The Oncogenic PI3K-Induced Transcriptomic Landscape Reveals Key Functions in Splicing and Gene Expression Regulation

The phosphoinositide 3–kinase (PI3K) signaling pathway governs multiple cellular processes such as cell growth, proliferation, metabolism, translation, survival, and apoptosis. Aberrant hyperactivation of this pathway downstream of gain-of-function mutations in PIK3CA, the gene coding for the catalytic subunit of PI3K (p110α), contributes to multiple developmental disorders, cancer, and treatment resistance in different tumor types. Mutations in PIK3CA are found in up to 40% of estrogen receptor–positive (ER⁺) HER2-negative primary and metastatic breast tumors (1–4), and the combination of the PI3Kα-specific inhibitor alpelisib with the ER degrader fulvestrant is approved by the FDA for the treatment of patients with ER⁺, PIK3CA-mutated breast cancer (5).

More than 95% of human genes encode multiple mRNA isoforms as a result of alternative splicing, i.e., the differential selection of intronic/exonic sequences and use of alternative promoters and 3′-end polyadenylation sites, which increases protein diversity (6, 7). RNA splicing is a cotranscriptional enzymatic process through which a precursor mRNA is converted into the mature mRNA by removal of the noncoding regions, the introns, and ligation of the coding ones, the exons (8). Alternative RNA splicing determines cell differentiation, development, and tissue identity (9), and dysregulated splicing can contribute to tumorigenesis as well as tumor evolution and resistance to therapy (8, 10).

Previous studies have reported that the PI3K/AKT/mTOR pathway regulates the function of the serine/arginine-rich (SR) splicing factors directly (11–14). In addition, the PI3K pathway has been shown to affect alternative splicing indirectly, by modulating the activity of the SR protein kinases, SRPKs (15–18). AKT has been shown to interact with SRPKs, inducing their autophosphorylation and nuclear translocation to activate SR proteins and affect alternative splicing (17). Dual inhibition of PI3K and mTOR has also been shown to upregulate the splicing factor hnRNPM and to induce alternative splicing in Ewing sarcoma cells (19). However, compared with the extensive studies on protein modifications and signaling events that occur downstream of oncogenic PI3K variants, a systematic and comprehensive interrogation of the transcriptome and the genome-wide alternative splicing modulated by mutant PI3K in breast cancer has not yet been undertaken.

In this study, we aimed to comprehensively elucidate the genome-wide transcriptional changes and alternative RNA splicing and isoform usage dynamics driven by PI3K pathway activation in PIK3CA-mutant models and to interrogate their translational relevance in breast cancer patients. Our study provides the first comprehensive analysis of widespread splicing alterations driven by oncogenic PI3K in breast cancer.

Isogenic parental and PIK3CA^H1047R heterozygous mutants were purchased from Horizon Discovery. MCF-10A cells were maintained in DF-12 media supplemented with 5% filtered horse serum (Invitrogen), EGF (20 ng/μL; Sigma), hydrocortisone (0.5 mg/mL; Sigma), cholera toxin (100 mg/mL; Sigma), insulin (10 μg/mL; Sigma), and 1% penicillin/streptomycin. MCF10A parental and mutant cells were seeded in 6-multiwell plates in regular culture conditions to allow correct attachment and ensure ∼75% confluency at harvesting day. Twenty-four hours after seeding, cells were washed twice with PBS before adding the starvation media (without serum, EGF, and insulin). Where indicated, cells were treated with DMSO as control or alpelisib (1 μmol/L), taselisib (100 nmol/L), GDC0077 (100 nmol/L), GDC0068/ipatasertib (1 μmol/L), or RAD001/everolimus (100 nmol/L). For RNA sequencing (RNA-seq) and RT-qPCR experiments and to control PI3K signaling by Western blot, cells were treated for 4 hours, while for Western blot analyses of PTK2/focal adhesion kinase-related non-kinase (FRNK), cells were treated for 24 hours.

Mouse embryonic fibroblasts (MEF) were obtained via an Material Transfer Agreement (MTA; ref. 20) and infected with adenoviral GFP (parental cells) or adenoviral Cre to obtain homozygous Pik3ca^H1047R mutation (mutant cells). MEFs were cultured in DMEM supplemental with 10% FBS, 1% penicillin/streptomycin.

MCF7 were purchased from ATCC (ATCC HTB-22) and grown in DMEM/F12 supplemented with 10% FBS, penicillin/ streptomycin 1% under standard conditions.

T47D were purchased from ATCC (HTB-133) and grown in RPMI supplemented with 10% FBS, penicillin/ streptomycin 1% under standard conditions. All cells were routinely tested negative for Mycoplasma.

The PI3Ka-specific inhibitors alpelisib and GDC0077, the PI3Kα/γ/δ taselisib, the pan-AKT inhibitor GDC0068/ipatasertib, the mTORC1 inhibitor RAD001/everolimus were purchased (Selleckchem).

The ON-TARGETplus short-interference RNA (siRNA) smart pool against SRSF1 (L-018672–01–0005), SRSF2 (L-019711–00–0005), SRSF3 (L-030081–00–0005), SRSF7 (L-015909–00–0005), CELF1 (:L-020166–00–0005), HNRNPC (L-011869–03–0005), TRA2B (L-007278–00–0005), AFMID (L-032645–02–0005), and PTBP1 (L-003528–00–0005), or nontargeting (D-001810–10–05) were purchased from Horizon Discovery. Mutant MCF10A cells were plated in 6-well plates at 30% confluency and 16 hours later siRNA oligonucleotides were transfected using Lipofectamine RNAiMax (Thermo Fisher Scientific, 13778150) according to the manufacturer's protocol at a final concentration of 20 nmol/L. After 72 hours, mutant MCF10A cells were treated with alpelisib (1 μmol/L) in starvation media for 4 hours and then collected for RNA extraction.

