Cohort description and patients’ clinical-pathological characteristics

Between 2019 and 2023, 176 patients were enroled in the Side-Out initiative powered by Perthera. Patients were referred by NCI-designated cancer centres in academic institutions (NCI-CCC) (n = 65) as well as community hospitals and large non-NCI-CCC academic centres (n = 111). A CONSORT diagram describing the study population is provided in Fig. S2. Genomic profiles were generated for 171 patients enroled in the study, and phospho-proteomic data were collected by two independent laboratories, Neogenomics and Theralink Technologies, for 32 and 69 patients, respectively (Fig. 1a). An overview of the overall study design and approach can be found in Fig. 1b.

Fig. 1: Study workflow and genomic characterisation of 176 MBCs retrospectively identified from a real-world evidence database.

Summary of the molecular information collected from each of the 176 patients enroled in the study (a). Overall study design along with the molecular information available for each cohort and data processing pipeline (b). Packed bubble graph depicting the most frequent genomic alterations identified across the 176 patients in the Side-Out cohort (c); the dimension of each circle is proportional to the number of occurrences. Red circles indicate genes detected in at least 10% of the study population. Matrix illustrating frequency and co-occurrences of the ten most frequently mutated or amplified genes in the Side-Out cohort (d and e, respectively).

The median age of participants at diagnosis was 59 (range 28–85), and the cohort included 173 females and three males (Table 1). Race and ethnicity information were available for 119 patients, of which 96 were White (80.7%), 12 were Black (10.1%), seven were Asians (5.9%) and four were Hispanic (3.4%) (Table 1).

Table 1 Demographic and clinical characteristics of patients enroled in the study.

Of the 176 participants, 110 (62.5%) were affected by HR+/HER2- tumours, 48 (27.3%) by TNBCs, 13 (7.4%) by HER2+ tumours, including nine HR+/HER2+ (5.1%) and four by HER2-enriched (HR-/HER2+) (2.3%) lesions (Table 1).

Of the 176 patients, 65.3% (n = 115) had stage 4 metastatic disease at the time of diagnosis. The liver (n = 27, 25.0%) was the most frequent site of metastasis, and 23 patients (21.3%) had secondary lesions at more than one site (Table 1). The number of treatments before study participation ranges from one to six (Table 1).

The PIK3CA pathway mutational landscape in a real-world evidence cohort of metastatic breast cancer patients mimics those of well-characterised cohorts

We first assessed the frequencies of the most common genomic alterations in the Side-Out cohort of MBC patients (Fig. 1c; Fig. S3; Fig. S4; Table S1). For nine of 171 patients (5.3%) with genomic data, none of the genes measured by NGS harboured oncogenic alterations (Fig. 1d). As expected, TP53 (n = 75, 43.9%) and PIK3CA (n = 58, 33.9%) mutations were the most frequently identified pathogenic variants (Fig. 1d), followed by amplifications of FGF3 (n = 24, 14.0%), FGF4 (n = 23, 13.5%), CCND1 (n = 22, 12.9%), MYC (n = 21, 12.3%) and mutations of ESR1 (n = 21, 12.3%) and GATA3 (n = 19, 11.1%) (Fig. 1d, e). The complete genomic landscape of the entire cohort can be found in Fig. S3 and S4.

PIK3CA alterations were found in 43.5% of HR+/HER2- tumours (n = 47), 50% of HER2-enriched lesions (n = 2), 33.3% of HR+/HER2+ (n = 3) and 11.1% of TNBCs (n = 5) (Fig. 2a). The three most frequent PIK3CA gain-of-function pathogenic variants were the H1047R (n = 26, 15.2%) mutation of the kinase domain, and the E545K (n = 15, 8.8%) and E542K (n = 6, 3.5%) mutations of the helical domain (Fig. 2a). Within the same sample, multiple pathogenic alterations of the PIK3CA gene were found in one TNBC (2%) and 10 HR+/HER2- tumours (9%) (Fig. 2a). PIK3CA amplifications were found in ~2% (n = 3) of patients including two patients with HR+/HER2- MBCs and one patient with a TNBC tumour. PIK3CA amplifications were exclusively found in tumours that harboured PIK3CA gain-of-function mutations affecting the H1047R and E542K residues. Genetic alterations of the PIK3CA gene co-occurred with pathogenic alterations of ESR1 and TP53 in 7% (n = 12) and 10.5% (n = 18) of patients, respectively (Fig. 1d).

