Targeting a Lineage-Specific PI3Kɣ/AKT Signaling Module in Acute Myeloid Leukemia Using a Heterobifunctional Degrader Molecule

Abstract

Dose-limiting toxicity poses a major limitation to the clinical utility of targeted cancer therapies, often arising from target engagement in non-malignant tissues. This obstacle can be minimized by targeting cancer dependencies driven by proteins with tissue- and/or tumor-restricted expression. In line with another recent report, we show here that in acute myeloid leukemia (AML), suppression of the myeloid-restricted PIK3CG/p110𝛾-PIK3R5/p101 axis inhibits AKT signaling and compromises AML cell fitness. Further, silencing PIK3CG/p110𝛾 or PIK3R5/p101 sensitizes AML cells to established AML therapies. Importantly, we find that existing small-molecule inhibitors against PIK3CG are insufficient to achieve a sustained long-term anti-leukemic effect. To address this concern, we developed a PROteolysis-TArgeting Chimera (PROTAC) heterobifunctional molecule that specifically degrades PIK3CG and potently suppresses AML progression alone and in combination with venetoclax in human AML cell lines, primary AML patient samples, and syngeneic mouse models.

INTRODUCTION

Acute myeloid leukemia (AML) is a hematologic malignancy characterized by clonal proliferation of abnormal myeloid progenitors in the bone marrow and peripheral blood. The five-year overall survival rate for patients diagnosed with de novo AML is 30%, with rates that exceed 50% in younger patients but are below 10% in patients diagnosed after the age of 60¹. This disparity is partially driven by the inability of older patients to tolerate highly toxic chemotherapies traditionally used as first-line induction regimens². Recent translational developments have led to the approval, since 2017 by the United States Food and Drug Administration (FDA), of more than 10 new drugs for treating AML, several of which are first-in-class therapies designed against specific, aberrantly-activated pathways¹. For example, AML driven by mutations in FLT3, IDH1, or IDH2 can now be treated with small-molecule inhibitors targeting the mutated enzyme, abrogating downstream leukemogenic signaling³. Elderly patients ineligible for standard 7+3 cytarabine- and daunorubicin-based chemotherapy are treated with the BCL2 inhibitor, venetoclax, combined with azacitidine or decitabine, which has improved overall survival^4,5. Unfortunately, most patients diagnosed with AML lack actionable mutations, underscoring the need to identify additional therapeutic targets, ideally those agnostic to clinical subtype or mutational profile, and potentially inclusive of those entrenched deep within the cancer’s identity.

A cancer’s tissue-of-origin templates key features of its biology; certain lineage-specific programs may be co-opted in neoplastic precursor cells to support tumorigenesis and tumor progression⁶. Multiple successful anticancer therapies function by targeting lineage-specific survival factors^7–10. This approach differs from most traditional targeted therapies, which engage ubiquitously expressed targets and are thus vulnerable to broad side effect profiles. Targeting nodes whose expression and/or function are unique to the cell lineage producing malignancy provides an opportunity to curb tumor survival while minimizing collateral damage in other tissues.

The effectiveness and tolerability of agents targeting lineage-restricted survival factors, approved as anticancer therapies, are exemplified in hormone-responsive cancers such as receptor-positive breast and prostate cancers^8,9. In hematological malignancies, the potential of lineage-directed therapies is evidenced in chronic lymphocytic leukemia (CLL), which has been transformed by the introduction of multiple agents targeting B-cell lineage dependencies⁷. For instance, inhibitors of Bruton’s tyrosine kinase (BTK), a B-cell restricted enzyme that relays essential cell survival and migration signals in CLL and other B-cell malignancies, induce durable remissions in CLL patients¹¹. Importantly, BTK inhibitors like ibrutinib have a narrow side effect profile consistent with BTK’s restricted expression, therefore permitting chronic use.

Phosphoinositide 3-kinase (PI3K) retains pleiotropic, essential functions and is a frequently altered driver of malignant progression. In particular, PIK3CA and PIK3CB, which encode catalytic isoforms of PI3K, are recurrently mutated and/or amplified in cancer, and PTEN, encoding a prominent negative regulator of the pathway, is frequently deleted¹². These alterations are prevalent in various tumor types, prompting extensive efforts to develop inhibitors of PI3K and its downstream effectors^13–15. Unfortunately, widespread PI3K expression across non-malignant tissues leads to the emergence of dose-limiting toxicities secondary to their inhibition in these tissues^16–18. Subsequent studies pinpointed the presence of specific PI3K isoforms in distinct malignant tissue types, suggesting that modifying the function of a pivotal enzyme like PI3K might be achievable in a tissue-selective manner. To this end, advances have been made in the development of inhibitors targeting the delta isoform of PI3K (encoded by PIK3CD), abundantly expressed in lymphoid cells. Clinical trials with PIK3CD inhibitors have demonstrated clinically-significant activity in patients with relapsed/refractory CLL, although improvements to their specificity may be required to further optimize their toxicity profile¹⁹. This example demonstrates targeting the essential enzyme, PI3K, in a tissue-specific manner based on isoform-dependency as a viable therapeutic strategy.

The specific isoform of PI3K, PIK3CG/p110𝛾 (also known as PI3Kγ), was previously described as a critical regulator of tumor immune evasion in myeloid cells, where it is the most prominently expressed isoform^20–22. In solid tumor studies, due to its immune function, PIK3CG was also reported as a regulator of innate immunity during inflammation and cancer. Consequently, PIK3CG inhibition using small-molecule inhibitors such as IPI-549 has been used in combination with anti-PD-1 checkpoint inhibition to drive therapeutic antitumor immunity^20,23. These findings have spurred the development of phase 1 clinical trials investigating the antitumor activity of IPI-549 in combination with PD-1/PD-L1 inhibitors²⁴.

