CA mutant tumors based on their mutational signature. Overall, the double-PIK3CA mutations were frequently identified to be clonal across the different mutational signature groups, similar to those identified in the overall TMB-low cohort (Fig. 2C). In addition, nearly half of the double-PIK3CA mutations identified in TMB-high tumors with an APOBEC or MMR signature were determined to be in cis, consistent with our findings in the TMB-low cohort (Fig. 2C).

These findings of the APOBEC signature prompted us to further assess the TMB-high cohort based on the APOBEC signature status. Overall, both single- and multi-PIK3CA mutant tumors were more frequent in TMB-high tumors with APOBEC signature compared with other TMB-high or TMB-low tumors; multi-PIK3CA tumors had a prevalence of 7.9% in APOBEC-positive samples compared with 2.4% in TMB-high tumors that lacked the APOBEC signature (Fig. 2D; Supplementary Table S11). This trend was more pronounced in breast tumors, with multi-PIK3CA mutants identified in nearly 19% (290/1,531) of the APOBEC-positive tumors, compared with only 4% (1,243/31,224) in both the TMB-high tumors lacking the APOBEC signature and the TMB-low cohorts (Supplementary Fig. S6; Supplementary Table S11). Although cervical tumors showed a higher prevalence of multi-PIK3CA mutants in the TMB-high cohort, we only observed a slightly higher prevalence in APOBEC-positive tumors (13.7%) compared with those that lacked the APOBEC signature (11.3%; Supplementary Fig. S6; Supplementary Table S11). Taken together, these findings suggest an enrichment of multi-PIK3CA mutations in samples with high TMB, particularly in tumors also harboring the APOBEC signature.

We next investigated the specific codons present in multi-PIK3CA mutant tumors (Fig. 3). To reduce the potential influence of high TMB, we excluded samples with a TMB ≥ 10 mutations/Mb from this analysis. In the overall pan-cancer cohort of samples with two PIK3CA mutations, several mutation pairs consisted of the three frequent hotspot mutations in the helical domain (E542K, E545K) and the catalytic domain (H1047R) of the PIK3CA oncogene (Fig. 3A; Supplementary Table S12). Double-PIK3CA mutants with the mutations E542K:E545K were the most frequent, identified in 5% (135/2,727) of all pairs. In addition, mutation pairs consisting of a hotspot mutation and the E726K mutation were also frequent (3.7% E545K:E726K, 3.3% H1047R:E726K, and 2.5% E542K:E726K). In addition to the hotspot mutations, mutations in codons R88 (adaptor binding domain), E453 (C2 domain), Q546 (helical domain), E726 (kinase domain), and M1043 (kinase domain) were frequently observed in the PIK3CA mutation pairs (Fig. 3B; Supplementary Table S13). A codon enrichment analysis further revealed these codons as most significantly enriched in multi-PIK3CA mutant tumors compared with the single-PIK3CA mutant tumors (Fig. 3C; Supplementary Table S14), with E726 (OR = 16.7, P < 10⁻⁵) and E453 (OR = 6.8, P < 10⁻⁵) being the most statistically significantly associated codon mutations.

To interrogate whether double-PIK3CA mutants exhibit increased general PI3K pathway signaling in a tissue agnostic manner, we performed phosphoproteomics on NIH-3T3 mouse fibroblasts bearing PIK3CA mutations E726K/H1047R and H1047R. Notably, many known PI3K pathway serine and threonine phosphorylated substrates including AKT1, AKT2, eIF4B, CTNNB1, METTL1 (25), MAP2K4 (26), and NT5C (27) are enriched in E726K/H1047R versus H1047R (Supplementary Fig. S7).

Given the differences observed in the prevalence and configuration of multi-PIK3CA mutations across tumor types, we hypothesized that codon enrichment patterns may also show tumor histology-specific patterns. As expected, codon enrichment patterns exhibited unique patterns in each tumor histology (Fig. 3D; Supplementary Tables S12, S13, and S14). In breast tumors, double-PIK3CA mutants with E726K and the top three hotspots were the most frequent, accounting for nearly 14% of all observed pairs (Supplementary Table S12). Codons E726 and E453 were also significantly enriched in multi-PIK3CA mutants compared with single mutants (Supplementary Table S14). A predominant presence of E542K:E545K mutation pairs was identified in cervical tumors, accounting for 27% of all the pairs seen in this tumor type (Supplementary Table S12). Mutation pairs involving codons E542, E726, and E453 were significantly enriched in the multi-PIK3CA mutant cervical tumors (Supplementary Table S14). In contrast, uterine/endometrial tumors showed several pairs involving codon R88 and R93 (e.g., with R88Q; Supplementary Table S13). In addition, multi-PIK3CA mutants in uterine/endometrial tumors were also enriched for mutations in codons R38 and E453, among others (Supplementary Tables S13 and S14).