Fig. 2: Frequencies of oncogenic alterations of members of the PIK3CA pathway in the Side-Out study compared to two cohorts of MBC patients enroled in a single academic institution.

Matrix illustrating frequencies of mutations and amplifications across ten members of the PIK3CA pathway in the Side-Out cohort (a); tumours were subdivided based on the molecular subtypes. HR+/HER2- tumours are shown in the top panel (a). Cumulative number of cases with genetic alterations of members of the PIK3CA pathway by tumour subtype in the Side-Out cohort (b) and the Inserm cohort retrieved from the GENIE database (c). Matrix illustrating frequencies of mutations and amplifications across ten members of the PIK3CA pathway in the Inserm cohort (d); tumours were subdivided based on the molecular subtypes; HR+/HER2- tumours are shown in the top panel.

Given the primary role of the PI3K/AKT signalling pathway in breast cancer onset and progression, especially in HR+/HER2- tumours, we next extended our analysis to other genes encoding for members of this signalling axis. Of the 92 genes listed in the Pathcards database as members of the PIK3CA/AKT ‘SuperPath’ [23], oncogenic alternations of 12 genes, namely PIK3CA, AKT 1, 2 and 3, PTEN, MDM2, PIK3R1, RICTOR, RPTOR, FOXO1, FOXO3 and SGK1, were identified in our cohort of patients (Fig. 2a, b).

Oncogenic alterations of genes belonging to the PIK3CA/AKT ‘SuperPath’ were identified in 85 (49.7%) of the 171 patients analysed (Fig. 2a). Sixteen of the 171 patients (9.4%) had aberrations of more than one gene belonging to the PIK3CA pathway (Fig. 2a). Oncogenic alterations of AKT1, 2 and 3 were found in 16 patients (9.4%), including nine AKT1 E17K point mutations, five AKT3 and two AKT2 amplifications (Figs. 1d, 2a). Of interest, PIK3CA oncogenic alterations were found in five of the seven patients with amplified AKT2 or AKT3. Loss-of-function alterations of the tumour suppressor PTEN were found in 15 patients (8.8%) and co-occurred with PIK3CA mutations in five patients. Additional alterations of members of the PIK3CA/AKT pathway included mutations of MDM2 in five patients (2.9%) and of PIK3R1 and RICTOR in two (1.2%) patients. Oncogenic mutations of FOXO1, FOXO3, RPTOR and SGK1 were found in one patient each across the entire cohort (0.6%, respectively) (Fig. 2a).

We next compared mutation rates of members of the PIK3CA pathway and their distribution in the Side-Out cohort against two publicly available MBC datasets retrieved from the GENIE database, the Inserm and the MSK cohort (Fig. 1b). While the MSK cohort only included HR+/HER2- tumours (n = 1116), biospecimens in the Inserm cohort (n = 216) contained all major breast cancer subtypes including 143 (66.2%) HR+/HER2-, 51 (23.6%) TNBC and 14 (6.5%) HER2-enriched MBCs. The molecular subtype of the remaining eight tumours was unknown.

PIK3CA pathogenic alterations were found in 31.5% (n = 68) of patients in the Inserm cohort and 40.1% (n = 448) in the MSK cohort (p = 0.03) (Fig. 2c, d, Fig. S5, Table 2). When the analysis was restricted to HR+/HER2- tumours, no differences were detected across the three cohorts (p > 0.05; Table 2).

Table 2 Pathogenic alterations of the PIK3CA pathway in the Side-Out, Inserm and MSK cohorts.