In this study, we establish that PIK3CG/p110𝛾 and its exclusive cognate regulatory subunit, PIK3R5/p101, exhibit lineage-restricted expression and represent critical vulnerabilities in AML, consistent with the findings of a recently published study²⁵. Further, we confirmed that tissue-specific expression and dependency patterns of the PIK3CG/PIK3R5 heterodimer confer isoform-specific control of downstream AKT signaling and, when lost, increase sensitivity to venetoclax. Surprisingly, small molecules targeting PIK3CG recapitulate neither the proliferative deficits nor the inhibitory signaling observed with genetic suppression of PIK3CG and PIK3R5. This motivated the development of a PROteolysis TArgeting Chimera (PROTAC) degrader molecule, which demonstrates efficacious PIK3CG degradation and markedly impairs AML cell survival in vitro and in murine models.

RESULTS

PIK3CG/PIK3R5 Exhibit a Myeloid-Biased Expression Profile

To explore lineage-specific genetic dependencies in AML, we used publicly available primary patient data from The Cancer Genome Atlas (TCGA) and healthy donor tissue-derived data from the Genotype-Tissue Expression Project (GTEx) to compare gene expression in primary tumor and healthy samples across different tissue types. We observed upregulation of PIK3CG expression (encoding PIK3CG/p110𝛾) in the myeloid compartment relative to other tissue types (Figure 1A). This observation was confirmed by comparison of the GTEx data with the primary AML patient sample collection data from the BEAT-AML cohort²⁶ (Extended Data Fig. 1A). By contrast, PIK3CA, PIK3CB, and PIK3CD, encoding the PI3K isoforms PIK3CA/p110𝛼, PIK3CB/p110𝛽 and PIK3CD/p110δ, respectively, displayed widespread expression across many tissue types, confirming in agreement with previous studies, a specific role for PIK3CG in the myeloid compartment^20,22 (Figure 1A). The restricted expression of PIK3CG is complemented by the expression of its exclusive, cognate regulatory subunit, PIK3R5 (encoding PIK3R5/p101), which exhibits similarly restricted expression in myeloid normal tissues and upregulation in AML. Within the BEAT-AML cohort, PIK3CG and PIK3R5 were significantly upregulated in patients subclassified in the French-American-British (FAB) monocytic 4 and 5 categories compared to other FAB subgroups (Extended Data Fig. 1B). No significant difference was observed in the expression of PIK3CG or PIK3R5 among patients with NPM1-mutated AMLs, those carrying the t(8,21) and MLL-fusion alterations, or between patients at diagnosis and relapse after chemotherapy (Extended Data Figs. 1C and 1D). Only PIK3R5 expression exhibited a significant increase in patients harboring the inv(16) alteration. A deeper correlative analysis into leukemia cell states revealed that PIK3CG and PIK3R5 expression is positively correlated with expression of monocytic- and dendritic cell-like transcriptomic features in AML cells²⁷, and negatively correlated with those related to the progenitor-like or the LSC17 gene signatures²⁸ (Extended Data Figs. 1E and F).

Figure 1: AML Cells Exhibit a Selective Dependency on PIK3CG and its Regulatory Subunit PIK3R5.

A. Expression of catalytic and regulatory Class I PI3K isoforms in normal and tumor tissue samples. Primary tumor expression data from TCGA database; normal expression data from GeTex expression database. Data accessed through Gepia gene expression portal.

B. Heatmap of normalized read density of H3K27ac ChIP-seq of 66 AML patient samples and four CD34⁺ healthy donors’ samples. Bar graphs represent the sum of depicted region signal for each sample.

C. DepMap CRISPR/Cas9 dependency profiling data depicting essentiality of the four catalytic and regulatory isoforms of PI3K. Data are ordered by unsupervised hierarchical clustering of Min-Max normalized averaged dependencies across cell lines in a given cancer type. Only cancer types with > 5 cell lines were included in analysis.

D. Cell growth of OCI-AML2 cells following shRNA-mediated depletion of PIK3CA, PIK3CB, PIK3CD, or PIK3CG. Data were normalized to values obtained from the corresponding – doxycycline condition. Error bars represent mean ± SD of three biological replicates after three days of seeding. P-values calculated using two-tailed Welch t test.

E. PIK3CG levels by western blot in OCI-AML2 cells transduced with a non-targeting control and two PIK3CG-directed sgRNAs. ACTIN used as a loading control.

F. Cell growth over time of indicated AML cell lines transduced with either a non-targeting control or two PIK3CG-directed sgRNAs. Error bars represent mean ± SD of three biological replicates composed of seven technical repeats.

G. PIK3R5 levels by western blot in OCI-AML2 cells transduced with a non-targeting control and two PIK3R5-directed sgRNAs. ACTIN used as a loading control.

H. Cell growth over time of indicated AML cell lines transduced with either a non-targeting control or two PIK3CG-directed sgRNAs. Error bars represent mean ± SD of three biological replicates composed of seven technical repeats.

I. Colony formation from indicated AML cell lines infected with either a control, or two sgRNAs directed against PIK3CG or PIK3R5. Error bars represent mean ± SD of three biological replicates composed of five technical repeats.