RNA was isolated using the QIAGEN RNeasy Kit and retrotranscription was performed using the iScript cDNA synthesis Kit from Bio-Rad, following the manufacturer's instructions. cDNA was amplified by real time quantitative PCR in a ViiA 7 Real-Time PCR system, using SYBR Select Master Mix from Applied Biosystems. Sequences of the primers used are shown in Supplementary Table S1. For validation experiments, MCF10A parental and mutant cells were seeded in 6-multiwell plates in regular culture conditions to allow correct attachment and ensure ∼75% confluency at harvesting day. Twenty-four hours after seeding, cells were washed twice with PBS before adding the starvation media (without serum, EGF, and insulin). Cells were then treated with DMSO or with the different inhibitors of the PI3K pathway, in starvation media for 4 hours at –37°C. Cells were then flash-frozen and stored at –80°C until RNA extraction. The experiments were performed in triplicate. RNA from MCF10A cells was sequenced using Illumina HiSeq 2000 instrument at 2×150 nucleotide paired-end reads. RNA from MEF cells were sequenced using Illumina HISeq 2000 instrument at 2×100 nucleotide paired-end reads. Information on RNA-seq analyses can be found at Supplementary Data.

Cells were washed with PBS twice, and lysed in RIPA buffer supplemented with protease and phosphatase inhibitors (Roche). Total protein lysates were run in Nupage 4% to 12% Bis-Tris gradient precast gels (Invitrogen) in MOPS buffer. Membranes were probed using specific antibodies: PTK2/FAK (#3285S; RRID: AB_2269034) phosphorylated AKT (pAKT) (S473; #4060; RRID: AB_2315049), pAKT (T308; #13038; RRID: AB_2629447), total AKT (#4691; RRID: AB_915783), pS6 (S240/244; #5364; RRID: AB_10694233), pS6 (S235/236; #4858; RRID: AB_916156), HA-tag (#3724; RRID: AB_1549585), V5-tag (#13202; RRID: AB_2687461), β-actin (#4970S; RRID: AB_2223172), and vinculin (#13901; RRID: AB_2728768) were purchased from Cell Signaling Technology. The same antibody was used to detect both PTK2/FAK (125 kDa) and FRNK (43 kDa). All primary antibodies were diluted 1:1,000 and anti-rabbit IgG secondary antibody (GE Healthcare; RRID: AB_772206) (1:10,000) was used.

The AFMID-202 isoform was ordered and synthesized by Genewiz in a pUC57 vector with attB1 and attB2 sites in 5′ and 3′ of cDNA, respectively. The donor clone was recombined with pLx302 (Addgene #25896) or pIND20 (Addgene #44012) lentiviral vectors, using standard Clonase II LR mix (Thermofisher; catalog no. 11791100). The Gateway LR reaction was incubated at room temperature overnight. Following efficient transformation and DNA purification, the clones were sequenced before use. As a vector control for the pLx302 AFMID-202, pLx302 Luciferase was used.

MCF10A PIK3CA^H1047R, MCF7, and T47D cells stably expressing Tet-FRNK-EGFP, or pIND20 AFMID-202 were seeded evenly in a 48-well plate (10⁴), in 3 biological replicates for each condition, 24 hours prior to the initiation assay. For the cell lines stably bearing doxycycline-induced vectors, full or 1 mg/mL doxycycline-containing media was added. The cell number was counted at day 0, 3, and 6 with a Countess 3 Automated Cell Counter (Thermofisher Scientific).

MCF10A PIK3CA^H1047R mutant cells stably bearing a doxycycline-inducible vector for FRNK-EGFP overexpression were seeded in a 12-multiwell plate (1.5×10⁵) 48 hours prior to the assay, to obtain a completely confluent monolayer at the day of the assay. Twenty-four hours after plating, 1 mg/mL doxycycline was added to the media to induce FRNK-EGFP expression. Twenty-four hours after induction, sufficient FRNK-EGFP expression was assessed using a fluorescent microscope (Evos M5000, Thermo Fisher Scientific). Next, a straight wound was scraped at the center of each well using a sterile p200 tip, to create a gap in the confluent monolayer. Wells were then gently washed with starvation media (without serum, EGF and insulin) to remove detached cells and replenished with fresh starvation media. Wounds were imaged using Evos M5000 microscope (Thermo Fisher Scientific) immediately after scraping (Time 0 hour), and at different time points for a duration of maximum 24 hours to reduce the contribution of cell proliferation to fill the gap. The wound area was quantified using FiJi (ImageJ).

MCF7 and T47D parental cells were counted and 5,000 cells/well were plated in 96 well plates and incubated overnight in full media. The next day, the cells were washed with 1x PBS before adding the starvation media (without FBS). After 24 hours, the cells were treated with increasing doses of alpelisib in the presence of H3B-8800 (50 nmol/L), E7070 (100 nmol/L) or DMSO control for 5 days. The cell viability was measured with the Cell proliferation MTT Kit (Sigma-Aldrich cat. no. 11465007001) following the manufacturer's protocol.