In the Inserm cohort, PIK3CA oncogenic alterations were detected in 55 (38.5%) HR+/HER2- tumours, 6 (11.8%) TNBC, 3 (21.4%) HER2-enriched lesions, four tumours with unknown subtype (Fig. 2c, d). Amplification of the PIK3CA gene was observed in 5 HR+/HER2- tumours (3.5%) and one TNBC patient (2%). When the analysis was extended to the other members of the PIK3CA/AKT ‘SuperPath,’ members of the PIK3CA pathway were found to be altered in 127 patients (58.8%) in the Inserm cohort. Oncogenic alterations of more than one gene were found in 45 biospecimens (Fig. 2d). AKT alterations were detected in 34 (15.7%) patients. PTEN mutations were present in 14 (6.5%) cases (Fig. 2d). Mutations of PIK3R1 were found in 16 (7.4%) individuals and of RICTOR and RPTOR in 13 (6%) and 12 (5.6%) patients, respectively (Fig. 2d). Pathogenic alterations of the transcription factors FOXO1 and FOXO3 were found in four (1.9%) and eight (3.7%) patients, respectively.

In the MSK cohort, oncogenic aberrations of the PIK3CA pathway were detected in 663 patients (59.4%), of which 148 had alterations of multiple genes. AKT was altered in 131 (11.7%) patients and PTEN in 95 (8.5%) patients (Fig. S3). Mutations of PIK3R1 were found in 26 (2.3%) patients and of RICTOR (2.2%) and RPTOR (2.2%) in 25 patients, respectively. Frequencies of genomic alterations of the main members of the PIK3CA pathway were at large similar among the three cohorts, with the exception of a few genes that were relatively infrequent across study sets and of PIK3CA when the analysis was not restricted to the HR+/HER2- tumours, which is the only subtype represented in the MSK cohort (Table 2).

AKT functional activation in metastatic breast cancers cannot be solely predicted by the underlying genomic profile

Next, we assessed whether the underlying genomic profile could predict the activation of critical nodes on the PI3K/AKT/mTOR signalling axis. We used two orthogonal proteomic methods to determine the association between genomic alterations and activation of the PIK3CA/AKT/mTOR pathway with standard IHC and RPPA.

We first assessed whether AKT (pAKT) activation was more frequently detected in tumours harbouring genetic alterations of the PIK3CA/AKT/mTOR axis compared to wild-type lesions using the IHC data. Of the 32 samples analysed, 25 had high activation of pAKT. The proportion of patients with high pAKT activity was similar in tumours harbouring pathogenic alterations of any member of the PIK3CA pathway compared to wild-type tumours (79% versus 77%, respectively; p > 0.05) (Fig. 3a). Approximately three quarters (76.9%) of PIK3CA wild-type tumours had pAKT levels that mimicked those of mutant lesions suggesting that AKT activity in MBCs is driven by genomic-dependent and independent events (Fig. 3b). Similar associations were also found when the analysis was restricted to the aberration of PIK3CA, AKT and PTEN, individually (Fig. 3a).

Fig. 3: Functional activation of the AKT-mTOR signalling axis in MBCs based on underlying oncogenic alterations of genes encoding for members of the PIK3CA pathway.

Mosaic plots showing correlations between mutational status and phosphorylated levels of AKT (pAKT) measured by IHC and classified on a dichotomous scale (high versus low) (a). The proportion of patients with high AKT activity was first compared between wild-type and PIK3CA, AKT and PTEN mutant/amplified tumours. The analysis was then extended to compare AKT activity in wild-type tumours and lesions with any genetic alterations of the PIK3CA pathway (PIK3CA pathway alterations). Tile plots summarising frequencies of genetic alterations of the PIK3CA pathway along with phosphorylated AKT levels measured by IHC (b). Unsupervised hierarchical clustering Ward’s method assessing activation of six members of the PI3K/AKT/mTOR signalling axis in wild-type tumours and lesions harbouring oncogenic alterations of the PIK3CA pathway (c). Functional protein activation was measured on a continuous scale using RPPA percentile scores. Tumours’ molecular subtypes are listed and colour-coded on the x-axis. On the y-axis, samples with underlying genomic alterations are shown in black. RPPA continuous data are shown on a blue (low activation) to red (high activation) scale. Violin plots comparing activation of the signalling molecules in samples harbouring alterations of any gene of the PIK3CA axis and wild-type tumours (d); sample median is shown for each plot and asterisks denote comparisons that were statistically different (p < 0.01). The receiver operating characteristic (ROC) curve shows the performance of mutations of the PIK3CA pathway as potential classifiers for predicting AKT phosphorylation levels in MBCs along with the corresponding area under the curve (AUC) (e). Violin plots comparing activation levels of signalling molecules across the three most frequently detected PIK3CA oncogenic mutations (f).