J. Bioluminescence signal detected by in vivo IVIS imaging of OCI-AML2 cells infected with a control or a PIK3CG-directed sgRNA and injected into NSG mice (n=10 mice / group) across multiple time points (days 7, 13, 18, 22). Error bars represent mean ± SD. P-values calculated using two-way ANOVA.

F, H, and I. P-values calculated using one-way ANOVA and reported on the figure panels. Different shapes of dots indicate distinct biological replicates.

E and G. Experiments were performed at least twice with similar results.

We then mined publicly-available ChIP-seq data to establish the binding pattern of the H3K27ac histone mark, associated with active transcription, at the promoter and gene body regions of PIK3CG and PIK3R5 in primary samples from AML patients and cord-blood-derived CD34⁺ cells from healthy donors (Figure 1B). Several genomic regions in PIK3CG and PIK3R5 were marked by H3K27ac in a significant proportion of AML patients and healthy donors. Notably, the increased signal of H3K27ac across the PIK3R5 gene was more pronounced in many AML patient-derived primary cells compared to normal CD34⁺ cells. This suggests heightened PIK3R5 transcriptional activity in the malignant context, consistent with the substantial RNA upregulation of PIK3R5 observed in AML relative to its healthy counterpart (Figures 1A and 1B).

To explore whether the increased expression of PIK3CG and PIK3R5 in myeloid cells is linked to a potential survival requirement in myeloid leukemias, we examined the DepMap dataset to establish the dependency profile of the catalytic and regulatory subunits of PI3K across cancer types. Of the four PI3K isoforms, PIK3CG was most essential in AML (and secondarily in ALL) with minimal essentiality in malignancies of non-hematopoietic origin (Figure 1C). By contrast, PIK3CA and PIK3CB were essential across malignancies of diverse tissue types. A selective dependency on PIK3R5 was also observed in AML and Chronic Myeloid Leukemia (CML), both related myeloid malignancies, while the regulatory subunits associated with PIK3CA, PIK3CB, and PIK3CD exhibited widespread essentiality across cancers of other tissues.

To validate the isoform-selective essentiality of PIK3CG in AML, we designed doxycycline (dox)-inducible short-hairpin RNAs (shRNAs) to knockdown PIK3CA, PIK3CB, PIK3CD, and PIK3CG. AML cell viability was reduced following shRNA-mediated depletion of PIK3CG but not other catalytic members of the PI3K family.(Figure 1D). We then introduced a doxycycline-inducible CRISPR-Cas9 single-guide RNA (sgRNA) system targeting the catalytic subunit of PIK3CG in various AML cell lines (Extended Data Fig. 2A). Knockout of PIK3CG resulted in significant growth impairment of AML cell lines, further evidencing a dependency on PIK3CG for the growth of AML cells in agreement with a recently published study²⁵ (Figures 1E and 1F). A similar CRISPR-Cas9 approach was used to knockout the expression of PIK3R5 and yielded a marked reduction in AML cell growth (Figures 1G–1H, and Extended Data Fig. 2A). We further evidenced that both PIK3CG and PIK3R5 gene suppression significantly dampened the colony-forming ability of multiple AML cell lines (Figure 1I). Finally, we assessed the bioluminescence of NOD-SCID gamma (NSG) mice injected with luciferase-expressing OCI-AML2 cells infected with either a non-targeting control (sgControl) or a PIK3CG-directed (sgPIK3CG) sgRNA. Mice transplanted with AML cells expressing sgPIK3CG demonstrated significantly lower disease burden over time compared to animals injected with sgControl OCI-AML2 cells, indicating that AML cell dependency on PIK3CG signaling is conserved in vivo (Figures 1J and Extended Data Fig. 2B).

PIK3CG/PIK3R5 Repression Potentiates the Effect of Venetoclax

We first correlated the expression of PIK3CG and PIK3R5 with response to 123 small-molecule inhibitors which were screened across primary AML samples as part of the BEAT-AML project²⁶ (Figure 2A). This led us to identify that patients with high PIK3CG or PIK3R5 expression exhibit a higher resistance to the BCL2 inhibitor, venetoclax, recently approved by the FDA for the treatment of elderly patients with AML. Further, decreased venetoclax sensitivity was predominantly found in patients classified in the FAB 4 and 5 subcategories, who also exhibit higher levels of PIK3CG and PIK3R5 expression (Figures 2B and Extended Data Fig. 1B). Therefore, we hypothesized that suppression of PIK3CG or PIK3R5 may potentiate the anti-leukemic effect of venetoclax.

Figure 2: PIK3CG and PIK3R5 Suppression Potentiates the Anti-Leukemia Effect of Venetolax.

A. Spearman correlation between PIK3CG and PIK3R5 expression and AUC responses to small- molecules screened as part of the BEAT-AML project²⁶.

B. Venetoclax sensitivity in patients from the BEAT-AML cohort who are part, or not part, of the FAB4 and 5 subgroups. Median with hinges at first and third quartiles, whiskers extend to 1.5 times the interquartile range, points indicate individual values. n values displayed on graph.

A and B. Two-sided p-values were corrected for multi-testing using the Benjamini & Hochberg and reported on the figure panels.

C. Cell sensitivity represented as half-maximal inhibitory concentration (IC50s, top) and Area Under the Curves (AUCs, bottom) of OCI-AML2 cells transduced with either a non-targeting control, two PIK3CG-directed, or two PIK3R5-directed sgRNAs and treated with increasing concentrations of the indicated targeted therapies or chemotherapy drugs. Error bars represent mean ± SD of two biological replicates composed of four technical repeats after three days of seeding.