For the T47D xenograft study, 0.72 mg/90d-release estrogen pellets were implanted into 6-week-old female athymic nude mice 3 days prior to the tumor cell transplantation. Ten million T47D cells per mouse were orthotopically transplanted.

For the MCF7 xenograft study, 0.18 mg/90d-release estrogen pellets were implanted into 6-week-old female NOD scid gamma mice 3 days prior to the tumor cell transplantation. Ten million MCF7 cells per mouse were subcutaneously transplanted.

The PIK3CA^H1047R mutant patient-derived breast cancer xenograft was generated as follows: 6-week-old NOD scid gamma female mice were implanted subcutaneously with specimens freshly collected from patients at Memorial Sloan Kettering (MSK) under an MSK-approved Institutional Review Board (IRB) biospecimen protocol #06–107. Tumors developed within 2 to 4 months and were expanded into additional mice by serial transplantation.

Treatment was started when tumor volumes reached approximately 200 mm³. Xenografts were randomized and dosed with alpelisib (25 mg/kg, orally 5 days a week), or vehicle as control (0.5% methocel, orally 5 days a week). At the end of the treatments (20–30 days), animals were sacrificed, and tumors were collected for Western blot and RNA analyses.

Mice were cared for in accordance with guidelines approved by the MSK Institutional Animal Care and Use Committee and Research Animal Resource Center.

Patients with known PIK3CA mutations were enrolled in an open-label, phase I dose escalation study of oral daily GDC-0077 alone (Arm A), in combination with endocrine therapy with or without palbociclib (Arm C: GDC-0077 + Letrozole; ARM D: GDC-0077 + Fulvestrant; Arm B: GDC-0077 + Letrozole + Palbociclib; Arm E: GDC-0077 + Fulvestrant + Palbociclib; Arm F: GDC-0077 + Fulvestrant + Palbociclib + Metformin prophylaxis; NCT03006172).

Written informed consent was obtained from each patient. The studies were conducted in accordance with recognized ethical guidelines (Belmont Report). Patients were enrolled to MSK-IRB # 12–245 and a core biopsy was collected and utilized for RNA-seq analyses. More details on patients’ enrollment in this trial are found at Supplementary Data file. Patient cohort information are found in Supplementary Table S2.

For each independent in vitro experiment, a minimum number of three experiments were performed to ensure adequate statistical power. One-way ANOVA and unpaired t test was performed on GraphPad PRISM, to compare samples in RT-qPCR and RIP-qPCR experiments, respectively (*, P < 0.05; **, P < 0.01; ***, P < 0.001). All information regarding statistical tests performed, ‘n’ values and P values per experiment can be found in the figure legends, figures, and results. The confidence level used for all the statistical analyses was 0.95 (α  =  0.05), unless otherwise specified. Computational analyses and statistics of all assays are provided in Supplementary Data file.

The datasets generated during this study [RNA-seq/ assay for transposase-accessible chromatin using sequencing (ATAC-seq) fastq files] are available at the Gene Expression Omnibus database under accession number GSE157858. The patient RNA-seq fastq files are available at dbgap database under the number: phs002840.v1.p1.

To understand the global transcriptomic regulation downstream of the PI3K pathway, we performed paired end, long read RNA-seq in two isogenic knock-in cellular models bearing the PIK3CA^H1047R activating mutation: MEFs and human nonmalignant mammary epithelial cells (MCF10A). (Supplementary Fig. S1A; Supplementary Fig. S1B and S1C). To assess effects of PI3K inhibition we also performed RNA-seq in both models after treatment with PI3Kα-specific inhibitors. An analysis of differential gene expression identified thousands of genes were altered by PIK3CA^H1047R as compared with parental control cells in MCF10A and MEF (Fig. 1A; Supplementary Fig. S1D–S1F), whose expression was also reversed by PI3Kα inhibition (Fig. 1B; Supplementary Fig. S1G; Supplementary Table S3). These observations established that PIK3CA^H1047R was primarily responsible for gene expression alterations in these isogenic models.

Gene set enrichment analysis (GSEA; ref. 21) identified cellular proliferation (MYC and E2F targets, G₂–M checkpoint) and metabolism (glycolysis, fatty acid metabolism, adipogenesis) as the most significantly altered pathways in PIK3CA^H1047R compared with parental cells (Supplementary Table S4). In addition, RNA binding, RNA processing, and mRNA splicing were also identified by GSEA in the Reactome database as significantly enriched pathways in mutant compared with parental cells, as well as negatively enriched in response to PI3Kα inhibition (Fig. 1C; Supplementary Fig. S1H; Supplementary Table S4), indicating that RNA splicing could be an important effector pathway downstream of oncogenic PI3K.