These findings highlight the underlying codon-based differences among multi-PIK3CA mutants based on tumor histology that might alter different domains of the PIK3CA oncogene, thereby resulting in different levels of functional activation.

We next sought to characterize how multi-PIK3CA mutant tumors respond to PI3K inhibitors and if specific codon mutations may cause differential responses to therapy. We analyzed ctDNA from patients randomized to the SANDPIPER clinical trial, a phase III clinical trial testing the addition of the PI3K inhibitor taselisib (GDC-0032) to fulvestrant in patients with ER⁺, HER2⁻ advanced breast cancer. In the entire multi-PIK3CA mutant population (n = 66; Fig. 4A), 35% (n = 23) received placebo + fulvestrant (Fig. 4B and E) while 65% (n = 43) received taselisib + fulvestrant (Fig. 4C and F). We examined allele-specific differences with respect to hotspot mutations by binning multi-PIK3CA mutant tumors by the presence of hotspot mutations (E542X, E545X, H1047X) or non-hotspot mutations (other mutations). Numerically, patients with multi-PIK3CA mutations responded more to taselisib + fulvestrant compared with placebo + fulvestrant with ≥ 2 hotspots (33.3% vs. 0%), H1047X + non-hotspot (23.5% vs. 0%), E542X/E545X + non-hotspot (33.3% vs. 11.1), or ≥2 non-hotspots (50% vs. 16.7%), as evaluated by objective response rate (ORR) per investigator's assessment (Fig. 4B and C). We also examined allele-specific differences with respect to E726X, a minor hotspot mutation among the most frequent double mutant–containing mutations (Fig. 4D). Numerically, patients with multi-PIK3CA mutations responded more to taselisib + fulvestrant compared with placebo + fulvestrant in the hotspot plus E726X (26.7% vs. 0.0%), and hotspot without E726X (30.8% vs. 8.3%) groups (Fig. 4E and F). There was no statistical difference in responses among multi-PIK3CA mutant allele-specific combinations to placebo + fulvestrant (P = 0.6323; Fig. 4B) or taselisib + fulvestrant (P = 0.7970; Fig. 4C).

We finally investigated whether our initial observation from the SANDPIPER study of increased ORR in patients with multi-PIK3CA mutated ctDNA treated with taselisib + fulvestrant (TAS + FUL) compared with placebo + fulvestrant (PBO + FUL; ref. 13) translated to PFS outcomes. Within both the multi-PIK3CA mutant group [TAS + FUL: 7.4 mos vs. PBO + FUL: 3.6 months; HR = 0.48 (0.26–0.90); P = 0.026; q = 0.051] and single-PIK3CA mutant group [TAS + FUL: 6.6 months vs. PBO + FUL: 3.7 months; HR = 0.69 (0.51–0.93); P = 0.019; q = 0.051], there was no statistically significant difference in PFS observed upon treatment with taselisib + fulvestrant compared with placebo + fulvestrant (Supplementary Fig. S8). With that said, a trend toward a larger treatment effect was observed in the multi-PIK3CA mutant group compared with the single-PIK3CA mutant group (Supplementary Fig. S8). The discordance between ORR and PFS among the patients with multi-PIK3CA mutant tumors may be a consequence of an underpowered PFS analysis due to the small size of the multi-PIK3CA mutated population in each treatment arm.

In this work, we comprehensively characterized multi-PIK3CA mutations in 66 different tumor types and measured allele-specific responses to PI3K inhibition in patients with breast cancer. Multi-PIK3CA mutations are frequent events found in 9% of all PIK3CA-mutant tumors with low TMB and 18% of all PIK3CA-mutant tumors with high TMB, and are most frequently found in breast and gynecologic cancers, with enrichments in ER⁺ HER2⁻ lobular breast cancers and endometrioid endometrial cancers. In most multi-PIK3CA mutant tumors, both PIK3CA mutations are clonal and in cis on the same allele. Double-PIK3CA mutations amplify PI3K pathway signaling in cells compared with single mutants. For double-PIK3CA mutant tumors, codons 726 and 453 are the most frequently associated secondary mutations observed pan-cancer, with amino acids in different domains appearing as the predominant secondary mutations in specific histologies. Multi-PIK3CA mutations are mutually exclusive with alterations in many oncogene and tumor suppressor drivers including genes in the PI3K pathway. Multi-PIK3CA mutant tumors are associated with a higher TMB and an APOBEC mutational signature. Data from the SANDPIPER trial showed that multi-PIK3CA mutant breast tumors respond similarly to fulvestrant with a PI3K inhibitor regardless of the combination of PIK3CA mutations.