D. Venetoclax sensitization effect (representative dose response curves, left panel, and AUCs, right panel) of two PIK3CG- and PIK3R5-directed sgRNAs in MOLM-14 and OCI-AML3 cell lines treated with increasing doses of venetoclax. Error bars represent mean ± SD of two biological replicates composed of five technical repeats after three days of seeding.

E.Colony formation from various AML cell lines infected with either a control, or two sgRNAs directed against PIK3CG or PIK3R5, and treated with 250nM, 5μM, or 150nM venetoclax in OCI-AML2, OCI-AML3, and MOLM-14, respectively. Error bars represent mean ± SD of three biological replicates composed of four technical repeats after seven days of seeding.

F. PIK3CA, PIK3CB and PIK3CD levels by western blot in OCI-AML2 cells transduced with a non-targeting control and two PIK3CG-directed sgRNAs. ACTIN used as a loading control. Experiment performed at least twice with similar results.

G. Representative growth inhibition curve (left panel) and AUCs (right panel) of OCI-AML2 cells infected with a non-targeting control and a PIK3CA-, PIK3CB-, or PIK3CD-directed sgRNA and treated with increasing venetoclax. Error bars represent mean ± SD of three biological replicates composed of seven technical repeats after three days of seeding.

C-E, G. P-values calculated using one-way ANOVA and reported on the figure panels. Different shapes of dots indicate distinct biological replicates.

H. Proportion of GFP-positive OCI-AML2 cells infected with a control or a PIK3CG-directed sgRNA and injected in NSG mice treated with either vehicle or 100mg/kg venetoclax (n=8 mice / group, mean ± SD). P-values calculated using two-tailed Mann-Whitney and reported on the figure panel.

I. Bioluminescence signal detected over multiple indicated time points of OCI-AML2 cells infected with a control or a PIK3CG-directed sgRNA and injected into NSG mice (n=8 mice / group, mean ± SD) and treated with either vehicle or 100mg/kg venetoclax. P-values calculated using two-way ANOVA and reported on the figure panel.

J. Overall survival of mice transplanted with either OCI-AML2 cells infected with a non-targeting or a PIK3CG-direct sgRNA alone or in combination with 100 mg/kg venetoclax daily from day 25 (n=10 mice/ group). Statistical significance by log-rank (Mantel-Cox) test and reported on the figure panel.

C-E, G. Different shapes of dots indicate distinct biological replicates.

We thus examined the effect of PIK3CG and PIK3R5 knockouts on AML cell response to venetoclax and to front-line chemotherapies, such as azacitidine, cytarabine, daunorubicin, in addition to FLT3 and KIT inhibitors. Suppression of PIK3CG and PIK3R5 consistently and significantly decreased the area under curve (AUC) and half-maximal inhibitory concentration (IC50) of venetoclax, but had no major effect on the other small-molecules tested, including FLT3 inhibitors; gilteritinib and sorafenib, and KIT inhibitors; amuvanib and telatinib (Figures 2C, Extended Data Figs. 3A and 3B). The potentiating effect of PIK3CG or PIK3R5 loss on venetoclax sensitivity was also validated in two additional AML cell lines, MOLM-14 and OCI-AML3 (Figure 2D). Venetoclax combined with PIK3CG- or PIK3R5-targeting sgRNAs also reduced the colony-forming capacity more than either perturbation alone (Figure 2E). In contrast, no further sensitization of AML cells to venetoclax was observed upon knockout of PIK3CA, PIK3CB, and PIK3CD (Figures 2F and 2G).

Finally, we injected a pool of luciferase-expressing GFP/Cas9-positive OCI-AML2 cells infected with either a non-targeting control or a PIK3CG-directed sgRNA into NSG mice (Figures 2H and 2I). We observed a profound decrease in the proportion of GFP-positive leukemic cells in the bone marrow of mice transplanted with PIK3CG knockout cells along with a significant sensitization to venetoclax treatment, compared to mice treated with venetoclax alone (Figure 2H). Mice injected with PIK3CG knockout cells exhibited lower overall disease burden compared to recipient animals injected with PIK3CG wild-type cells, an effect potentiated with venetoclax treatment (Figures 2I and Extended Data Fig. 3C). Importantly, animals injected with PIK3CG-silenced AML cells exhibited increased overall survival, enhanced by addition of venetoclax, confirming that specific targeting the PIK3CG-PIK3R5 signaling axis can potentiate the effect of venetoclax in vivo (Figure 2J).

AKT Signaling is Dependent on PIK3CG/PIK3R5 in AML

To characterize the PIK3CG-PIK3R5 signaling network, we performed tandem purification and mass spectrometry-based interactomic profiling of PIK3R5 in AML cell lines: MV4–11, OCI-AML2, and NOMO-1. We identified 170 top-scoring PIK3R5 interactors based on significant enrichment compared to the control (Extended Data Fig. 4A) and specificity of peptides on identification only into PIK3R5 conditions (Extended Data Fig. 4B) in at least 2/3 cell lines profiled (Extended Data Fig. 4C). We then constructed a STRINGdb-based physical interaction network (score > 0.5), categorized into distinct nodes which displayed pertinent sets of PIK3R5 interactors (Figure 3A).

Figure 3: PIK3CG and PIK3R5 Selectively Control AKT Signaling in AML.

A. Network of the top PIK3R5-interacting proteins in at least two AML cell lines from a tandem affinity purification of PIK3R5/p101. Known physical interaction from StringDB (score > 0.5).