Seeking to define the network of transcription factors (TF) regulated by PI3K activation, we performed TF motif analyses in our models. We identified the AKT substrates FOXO (22), NRF1 (23), and AR (24), as well as E2F proteins (25), MYC (26), ELK1 (27), STAT5 (28), TCF4 (27), that have been previously described to be regulated by the PI3K pathway. Our analyses also predicted a panel of TFs such as PR, GR, RORA, YY1, AP1 (Supplementary Table S4), whose regulation through the PI3K pathway had not been previously reported. To further define the relevance of these TFs in modulating transcriptional outputs, we interrogated the chromatin landscape of mutant and parental MCF10A cells using the ATAC-seq. Widespread (n = 8,330) chromatin accessibility changes were mediated by mutant PI3K in comparison to parental cells (Supplementary Fig. S1I). Consistent with our RNA-seq TF activity analyses, de novo motif analysis identified AP1, ZEB1, YY2, nuclear receptor, ELK, TCF, and STAT among the top TFs motifs enriched upon PI3K pathway activation (Supplementary Fig. S1J). Altogether, our integrated RNA-seq and ATAC-seq analyses identified a robust set of TFs whose function is potentially regulated by activation of the PI3K pathway.

Previous studies have linked the PI3K/AKT/mTOR pathway to specific components of the splicing machinery (11–17). However, a systematic interrogation of splicing upon PI3K pathway activation and inhibition by RNA-seq has not been reported. To this end, we first performed RNA splicing analysis for differential usage of known exons and splice junctions using the JunctionSeq algorithm (29). The number of significantly altered events identified was greater in mutant MCF10A cells relative to parental cells (n = 4,444) and also in comparison to mutant cells treated with PI3Kα inhibitor (n = 1,539). The same analysis in MEF cells also identified widespread RNA splicing alterations mediated by the PIK3CA^H1047R mutation (n = 2092) and by PI3Kα inhibition (n = 735; Supplementary Table S5). Although, alternative splicing is known to be tissue- and species-specific (30), the intersection of differentially mis-spliced genes in our human and mouse models identified events in 6 genes, namely CCDN3, CKAP4, YPEL, RPL27, COG5, and IP6K2.

The inverse correlation of significant events between mutant cells versus parental and mutant cells versus PI3K inhibition provided evidence that PI3Kα inhibitor treatment can reverse RNA splicing events in both MEF and MCF10A lines (Fig. 1D; Supplementary Fig. S1K). We established a confident set of 618 isoforms in MCF10A cells and 283 isoforms in MEFs that produced statistically significant alternative splicing events that were inversely correlated between PIK3CA^H1047R and parental cells and also rescued by PI3Kα inhibition (Fig. 1D and E; Supplementary Fig. S1L). Annotation of these splicing events revealed that the most frequent isoform level changes were exon skipping from single and multiple exons and alternative transcription termination and start sites (Fig. 1F; Supplementary Fig. S1M; Supplementary Table S5). A gene ontology (GO) analysis of the 618 isoforms in MCF10A identified enrichment of several biological pathways, including: translation, RNA catabolic process, mRNA, and peptide metabolic process, ribosome biogenesis, and cell cycle (Supplementary Table S6).

In addition to JunctionSeq (29), we also confidently assessed RNA splicing and isoform usage alterations using the established algorithms IsoformSwitchAnalyzerR (31), rMATS (32), and SUPPA2 (33). For splicing validations, we selected the most significant candidates that were detected by at least two algorithms, with the same splicing event reversed by PI3Kα inhibition for each algorithm. Among the 71 genes found significantly spliced by at least two algorithms (Supplementary Table S7), 10 candidates with gene functions involved in splicing, migration, cellular growth, and metabolism were randomly selected and validated: ​​AFMID, encoding for arylformamidase, involved in tryptophan degradation; CCND3, encoding the cell-cycle promoting cyclin D3; HNRNPC, encoding the splicing factor heterogeneous nuclear ribonucleoprotein C; IP6K2, encoding the inositol hexaphosphate kinase 2; MTUS1, encoding the microtubule-associated scaffold protein 1; PTK2/FAK, encoding the focal adhesion kinase; SRP68, encoding the Signal Recognition Particle 68; SPRK2, encoding the splicing factor SRSF kinase 2; RRAS2, encoding the signal transducer RAS related 2; and UNKL, encoding a putative E3 ubiquitin ligase Unk like zinc finger (Fig. 1G).

Next, we validated splicing events within the 10 candidates in parental and mutant MCF10A cells with or without alpelisib treatment. We observed a depletion of the spliced regions of AFMID, CCND3, MTUS1, SRP68, RRAS2, UNKL, PTK2/FAK, and SRPK2 as well as an enrichment of the spliced regions of HNRNPC and IP6K2 isoforms in mutant cells compared with parental cells, that were reversed by PI3K inhibition, which was consistent with our computational analyses (Figs. 1G and 2A). As a control, the levels of selected unspliced regions for each transcript remained constant. We next investigated the alternative splicing events in MCF7 (PIK3CA^E545K), MDA-MB-453 (PIK3CA^H1047R) and T47D (PIK3CA^H1047R), and in vivo using MCF7 and T47D-derived xenografts and a PIK3CA^H1047R patient-derived xenograft (PDX) model treated with alpelisib or vehicle daily for more than 15 days (Supplementary Fig. S2A and S2B). In three PIK3CA-mutated breast cancer cell lines, the alternative splicing events of AFMID, CCND3, HNRNPC, PTK2, RRAS2, SRPK2, and UNKL were rescued upon alpelisib treatment in at least two models (Fig. 2B; Supplementary Fig. S2C). An in vivo validation of the candidate splicing patterns in MCF7 or T47D-derived xenograft tumors and in the PDX models demonstrated that the subsequent splicing changes of a number of candidates were also observed in at least two models upon alpelisib treatment (Supplementary Fig. S2B and S2C). Taken together, these data indicate that the PI3K signaling pathway is a crucial regulator of RNA splicing.