Our data regarding frequency, clonality, codon enrichment, and clinical features are consistent with existing analyses of multi-PIK3CA mutant tumors (13, 14, 28). However, given our large sample size we were able to discern many higher-order correlations with respect to tumor histology subtypes, gene mutual exclusivity, and mutational signatures. PI3K inhibitors have been tested in basket trials (29, 30) and are being investigated in other tumor types. Thus, it is important to identify specific disease subtypes that may be more sensitive to PI3K inhibition. Our identification of multi-PIK3CA mutant enrichment in lobular breast cancers and endometrioid endometrial cancers has translational relevance for histology-specific clinical trials testing PI3K inhibitors. Notably, both these tumor types are driven by ER signaling. ER and PI3K signaling pathways have significant cross-talk, where inhibition of PI3K increases ER-dependent transcription (31, 32), raising the idea of combining additional antiestrogen therapies with PI3K inhibitors in these tumor types. One complicating factor is that in endometrial cancer, multi-PIK3CA mutants are frequently coaltered with PTEN, suggesting that additional genes may need to be assessed to optimally define an endometrial cancer population sensitive to PI3K inhibitors given that PTEN loss is a biomarker of resistance to PI3K inhibitors (33).

We previously phased the allelic configuration of double mutant tumors using long-range single-molecule real-time sequencing. Here, we phased cis/trans configurations for many tumors using next-generation sequencing alone which showed that in addition to breast cancer, a significant proportion of endometrial and colorectal cancers have multiple mutations in cis. We also showed that E726 and E453 alterations are most associated with multiple mutant tumors. We propose combining these two methods as a clinical tool to identify multi-PIK3CA mutant tumors in cis; while this approach would not identify all such patients, it would avoid the need for long-range sequencing for every patient which may not be feasible in a clinical trial. Alternatively, long-range sequencing could be used to prioritize PI3K inhibitor therapy for patients with tumors harboring double cis mutations in histologies with less clonality, such as colorectal, NSCLC, and cervical tumors. Because these generate different protein products, the tumor-type specific allelic configurations may have implications on the targeting of these alterations. The concurrent presence of clonal-clonal, cis double-PIK3CA mutations in non-breast cancers may further elevate PI3K activation and downstream signaling in these tumors. In breast cells, double mutations in trans confer similar levels of PI3K pathway signaling, growth, and PI3K inhibitor sensitivity compared with single mutations, and are less oncogenic by these measures compared with mutations in cis (13). However, additional studies are warranted for the comparison of these cis, clonal double-PIK3CA mutations with their subclonal or trans counterparts.

Functional modeling of double-PIK3CA mutations has demonstrated that it can drive growth factor–independent cell growth and tumor growth in mice (13, 14). Consistent with its role as a bona fide oncogene, double- and multi-PIK3CA mutations are mutually exclusive with most oncogenes and tumor suppressors including other genes in the PI3K pathway. ERBB2 alterations co-occurred with single-PIK3CA mutant but not multi-PIK3CA mutant tumors. These observations suggest a continuum model where large increases in oncogene dosage may negatively correlate with tumorigenesis (34). Single-PIK3CA mutations in mouse models do cause cancers, albeit with high latency periods and mixed histologies. Thus, multi-PIK3CA mutations may more accurately model PI3K pathway hyperactivity in vivo (9).

Which evolutionary pressures lead to multiple PIK3CA mutations in tumors is an outstanding question. Our work implicates both TMB and APOBEC as two features associated with multi-PIK3CA mutations. Notably, E453K, E542K, E545K, and E726K mutations all result from cytidine to thymidine transitions, which are the nucleotide hallmark of the APOBEC signature (35). The correlation between TMB and multi-PIK3CA mutations raises the possibility of testing PI3K inhibitors in combination with anti-PD1/PDL1 therapy.