B. Correlation between gene dependency scores of AKT1 and AKT2 versus PIK3CG in a panel of 15 AML cell lines. Data from DepMap CRISPR/Cas9 dependency profiling. Pearson correlation coefficient (ρ) and two-tailed p-value are reported on the figure panel.

C. Top MK-2206 sensitizer and resistor candidate genes identified from a pooled drug-modifier screen conducted in OCI-AML2 cells. Gene-level scores were obtained by averaging sgRNA-level comparisons. Genes with a log₂ depletion of < - 0.5 or enrichment of > 0.5 designated in orange and blue dots, respectively. Red dot denotes BCL2. Non-targeting control genes depicted as grey dots.

D. Expression and phosphorylation levels of PIK3CG, PIK3R5, AKT, and AKT substrates by western blot in OCI-AML2 and OCI-AML3 cells transduced with a non-targeting control, and one representative PIK3CG- and PIK3R5-directed sgRNA.

E. AKT1 and AKT2 phosphorylation levels by western blot in OCI-AML2 cells following knockout of PIK3CG and PIK3R5.

F. Indicated pro- and anti-apoptotic protein levels by western blot following knockout of PIK3CG and PIK3R5 using one representative sgRNA in OCI-AML2 cells.

G. Catalytic PI3K subunit levels and phosphorylation of AKT by western blot following knockdown of indicated catalytic PIK3 subunits in OCI-AML2 cells. ACTIN used as a loading control.

H. Fold-change in mRNA expression levels of each regulatory PI3K subunit following their knockdown. Error bars represent mean ± SD of three biological replicates. P-values calculated using two-tailed Welch t test and reported on the figure panel.

I. AKT phosphorylation levels by western blot following knockdown of regulatory subunits of AKT in OCI-AML2 cells.

J. Phosphorylation levels of AKT and a downstream substrate, TSC2, by western blot following ectopic expression of wild-type PIK3CG in OCI-AML2 cells infected with an shRNA targeting the 3’ UTR region of PIK3CG.

K. Schematic of the PI3K signaling downstream G-coupled protein receptors (GPCR) and tyrosine kinase receptors (RTK). CXCL12 activates GPCRs whereas pertussis toxin blocks their activity.

L. AKT and TSC2 phosphorylation levels by western blot in indicated human AML cell lines treated with 100ng/mL pertussis toxin Ptx for 24 hours.

M. Phosphorylation levels of AKT and its downstream substrate PRAS40 by western blot in control or PIK3CG-depleted OCI-AML2 cells treated with 200ng/mL SDF1𝛼 in serum-free media for 30 minutes.

D-G, I-J, L-M: Experiments were performed at least twice with similar results.

A core node featured predominantly PIK3CG which consistently interacted with PIK3R5 across all profiled AML cell lines. This core complex was associated with critical kinase and transcription factor regulators of myeloid cells such as FLT3, RUNX1, and TET2, which are commonly found mutated and dysregulated in myeloid neoplasia²⁹. Notably, PIK3CG-PIK3R5 activity, predominantly controlled by G protein-coupled receptors (GPCRs), gathered various guanine nucleotide exchange-related factors (GEFs) to the PIK3CG/R5 core complex^30,31. This core complex was also physically associated with proteins linked to mitochondria and nucleopores, hinting at a biological link between PI3K gamma signaling, mitochondrial function and metabolism, and nucleocytoplasmic transport. Interestingly, the TORC2 complex comprising MTOR, RICTOR, and FKBP8, along with the R2TP/TEL chaperone complexes, were identified as significant interactors of PIK3CG/R5. These complexes were reported to regulate AKT activation levels. TORC2 directly phosphorylates AKT on Ser473 and facilitates Thr308 phosphorylation by PDK1, while R2TP/TEL2 stabilizes MTOR and modulates AKT through downstream regulation of AKT2 phosphorylation in the PI3K signaling pathway^32,33. Furthermore, activation of PIK3CG via GPCRs leads to the production of the versatile second messenger PtdIns(3,4,5)P3, serving as a docking site for pleckstrin homology domain-containing kinases, including AKT^23,34.

We then investigated whether the highly tissue-specific expression pattern of the PIK3CG-PIK3R5 signaling network in AML conferred isoform-specific control of downstream AKT signaling. In a panel of 15 AML cell lines, we found that PIK3CG dependency was correlated with dependency on AKT1 and AKT2 in AML cell lines but not cell lines from other tissues (Figures 3B and Extended Data Fig. 5A). Given the correlation between PIK3CG dependency and AKT1/2, we used a CRISPR/Cas9 drug-modifier screen to identify a common set of genes that significantly altered sensitivity to MK-2206, a pan-AKT inhibitor. Cells harboring sgRNAs targeting BCL2 were most depleted in MK-2206 modifier screens, while cells harboring sgRNAs targeting negative regulators of TOR signaling (TSC1, TSC2, NPRL2) were enriched (Figure 3C). These findings are consistent with our data and others’, positioning the venetoclax target, BCL2, and MTOR as protein candidates whose inhibition synergistically enhances the effects of PI3K pathway inhibition^35,36. In this cellular context, the potentiation of AKT inhibition appears to be selective to BCL2 suppression, as no other anti-apoptotic BCL2-family proteins scored as sensitizers. Furthermore, topoisomerase-encoding genes, which are targets of anthracyclines such as daunorubicin, and polymerase-encoding genes affected by cytarabine, did not score as sensitizers to MK-2206 (Extended Data Fig. 5B).