To single out the component of the PI3K pathway responsible for the validated splicing changes, we treated parental and PIK3CA^H1047R mutant MCF10A cells with a panel of inhibitors targeting PI3K, AKT or mTORC1. Interestingly, the splicing of PTK2/FAK, AFMID, CCND3, and RRAS2 were not rescued by mTORC1 inhibition; however, SRPK2 and IP6K2 splicing were rescued by all PI3K/AKT/mTORC1 inhibitors tested (Fig. 2C; Supplementary Fig. S2D). This suggests a gene-specific regulation of RNA splicing by distinct members of the PI3K/AKT/mTORC1 signaling cascade.

To investigate the spliceosomal factors responsible for the observed splicing changes due to PIK3CA activating mutations, we performed RNA-binding protein (RBP) motif analysis of the alternatively spliced regions validated across different breast models using RBP-map (34). Motifs within the SRSF family (SRSF1, SRSF2, SRSF3, SRSF7, TRA2B/SRSF10) CELF1, HNRNPC, and the splicing repressor PTBP1 (Fig. 2D; Supplementary Fig. S3A; Supplementary Tables S8 and S9) were significantly enriched (P value < 0.05) at the selected top PI3K-dependent splice sites.

To identify the RBPs responsible for the RNA splicing of the selected candidates, MCF10A PIK3CA^H1047R cells were transfected with siRNA targeting the RBPs found in the motif analysis and, 72 hours later, treated with alpelisib for 4 hours. The effective downregulation of each RBP was checked via qRT-PCR (Supplementary Fig. S3B). We considered only the RBPs whose knockdown reversed the splicing changes of the candidate tested (PTK2, AFMID, SRPK2, RRAS2, HNRNPC, IP6K2), without significantly affecting the stability of each transcript as a validation. According to these criteria, the majority of the splicing events appeared to be coregulated by members of the SRSF family, while the knockdown of PTBP1 markedly reduced the stability of all the genes tested. (Fig. 2E; Supplementary Fig. S4). Interestingly, although both PTK2/FRNK and AFMID appeared to be regulated by AKT and not by mTOR (Fig. 2C), FRNK splicing was regulated by SRSF3 but not SRSF7, while AFMID splicing was regulated by SRSF7 but not SRSF3, indicating a gene-specific modulation of splicing by AKT.

To illustrate the potential functional consequences of the differential splicing events modulated by the PI3K pathway we focused on two candidates that were validated in preclinical models and in patient samples. Their alternative splicing was previously reported, albeit not in a PI3K-dependent manner. PTK2 or focal adhesion kinase FAK is an intracellular protein tyrosine kinase involved in cell motility, invasion, survival, proliferation, and epithelial–mesenchymal transition (35). AFMID is an arylformamidase, critical for the Tryptophan catabolism pathway and nicotinamide adenine dinucleotide (NAD⁺) cofactors biogenesis (36, 37).

Detailed analysis of the PTK2/FAK transcripts revealed that an intronic region of the canonical PTK2–240 transcript was significantly excluded in PIK3CA^H1047R but not parental cells and reverted by alpelisib treatment (Fig. 3A). This event can be annotated as an alternative transcription start site (ATSS) or an intron retention to produce the short transcript PTK2–203. PTK2–203 loss was validated in PIK3CA^H1047R cells and conversely, a subsequent gain was observed during PI3K or AKT inhibition (Fig. 2A and C). Transcript PTK2–203 encodes a known protein isoform named FRNK, a dominant negative regulator of PTK2/FAK itself (38, 39). While lacking FERM and Kinase domains, FRNK retains its FAT domain, (Fig. 3B) and thus the ability to localize at focal adhesion plaques. This localization is needed to interact with PTK2/FAK binding partners and thus disrupts PTK2/FAK activation (40).

The expression of both PTK2/FAK and FRNK isoforms were investigated by Western blot in parental and PIK3CA^H1047R MCF10A cells treated with a panel of PI3K pathway inhibitors (Fig. 3C). FRNK isoform is ablated in PIK3CA^H1047R cells compared with parental cells but induced by different PI3Ka inhibitors (alpelisib, GDC0077, and taselisib) and the AKT inhibitor ipatasertib, but not the mTORC1 inhibitor everolimus (Fig. 3C). For this reason, we stably transduced a doxycycline-inducible FRNK in mutant MCF10A cells to study their cell migration and proliferation in the absence or in the presence of FRNK. As a consequence, alpelisib treatment or overexpression of an inducible GFP-tagged FRNK led to decreased cell migration of mutant MCF10A cells, as tested by wound healing assay (Fig. 3D; Supplementary Fig. S5A–S5D) and cell proliferation (Fig. 3E). Of note, pAKT was reduced by FRNK overexpression, in agreement with previous observations showing that PTK2/FAK activates the PI3K pathway through phosphorylation of the regulatory subunit p85 (Fig. 3C; ref. 41). These data demonstrated that PI3K/AKT activation abrogated the negative regulation of PTK2 via the regulation of alternative RNA splicing and FRNK isoform switching, thus unleashing PTK2-mediated downstream signaling. Accordingly, treatment with PI3K/AKT inhibitors restored FRNK, which attenuated PTK2 function in cell growth and migration.