We next compared activation levels of six proteins involved in the PI3K/AKT/mTOR signalling axis measured by RPPA on a continuous scale; RPPA values were calculated as percentile scores of the reference population [37]. We first used unsupervised hierarchical clustering to examine how key proteins belonging to the PI3K/AKT/mTOR signalling axis were distributed in 64 samples for which RPPA and NGS data were available (Fig. 3c). At large, subtypes were not associated with AKT/mTOR activation (Fig. 3c, Fig. S6). Samples harbouring mutations of genes encoding for proteins belonging to the PI3K/AKT/mTOR axis were spread across three clusters and had heterogeneous activation of the signalling molecules. Cluster 1 was characterised by increased activation of mTOR and its downstream signalling molecules. Of the 29 patients in this cluster, 15 had genomic alterations of target genes. Cluster 2 was characterised by higher activation levels of AKT and its downstream targets and contained ten samples, including three of the eight tumours harbouring PTEN loss of function mutations. Samples included in Cluster 3 had an overall low activation of the different signalling molecules regardless of the mutational status.

When tumours with any mutation of the PIK3CA pathway were compared to wild-type lesions, activation of AKT (T308) and (S473) was significantly higher in the mutant population (p < 0.01). However, a significant degree of overlap in the samples’ distribution was detected between the two groups (Fig. 3d). ROC analysis assessing the ability of the mutational status to predict AKT activation suggested that pathogenic alterations of the PIK3CA pathway are poor predictors of protein activity (AUC: 0.69; Fig. 3e). When the analysis was restricted to PIK3CA, AKT and PTEN individually, activation of the signalling molecules was not different between wild-type and mutant tumours (p > 0.05; data not shown). Likewise, when a pathway score was created using the RPPA data as previously described [18], activation of the signalling molecules by RPPA did not differ between the mutant and wild-type populations (data not shown). Similar trends were also observed when the analysis was restricted to the HR+/HER2- population (Fig. 4a, b). Lastly, activation of signalling molecules did not differ between tumours harbouring different PIK3CA pathogenic mutations (Fig. 3f) nor in tumours harbouring multiple pathogenic alterations of genes belonging to the PIK3CA pathway (Fig. S7).

Fig. 4: Activation of the AKT-mTOR signalling axis in HR+/HER2- MBCs and their implications for predicting survival to first-line treatment with a CDK4/6 inhibitor in combination with endocrine treatment.

Violin plots comparing activation levels, measured as RPPA percentile scores, of signalling molecules in HR+/HER2- MBCs harbouring alterations of any gene of the PIK3CA axis (a); sample median is shown for each plot, and asterisks denote comparisons that were statistically different (p < 0.02). The receiver operating characteristic (ROC) curve shows the performance of mutations of the PIK3CA pathway as potential classifiers for predicting AKT phosphorylation levels in HR+/HER2- MBCs along with the corresponding area under the curve (AUC) (b). Heat map capturing activation levels of AKT (S473) and (T308) across 29 cell lines publicly available in the DepMap database. HR+/Her2- cells are shown in blue (c). Heat map showing PIK3CA dependency in 38 cell lines that underwent CRISPR and Ribonucleic acid interference (RNAi) screening (d). Unsupervised hierarchical clustering using Ward’s method assessing activation of the PI3K/AKT/mTOR signalling axis in the MCF-7, T47D and T47D abemaciclib-resistant cells; functional protein activation was measured on a continuous scale using RPPA values (e). Cell viability line plot of MCF-7 and T47D cells treated with capivasertib (range from 0.01–10 μM) along with box plot depicting the RPPA continuous values of phospho-AKT (T308) and phospho-4EBP1 (S65) and (T70) (p < 0.01). Median, highest and lowest values of experimental replicates are shown; asterisks denote comparisons that are statistically significant (f). Cell viability line plot of T47D parental and abemaciclib-resistant cells treated with buparlisib (range from 0.007 to 1 μM) along with box plot depicting the RPPA continuous values of phospho-AKT (T308) and phospho-p70S6 Kinase (T389) (p < 0.01 and 0.05, respectively). Medians, highest and lowest values of experimental replicates are shown; the asterisks denote comparisons that are statistically significant (g). Kaplan-Meier plot along with hazard ratio for progression-free survival in days and 95% confidence interval for patients with genomic alterations of any members of the PIK3CA pathway (HR: 0.58; CI: 0.14–2.36), PIK3CA (HR:0.67; CI 0.19–2.38), and phospho-p70 S6 Kinase (T389) activity (HR: 4.18; CI: 1.19–14.63) (h). P70S6 kinase activity levels were classified on a binary scale (high/low) based on the population median of the continuous RPPA data. Diagram showing a workflow for integrating multi-omic-based profiling for allocating patients to targeted treatments against members of the PI3K/AKT/mTOR pathway (i).