To further confirm the connection between PIK3CG-PIK3R5 with AKT signaling, we showed that PIK3CG and PIK3R5 knockouts in OCI-AML3 and OCI-AML2 cells diminish the phosphorylation of AKT and its previously reported downstream substrates, PRAS40 and TSC2 (Figure 3D). The phosphorylation levels of both AKT1 and AKT2 isoforms were decreased upon suppression of PIK3CG and PIK3R5 (Figure 3E). PIK3CG silencing did not alter the expression levels of BCL2, or the pro-apoptotic proteins BAX, PUMA, and BIM, but we observed a slight decrease in MCL1 (Figure 3F). To determine whether AKT signaling relies selectively on PIK3CG in AML cells, we employed doxycycline-inducible shRNAs to knock down the expression of PIK3CA, PIK3CB, and PIK3CD, and revealed that PIK3CG suppression yielded predominantly a decrease in the activation of AKT, affecting both T308 and S473 phosphorylation sites, in contrast to the other PIK3 catalytic isoforms (Figure 3G). A similar approach to selectively reduce expression of the PI3K regulatory subunits confirmed that only shRNAs directed against PIK3R5 significantly dampened AKT phosphorylation compared to inhibition of the other regulatory subunits (Figures 3H and 3I). Conversely, overexpression of wild-type PIK3CG under a constitutive EF1α promoter resulted in hyperactivation of AKT signaling. We then introduced an shRNA directed against the 3’ UTR region of PIK3CG, capable of knocking down endogenous PIK3CG, without affecting exogenously expressed PIK3CG cDNA lacking the 3’ UTR. Expression of wild-type PIK3CG cDNA fully rescued AKT suppression secondary to the suppression of endogenous PIK3CG (Figure 3J).

We finally investigated whether the modulation of GPCR functions could affect PIK3CG-mediated AKT activation. CXCL12 (SDF1α) is the ligand for CXCR4, a GPCR that signals to downstream pathways including the PI3K/AKT pathway. CXCL12/CXCR4 signaling is implicated in leukemic niche maintenance³⁷, and both receptor and ligand are upregulated in AML cells relative to normal hematopoietic cells (Figures 3K and Extended Data Fig. 5C). The addition of pertussis toxin, which broadly impairs GPCR activity³⁸, to AML cell lines reduced AKT phosphorylation, demonstrating a predominant role of GPCRs over other ligand-binding receptors across AML cell lines in modulating AKT signaling (Figure 3L). Because PIK3CG-PIK3R5 is activated via GPCRs, whereas other isoforms are predominantly under the control of receptor tyrosine kinases (RTKs)³⁹, we interrogated the effect of PIK3CG suppression in the presence or absence of the CXCR4 GPCR-activating ligand, CXCL12 (Figure 3M). Following PIK3CG knockout, CXCL12 failed to activate downstream AKT signaling. These data provide evidence that PIK3CG mediates AML cell AKT signaling downstream of GPCRs, including CXCR4.

PIK3CG Inhibitors Do Not Sustain Long-Term AKT suppression

Given the PIK3CG dependency across several AML models, we anticipated that small-molecule inhibitors targeting PIK3CG would suppress AML cell growth. Surprisingly, pharmacological inhibition of PIK3CG using IPI-549 and AZ2, two PIK3CG-isoform selective small-molecule inhibitors, did not markedly reduce AML cell viability and colony-forming capacity compared to genetic suppression of PIK3CG/PIK3R5 (Figures 4A and 4B). Notably, no further reduction in cell growth was measured with the addition of fresh IPI-549 or AZ2 every 24 hours, suggesting that the blunted activity of these small molecules relative to PIK3CG/PIK3R5 knockout was not caused by drug instability (Figure 4C). We therefore postulated that these small-molecule inhibitors might not durably inhibit PIK3CG-mediated AKT signaling. Indeed, long-term suppression of AKT phosphorylation was achieved using sgRNAs targeting PIK3CG/PIK3R5 but not with IPI-549 or AZ2 treatment, which showed AKT reactivation by 72 hours (Figure 4D). This activation persisted despite retreatment of the cells with fresh drugs for one hour (Figure 4E). In contrast, the AKT inhibitor MK-2206, a positive control for AKT signaling inhibition, repressed AKT phosphorylation durably (Figures 4D and 4E). Overall these findings imply that existing small-molecule inhibitors of PIK3CG fail to durably suppress downstream AKT signaling, and consequently, they cannot attain the same level of cytotoxicity as PIK3CG and PIK3R5 knockout, which stably block AKT phosphorylation. Therefore, an alternative therapeutic approach for targeting PIK3CG, one that ensures a sustained disruption of PIK3CG/PIK3R5-mediated AKT signaling in AML cells, was needed.

Figure 4: PROteolysis Targeting Chimera (PROTAC)-Based PIK3CG Degradation Is a Promising Strategy to Durably Suppress AKT Signaling in AML.

A-B. Relative cell growth (A) and colony formation capacity (B) of OCI-AML2 cells following sgRNA-mediated suppression of PIK3CG or PIK3R5, or treated with indicated concentrations of small-molecule inhibitors of PIK3CG, IPI-549 and AZ2. Error bars represent mean ± SD of two biological replicates composed of seven technical repeats (A) for the knockout experiment and (B) three biological replicates composed of four technical repeats for the pharmacological study..

C. Growth inhibition, IC50 and AUC values, of OCI-AML2 cells treated once, twice or thrice iteratively with increasing concentrations of IPI-549, or AZ2. Error bars represent mean ± SD of two biological replicates composed of seven technical repeats.

D. AKT phosphorylation levels by western blot in OCI-AML2 cells either infected with PIK3CG- and PIK3R5-directed sgRNAs for 72 hours or treated with 500nM IPI-549, AZ2, or the AKT inhibitor, MK-2206, for one and 72 hours.