Analysis of the AFMID locus revealed that the exons 5–9 belonging to AFMID-202 transcript were excluded in MCF10A mutant cells but significantly included upon treatment with alpelisib (Fig. 3F). AFMID-202 is translated into the main protein product (isoform a, 303 aa), following the proper inclusion of exon 5–9 (Fig. 3G; ref. 42). Interestingly, this alternative spliced cassette encodes for the kynurenine formamidase (KFase) domain, containing a HGGWX motif (exon 5), an alpha/beta hydrolase domain (exons 4–9) and the catalytic active residues S162, D247 and H279, critical for the catalytic activity of AFMID (Fig. 3G; ref. 43). Given that kynurenine accumulation inhibits cell proliferation (44), we hypothesized that the inclusion of exon 5–9 splicing cassette in the mature AFMID-202 mRNA upon PI3Kα inhibition would affect cell proliferation of PIK3CA-mutant cell lines. We transduced MCF10A PIK3CA^H1047R, MCF7, and T47D cells with an inducible AFMID-202 vector (Fig. 3H; Supplementary Fig. S5E and S5F) and observed reduced cell proliferation upon the overexpression of AFMID-202. In addition, upon alpelisib treatment, AFMID depletion by a pool of siRNAs in mutant MCF10A cells significantly rescues alpelisib-induced inhibition of proliferation (Supplementary Fig. S5G and S5H). Altogether, our data suggest that the PI3K pathway uses alternative splicing to promote the switch towards oncogenic RNA isoforms, governing growth, migration, and metabolism in breast cancer and that these changes can be partially reversed by treatment with PI3K inhibitors.

Given that the PI3K-regulated alternatively spliced isoforms affect cell proliferation of PIK3CA-mutant breast cancer cells, we questioned whether spliceosome inhibitors currently in clinical trials, H3B-8800 (NCT02841540) inhibitor of SF3b complex (45), or the splicing inhibitor E7070 (NCT00165880), that targets the pre-mRNA splicing factor RBM39 (46), would sensitize breast cancer cells to alpelisib. Interestingly, splicing modulators increased the sensitivity of MCF7 and T47D cells to alpelisib (Fig. 3I; Supplementary Fig. S5I), providing a rationale for spliceosome-targeted therapy in combination with alpelisib in ER⁺ breast cancer.

To expand our findings in the clinical setting, we performed RNA-seq (100–150 million reads) in 8 tumor samples isolated from patients with PIK3CA mutated breast cancer that were enrolled in a phase I clinical trial of the PI3Kα inhibitor GDC0077 (inavolisib). In addition, we reanalyzed RNA-seq data obtained from 10 tumors isolated from breast cancer patients treated with alpelisib (47). Paired longitudinal tumor biopsies were collected before PI3Kα inhibitor therapy and after 14 days of daily treatment, 2 to 4 hours after drug administration (Fig. 4A). A differential gene expression heatmap of PI3K mutant and PI3Kα inhibitor treated patient biopsies demonstrated separation of pretreatment and on-treatment breast cancers (n = 4967, FDR adjusted P value < 0.05; Fig. 4B; Supplementary Table S10). GSEA analysis identified metabolic pathways such as fatty acid metabolism, adipogenesis, and cholesterol homeostasis to be among the top pathways enriched in samples subjected to acute pharmacologic PI3K inhibition (Supplementary Table S11).

GSEA motif signature enrichment found TFs such as STATs, FOXOs, ETS, the nuclear hormone receptors PPARG, AR, ER, RORA, the TFs FOXA1 (HNF3ALPHA), and GATA among others, to be positively enriched upon PI3Kα inhibition (Supplementary Fig. S6A). Of note, the FOXO TF family is a well-established direct target of the PI3K pathway, and we observed that its activity and the expression of FOXO targets genes (PDK4, IGF2, LIPG, INSR, CDKN3, APOC3, among others) were upregulated upon treatment, consistent with pathway inhibition (Supplementary Table S10; Supplementary Fig. S6A). In contrast, TF motifs of AP1, E2Fs, OCT, and PAX were depleted on-treatment (Supplementary Fig. S6A). An examination of TF motifs identified in ATAC-seq of two ER⁺ paired breast cancer biopsies that had been collected before and during alpelisib treatment (47), also identified responsive elements for ER, FOXA1, nuclear receptor, STATs, AP1, GATA TF and others (Supplementary Table S12), providing additional evidence that the activity of these TFs may be regulated by the PI3K pathway. The overlap between the TF motifs enriched in MEFs, MCF10A cells, and breast cancer patient samples upon PI3Kα inhibitor treatment point to common regulatory factors in all 3 systems (Supplementary Fig. S6B).

In the breast cancer patient cohort, we found that the oncogenic PI3K pathway promoted >4,000 alternative splicing events between pre- and on-treatment samples (Fig. 4C; Supplementary Table S5). This translated to the most frequent isoform alterations as combined single and multiple exon skipping, ATSS, and alternative transcription termination site (ATTS), and in accordance with cell lines (Fig. 4D). A GO analyses of patient candidate isoform gene signatures found an enrichment in fatty acid and carbohydrate metabolism (Fig. 4E; Supplementary Table S13).

SUPPA (33) was used to profile for any splicing events or isoforms, which are found to be differentially expressed in each patient pair of the GDC0077 cohort. These analyses determined alternative splicing events were primarily enriched for alternative first (AF) and exon skipping (ES; Supplementary Fig. S6C). Interestingly, in the tumor samples from patient 5, harboring a weaker and uncommon PIK3CA mutation Q546K, and patient 8, identified as PIK3CA wild-type, we found substantially fewer significant alternative events (Supplementary Fig. S6C). RNA-seq from ten tumor biopsies in both pre- and on-treatment with alpelisib (47), contained dramatic splicing changes favoring AF and ES as the two most common events (Supplementary Fig. S6D).