We next assessed whether PIK3CA mutations predicted AKT activation in 29 breast cancer cell lines retreated from the publicly available DepMap portal. At large, activation of AKT was not higher in cell lines harbouring a PIK3CA oncogenic mutation (Fig. 4c). We next looked at the significance of PIK3CA in relation to the lethality of its deletion or knockdown on a target across all 38 cell lines included in the dataset. Negative dependence scores suggest cell lines are heavily dependent on that gene. In line with our protein data, only two cell lines of the PIK3CA mutant cell lines, MDAMB361 and EFM19, showed dependency from the PI3K (Fig. 4d). To confirm the role of signalling molecules in predicting response to agents targeting the PI3K/AKT/mTOR pathway, we next compared the sensitivity to capivasertib in two HR+/HER2- breast cancer cell lines, namely T47D and MCF-7, harbouring hotspot mutations of PIK3CA (H1047R and E545K, respectively) [38] with diverse levels of activation of the PI3K/AKT/mTOR axis (Fig. 4e). Compared to the MCF-7, T47D cells had lower IC50 values (2.1 vs. 7.5 μM) and higher levels of phosphorylated AKT (T308) and 4EBP1 (S65) and (T70) (Fig. 4f). Given the increased levels of AKT activity in the T47D cell lines compared to the MCF-7 cells, we next assessed responses to the PI3K inhibitor buparlisib in parental and isogenic abemaciclib-resistant cells. As expected, resistant cells were more sensitive to buparlisib and increased sensitivity was associated with higher levels of phosphorylation of AKT (T308) as well as increased activation of the downstream substrate p70S6 Kinase (Fig. 4g).

Taken together, these data suggest that inferring activation of signalling molecules merely by the underlying genomic alterations of the PIK3CA pathway underestimates the activation frequencies of the PI3K/AKT/mTOR axis in MBCs. As AKT and PI3K inhibitors are currently being used in the clinical setting, capturing activity levels of signalling molecules belonging to this axis may help identify MBCs driven by the PI3K/AKT/mTOR signalling pathway and thus may benefit from these treatments.

Functional activation of AKT downstream substrates is associated with response to first-line treatment in HR+/HER2- metastatic breast cancers

As inhibition of AKT and PI3K activity has recently been shown to delay resistance to ET in combination with a CDK4/6 inhibitor in MBCs in preclinical studies [5, 19], we next assessed whether baseline activation levels of members of the PI3K/AKT/mTOR pathway are associated with response to first-line treatment with a CDK4/6 inhibitor in the Side-Out cohort. Of the 69 patients for which RPPA data were collected, 20 were treated with a CDK4/6 inhibitor in combination with ET in first-line. NGS data were available for 18 of the 20 patients included in this sub-analysis (Table S3). As expected, the mutation status of the PIK3CA pathway or PIK3CA was not associated with progression-free survival (PFS) (p > 0.05) (Fig. 4h). However, patients with p70S6 Kinase (T389) levels above the whole population median had shorter PFS compared to those with low p70S6 Kinase (T389) activity (HR: 4.18, 95% CI:1.19–14.63, p = 0.02; Fig. 4h).

Taken together, our data suggest that genomic alterations of the PIK3CA pathway are insufficient to predict protein activation in clinical samples and patients’ response to standard of care.