E. AKT phosphorylation levels by western blot of OCI-AML2 cells treated for one hour, 72 hours, or 72 hours followed by fresh addition of 500nM IPI-549, AZ2, or MK-2206 for one hour.

F. Simulation of the docking of AZ2 in the PIK3CG pocket. The parental AZ2 inhibitor is shown to be converted to ARM165 degrader by capitalizing on the presence of the acetyl moiety, that is orientated towards the outside of the kinase pocket, to which a linker and subsequent CEREBLON recruiting moiety is attached to generate the PIK3CG degrader compound, ARM165.

G. PIK3CG and PIK3R5 levels by western blot in OCI-AML2 cells treated with increasing concentrations of ARM165. ACTIN used as a loading control. Q: quantification of PIK3CG levels.

H. PIK3CG expression level by western blot in OCI-AML2 cells treated with 1μM ARM165 in combination with either 10nM bortezomib or 10μM lenalidomide for 24 hours.

I. PIK3CG expression level by western blot in HEK293T cells expressing either wild-type or a lysine to arginine mutant form of PIK3CG (K437R, K444R, K455R, K457R, K490R, K510R, K531R, K553R, K572R, K584R, K587R, and K597R). ACTIN used as a loading control.

J. Colony formation from OCI-AML2 cells harboring either a control or a PIK3CG-directed sgRNA and treated with 600nM ARM165. Error bars represent mean ± SD of two biological replicates composed of four and five technical repeats after seven days of seeding.

K. Proportion of mTagBFP-positive OCI-AML2 cells infected with a control (sgCT), and two CRBN-directed sgRNAs (sgCRBN_1 and sgCRBN_2) and treated with 3μM ARM165. Error bars represent mean ± SD of three biological replicates after three days of seeding. P-values calculated using two–tailed Welch t test and reported on the figure panel.

L. Quantitative proteomics analysis of PIK3CG-expressing HEK293T cells treated with 2μM ARM165 for 24 hours. Volcano plot threshold applied at p-value ≤ 10⁻⁴ and absolute log2 (FC) ≥ 0.8.

M. Phosphorylation levels of AKT and its substrate PRAS40 by western blot in OCI-AML2 cells infected with a control, a representative PIK3CG-, or PIK3R5-directed sgRNA or treated with 1μM ARM165 for 72 hours.

N. Representative growth inhibition curves (left panel), IC50s, and AUCs (right panel) in response to increasing concentrations of ARM165 of OCI-AML2 cells infected with either a control or a myristoylated form of AKT. Error bars represent mean ± SD of three biological replicates composed of seven technical repeats after three days of seeding. P-values calculated using two-tailed Mann Whitney and reported on the figure panel. Different shapes of dots indicate distinct biological replicates.

A, B, J, K. P-values calculated using one-way ANOVA and reported on the figure panels. Different shapes of dots indicate distinct biological replicates.

D-E, G-I, M. Experiments were performed at least twice with similar results.

Harnessing recent advancements in targeted protein degradation, we developed a PROteolysis TArgeting Chimera (PROTAC) system, in which AZ2, a PIK3CG-isoform selective inhibitor, was conjugated to a CEREBLON-targeting moiety. We capitalized upon the presence of an acetyl moiety on the AZ2 parental compound that is orientated towards the outside of the kinase pocket to attach a linker and subsequent CEREBLON-recruiting moiety (Figure 4F). We discovered a lead compound, designated as ARM165, which was synthesized from AZ2 through a 5-step chemical pathway (Extended Data Fig. 6). To confirm that ARM165 targets PIK3CG, we treated OCI-AML2 cells with increasing concentrations of ARM165 and observed 50% degradation of PIK3CG, but not PIK3R5, at 1μM ARM165 (Figures 4G). We did not notice any major hook effect up to 10μM ARM165, potentially attributable to a favorable protein/protein interaction that broadens the concentration range in which ternary complex formation prevails over binary complexes, as described in other systems^40–43. Therefore, PIK3CG and CEREBLON may bind cooperatively in the presence of ARM165. Co-treatment of cells with the proteasome inhibitor bortezomib fully rescued the loss of PIK3CG protein induced by ARM165, indicating proteasome-mediated PIK3CG degradation. Additionally, co-treatment of cells with ARM165 and lenalidomide, which binds to the same site on CEREBLON, competed with ARM165 to block PIK3CG degradation (Figure 4H). Consistent with ubiquitin-mediated degradation of PIK3CG by ARM165, a lysine to arginine PIK3CG mutant in which lysine residues between position 427 and 597 were mutated into arginine, was resistant to ARM165 (Figure 4I). Cells lacking PIK3CG, or CRBN (encoding CEREBLON), were resistant to the PROTAC degrader, indicating that the cytotoxic effect of ARM165 is attributable PIK3CG degradation (Figures 4J and 4K). Through an unbiased proteomic-based analysis conducted on HEK293T cells expressing exogenous PIK3CG, we confirmed that PIK3CG is the most significant ARM165 target (Figure 4L). Finally, cells treated with ARM165 demonstrated a reduction in PIK3CG comparable to that achieved by CRISPR-Cas9 targeting sgRNAs, resulting in a similar level of disruption in downstream AKT signaling (Figure 4M). Importantly, ectopic expression of a constitutively activated, myristoylated form of AKT alleviated markedly the ARM165-induced decreased AML cell growth (Figure 4N). Collectively, these findings support ARM165 as a newly developed selective heterobifunctional degrader of PIK3CG which durably alters AKT signaling in AML cells.