Importantly, 21 of the candidate isoforms previously identified in MCF10A cells and validated in breast cancer lines and xenografts, exhibited similar RNA splicing patterns in the pooled patient samples upon PI3K inhibition. For example, the AFMID-202 and PTK2–203 isoforms, which negatively regulate cell proliferation and invasion, were significantly enriched in the patient cohorts treated with GDC0077 (Fig. 4F) or alpelisib (Supplementary Fig. S6E), respectively, demonstrating the clinical relevance of our preclinical cell line data. We also overlapped the same RBP motifs that were validated in MCF10A lines with the exons of the alternative candidates in the GDC0077 cohort. This RBP analysis demonstrated significant enrichment of both SRSFs and CELF1 motifs within the exons of alternatively spliced isoforms of the GDC0077 (inavolisib) patient cohort (Fig. 4G).

This study provides comprehensive transcriptomic analysis of alternative RNA splicing, gene expression, and TF enrichment upon PI3K pathway activation and inhibition in cell lines, xenografts, and samples from patients with breast cancer treated with PI3Kα inhibitors. We discovered the PI3K pathway regulated a unique transcriptome and enabled alternative splicing of genes that were involved in growth, migration, metabolism, and splicing itself.

In preclinical models, we observed common TF motifs that were inversely enriched upon activation and inhibition of the PI3K pathway. While some of these TFs were expected to be associated with the PI3K pathway, we also identified novel TFs such as PR, GR, RORA, ZEB1, AP1, YY1, and others. Importantly, a substantial number of these TFs were also found in the analyses of patients with breast cancer, in particular the nuclear hormone receptors family.

Increasing evidence also suggests that splicing alterations contribute to tumorigenesis and they are now considered a full-fledged hallmark of cancer (8). Our comprehensive analysis of alternative splicing upon PI3K pathway modulation uncovered widespread splicing changes of isoforms involved in metabolism and cell survival, whose dysregulation may contribute to tumorigenesis. The rescued expression of top splicing candidates PTK2 or AFMID alternatively spliced isoforms in PI3K-mutant breast models led to impaired cell migration and proliferation. This suggested that the PI3K pathway can contribute to tumorigenesis also through alternative mRNA splicing. Our work highlights another level of regulation of metabolism and proliferation by the PI3K pathway, through both transcription of metabolic factors and the splicing of genes involved in metabolism, such as AFMID, across preclinical models and patient samples.

Aberrant splicing of other candidates involved in cell growth such as IP6K2, RRAS2/TC21, and CCND3 may also contribute to PI3K-dependent tumorigenesis. Of note, RRAS2/TC21 activates the PI3K pathway promoting breast tumorigenesis and mammary gland development in a PI3K-dependent manner (48), thus participating in a feedback loop. Our unbiased analysis revealed the significant enrichment of RRAS-206 isoform in preclinical and clinical models upon PI3Kα inhibition. This isoform is predicted to lack the GTP-binding and switch domains crucial for the stability and function of RRAS2 (49), identifying an additional pathway by which the oncogenic PI3K affects cell proliferation. While our work suggests that the PI3K pathway can contribute to tumorigenesis also through alternative mRNA splicing, our findings argue for a systematic interrogation of the biological function of the top modulated isoforms. In addition, the fact that splicing modulators further sensitize breast cancer cells to alpelisib suggests that the PI3K-mediated regulation of splicing could potentially be therapeutically exploitable. Whether PI3K-mediated aberrant RNA splicing may predict for a robust biomarker-driven transcriptomic signature of PI3K activation and response to PI3K inhibitors remains to be determined.

Interestingly, the PI3K pathway mediates the splicing of genes that code for members of the splicing machinery itself, such as SRPK2, HNRNPC, RBM5, and others. This postulates a regulatory feedback loop whereby PI3K-mediated splicing controls splicing. Notably, we discovered that the PI3Kα-dependent alternative RNA splicing of HNRNPC is possibly self-regulated and this may provide an additional mode of regulation in RNA processing controlled by the PI3K-pathway.

We also uncovered a robust signature of RBPs that mediated PI3K-dependent RNA splicing in breast cancer cell lines. Specifically, we identified several members of the SRSF family or CELF1, to regulate PI3K-mediated splicing of candidates such as PTK2, AFMID, HNRNPC, RRAS2, and IP6K2. It has been shown that SRSF proteins (11–13) and CELF1 (50) can be regulated by AKT through direct phosphorylation. In addition, AKT interacts with SRPKs to induce autophosphorylation and nuclear translocation to activate SR proteins and affect alternative splicing (17). A fraction of our validated candidates were rescued by PI3K or AKT inhibitors but not by the mTORC1 inhibitor. As an example, PTK2/FRNK or AFMID transcripts appear to be regulated by AKT and not by mTOR, and FRNK splicing is regulated by SRSF3 but not SRSF7, while AFMID splicing is regulated by SRSF7, but not SRSF3. These differences cannot be only explained by the direct or indirect regulation of SR proteins mediated by AKT, but suggest a gene-specific modulation of alternative splicing that needs to be dissected for each of the candidates. Thus, further work is necessary to determine the exact mechanisms of how the PI3K pathway affects RBP function to regulate splicing of each candidate in breast cancer.