While single-PIK3CA mutations are a biomarker of response to PI3K inhibitors, the clinical significance of the majority of rare multi-PIK3CA mutations is not well understood (9). With multiple mutations, this issue multiplies, with theoretically differential PI3K pathway activation levels for varying multi-PIK3CA mutation combinations, and our data demonstrate widely divergent patterns of PIK3CA mutation combinations in different tissue types. Recent biophysical data highlight a remarkable convergent mechanism of action of double-PIK3CA mutations involving both disengagement of p110:p85 interactions and increased membrane binding (36). Our data showing similar response rates of different allelic combinations to PI3K inhibition, albeit small numbers, suggests that as a biomarker, double/multi-PIK3CA may similarly converge with respect to clinical activity in breast cancer. Thus, even disparate multi-PIK3CA mutational patterns at the genomic level may result in similar levels of PI3K pathway dependence, functional, and pharmacologic sensitivity to PI3K inhibitors in vitro (13), and clinical response to PI3K inhibitors in patients (this work).

Multi-PIK3CA mutations correlate with increased clinical response to PI3K inhibitor therapy in breast cancer (13), based on retrospective analysis of a large prospective clinical trial (6). How this may translate to differences in overall survival warrants additional investigation. Ongoing clinical trials (NCT04589845, NCT05216432) now seek to determine whether the predicted benefit is observed across multiple cancer types by interrogating PI3K inhibitors in patients with multi-PIK3CA mutant tumors. As with any biomarker, moving beyond first-order approximations to identify biomarker-selected subsets of patients who experience better or worse responses upon treatment with a PI3K inhibitor is critical. Together, our findings describing the genetic heterogeneity and tissue specificities of multi-PIK3CA mutations will influence correlative analyses of patient samples from ongoing trials and will guide future PI3K inhibitor trial design.

S. Sivakumar reports personal fees from Foundation Medicine and other support from Roche during the conduct of the study; in addition, S. Sivakumar has a patent for clonality prediction methodology pending. D.X. Jin reports personal fees from Foundation Medicine, as well as other support from Roche during the conduct of the study. J. Ross reports personal fees from Foundation Medicine during the conduct of the study. L.C. Cantley reports grants from NCI R35 CA197588 and AACR/Stand Up to Cancer during the conduct of the study, as well as other support from Novartis, Faeth, and Loxo-Lilly outside the submitted work. M. Scaltriti reports personal fees from AstraZeneca, as well as other support from AstraZeneca outside the submitted work. J.W. Chen is a full-time employee of Genentech/Roche and holds stock in Roche. K.E. Hutchinson is a current employee of and stock owner in Roche/Genentech. T.R. Wilson reports other support from Genentech, Inc and Roche during the conduct of the study, as well as other support from Genentech, Inc and Roche outside the submitted work; in addition, T.R. Wilson has a patent for 16733928.2–1111 Methods of Treatment with Taselisib pending. E.S. Sokol reports other support from Foundation Medicine and Roche during the conduct of the study; in addition, E.S. Sokol has a patent for Foundation Medicine pending. N. Vasan reports other support from Novartis, Reactive Biosciences, Magnet Biomedicine, and Heligenics, Inc., as well as grants from Gilead outside the submitted work; in addition, N. Vasan has a patent for US20210189503A1 pending to Memorial Sloan Kettering Cancer Center. No disclosures were reported by the other author.

S. Sivakumar: Data curation, software, formal analysis, investigation, visualization, methodology, writing–original draft, writing–review and editing. D.X. Jin: Formal analysis. R. Rathod: Writing–original draft. J. Ross: Conceptualization. L.C. Cantley: Conceptualization. M. Scaltriti: Conceptualization. J.W. Chen: Investigation, writing–review and editing. K.E. Hutchinson: Investigation, writing–review and editing. T.R. Wilson: Writing–review and editing. E.S. Sokol: Conceptualization, resources, data curation, formal analysis, supervision, validation, investigation, visualization, writing–original draft, project administration, writing–review and editing. N. Vasan: Conceptualization, supervision, funding acquisition, investigation, writing–original draft, project administration, writing–review and editing.

We thank Ronald Hendrickson, Matthew Miele, Zhuoning Li, and the Memorial Sloan Kettering Cancer Center Microchemistry and Proteomics Core for assistance with phosphoproteomics. The authors acknowledge support from the NIH K08 CA245192 (N. Vasan) and the Susan G. Komen Career Catalyst Research Grant (N. Vasan).

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.