In patients with breast cancer treated with PI3Kα inhibitors, we uncovered at least 21 genes such as IP6K2, UNKL, AFMID, KDM2B, NMT2, BCL6, and others that displayed the same isoform alternatively spliced in MCF10A cells. It remains to be elucidated how distinct alterations in the PI3K pathway members such as AKT or PIK3CA mutations or loss of PTEN in breast and other malignancies contribute to control gene expression and RNA splicing. Our findings also argue for a systematic investigation of other frequently mutated oncoproteins that may alter splicing to promote tumorigenesis. It also remains to be discovered whether tumors addicted to oncogenic PI3K pathway may be more sensitive to spliceosome inhibitors. As alternative splicing is known to be tissue-specific (30), further investigation will be necessary to define roles of these splicing candidates in other cellular contexts.

Collectively, through a comprehensive transcriptomic analysis of preclinical and clinical models, our results highlight a key role for the PI3K pathway in regulating RNA splicing, uncovering new crosstalk into how the PI3K regulates proliferation and metabolism in breast cancer, which can be potentially therapeutically targetable with spliceosome inhibitors.

F. Michelini reports personal fees from AstraZeneca outside the submitted work. K. Jhaveri reports other support from Genentech, Novartis, Lilly Pharmaceuticals/Loxo Oncology, AstraZeneca, Pfizer, Zymeworks, Puma Biotechnology, Merck, Novita Pharmaceuticals, Debio Pharmaceuticals, ADC Therapeutics; personal fees from Novartis, Pfizer, AstraZeneca, Lilly Pharmaceutical/Loxo Oncology, Daiichi Sankyo, Seattle Genetics, Taiho Oncology, Jounce Therapeutics, Blueprint Medicines; and personal fees from Sun Pharma Advance Research Pvt. Ltd. outside the submitted work. P. Castel reports personal fees from Venthera outside the submitted work. L. Fairchild reports being a current employee at Novartis Institutes of BioMedical Research, but was affiliated with MSKCC during contribution to the manuscript. A. Arruabarrena-Aristorena reports other support from postdoctoral grant from the Education Department of the Basque Country Government in Spain during the conduct of the study. M. Sallaku reports personal fees from Loxo Oncology during the conduct of the study and personal fees from Loxo Oncology outside the submitted work. S. Chandarlapaty reports grants from BCRF, Daiichi-Sankyo, and grants from NIH/NCI during the conduct of the study; grants, personal fees, and nonfinancial support from Novartis; personal fees from Sanofi, Lilly, Inivata; grants and personal fees from Paige.ai, AstraZeneca; and grants from Ambryx outside the submitted work. O. Abdel-Wahab reports grants from Loxo Oncology, Lilly, H3 Biomedicine Inc., Nurix Therapeutics; other support from Aichemy Inc., Harmonic Discovery Inc.; and other support from Envisagenics Inc. outside the submitted work. M. Scaltriti reports other support from AstraZeneca and other support from AstraZeneca outside the submitted work. E. Toska reports grants from Jayne Koskinas Ted Giovanis Grant, NIH K22CA245487, NIH R21CA252530, Breast Cancer Alliance, and grants from Innovation to Cancer Informatics during the conduct of the study; grants and personal fees from AstraZeneca outside the submitted work. No disclosures were reported by the other authors.

E. Ladewig: Conceptualization, data curation, software, formal analysis, visualization, methodology, writing–review and editing. F. Michelini: Conceptualization, formal analysis, visualization, methodology, writing–original draft, writing–review and editing. K. Jhaveri: Resources, writing–review and editing. P. Castel: Conceptualization, writing–review and editing. J. Carmona: Investigation, writing–review and editing. L. Fairchild: Formal analysis. A.G. Zuniga: Investigation. A. Arruabarrena-Aristorena: Investigation, writing–review and editing. E. Cocco: Investigation, writing–review and editing. R. Blawski: Investigation, visualization. S. Kittane: Investigation, visualization. Y. Zhang: Investigation. M. Sallaku: Investigation. L. Baldino: Investigation. V. Hristidis: Investigation. S. Chandarlapaty: Conceptualization. O. Abdel-Wahab: Conceptualization, resources, writing–review and editing. C. Leslie: Conceptualization, resources, supervision, writing–review and editing. M. Scaltriti: Conceptualization, resources, supervision, funding acquisition, writing–original draft, writing–review and editing. E. Toska: Conceptualization, resources, supervision, validation, investigation, writing–original draft, project administration, writing–review and editing.

This work has been supported by NIH grants P30CA008748 and RO1CA190642–01A1, the Breast Cancer Research Foundation, the V Foundation and the National Science Foundation (to M. Scaltriti). It has also been supported by Innovation to Cancer Informatics to E. Toska and C. Leslie; and Jayne Koskinas Ted Giovanis grant, NCI K22CA245487, R21CA252530, Breast Cancer Alliance to E. Toska. E. Ladewig is supported by grant no. NCI K00CA212478. A.G. Zuniga is supported by NIH R25GM109441. The authors thank Wayne A. Phillips for providing the parental and Pik3ca H1047R knock-in MEFs. H3B-8800 was a gift from H3 Biomedicine to O. Abdel-Wahab.

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.