Degradation Over Inhibition of PIK3CG Improved Cytotoxicity

We first compared the impact of both ARM165 and the parental AZ2 with that of an analogous molecule, ARM204, featuring a single R group alteration—replacing the hydrogen atom of the imide function with a methyl group (Extended Data Fig. 7A). This minor structural modification rendered the compound incapable of binding to CEREBLON⁴⁴. Increasing concentrations of ARM165 significantly affected the growth of these cell lines, whereas AZ2 and ARM204 had minimal anti-leukemia effects (Figures 5A and 5B, and Extended Data Fig. 7B). Because the dependency on PIK3CG and its regulatory subunit PIK3R5 was primarily observed in myeloid cells, we hypothesized that the degrader’s efficacy would be more specific to this cell type. Indeed, we found unambiguously that non-AML cell lines were insensitive to ARM165 (Figure 5B). These data support ARM165 as a lineage-specific treatment in AML to avoid the dose-limiting effects of targeting PI3K in non-hematopoietic tissues. Using patient-isolated primary AML cells, we observed significantly compromised viability with ARM165 relative to parental compound, AZ2 (Supplementary Table 1 and Figure 5C).

Figure 5: Degradation of PIK3CG Demonstrated Superior Cytotoxic Performance Relative to Existing Small-Molecule Inhibitors of PIK3CG.

A. Representative growth inhibition curves, with IC50s and AUCs (n=7, mean ± SD) of two human AML cell lines treated with increasing doses of ARM204 (a non-PIK3CG targeting inactive small-molecule analog to ARM165), AZ2, and ARM165.

B. AUCs reflecting viability of multiple indicated human AML and non-AML cell lines treated with increasing ARM204, AZ2, and ARM165. Error bars represent mean ± SD of three biological replicates composed of seven technical repeats after three days of seeding. P-values calculated using one-way ANOVA and reported on the figure panel. Different shapes of dots indicate distinct biological replicates.

C. Growth inhibition, IC50 and AUC values, of four primary samples (n=4, mean ± SD) from patients with AML treated with increasing AZ2 and ARM165after five days of seeding.

D. Principal component analysis of OCI-AML2 and MOLM-14 cells treated with DMSO, 1μM and 500nM AZ2 or ARM-165 for 24 hours, respectively.

E. Representative upregulated and downregulated pathways along with their respective biological functions, using the top dysregulated genes (with q-values < 0.001) common to ARM165-treated OCI-AML2 and MOLM-14 cells.

F. Heatmaps of leading-edge genes of biological pathways reported in panel E and their expression levels compared to DMSO following AZ2 and ARM165 treatments of MOLM-14 cells.

G. Proportion of Annexin V-positive MOLM-14 and OCI-AML2 cells treated with 500nM and 1μM ARM165, respectively, for 48 hours. Error bars represent mean ± SD of three biological replicates. P-values calculated using two-tailed Welch t test and reported on the figure panel.

H. ARM165 gene signature representation in AML patients from the BEAT-AML cohort with low versus high PIK3CG expression. ON versus OFF ARM165 signatures were assigned for each patient based on their ES score below or above the median, respectively. The number of ARM165-ON patients in the PIK3CG high subset versus the PIK3CG low subset was compared. P-values using two-tailed Fisher’s t-test and reported on the figure panel.

To further assess ARM165 activity, RNA-sequencing data from cells treated with vehicle, AZ2, or ARM165 were projected onto a principal component analysis (PCA) plot, confirming that ARM165-treated cells displayed a markedly altered transcriptional profile compared to those treated with the parental compound AZ2 (Figure 5D). An open-ended overrepresentation analysis identified P53- and apoptosis-related gene sets as being significantly activated by ARM165, whereas oxidative phosphorylation-, cell cycle-, and AKT/MTOR-related gene sets were significantly repressed (Figure 5E). The expression levels of leading-edge genes from each of these representative biological pathways were indeed dysregulated by ARM165, but not by AZ2 (Figure 5F). We observed an increase in Annexin V-positive MOLM-14 and OCI-AML2 cells in response to ARM165, confirming that the PROTAC degrader promotes apoptosis in AML cells (Figure 5G). Employing single-sample Gene Set Enrichment Analysis (ssGSEA), we then queried the predicted response of patients to ARM165 based on their PIK3CG expression. Patients with elevated levels of PIK3CG were significantly more likely to respond to ARM165 treatment, further demonstrating ARM165’s favorable selectivity towards PIK3CG (Figure 5H).

PIK3CG Degradation Potentiates the Effect of Venetoclax

We investigated whether the ARM165 degrader would sensitize AML cells to venetoclax treatment. We determined that increasing doses of ARM165 significantly amplified the effect of venetoclax on reducing the growth and colony-forming capacity of various AML cell lines, in contrast to cells treated with the parental AZ2 compound (Figures 6A and 6B, and Extended Data Fig. 8A). ARM165 sensitized cells to venetoclax to a greater extent than to daunorubicin and cytarabine, for which ARM165 only mildly enhanced the anti-leukemia effect (Extended Data Fig. 8B). In addition, in OCI-AML2 cells, ARM165 maximized the cell growth defect induced by venetoclax, but not by MCL1 or BCL-X_L inhibition using S63845 (MCL1 inhibitor), WEHI-539 (BCL-X_L inhibitor), or A-1331852 (BCL-X_L inhibitor), despite basal sensitivity of these cells to MCL1 inhibition (Figure 6C). Although the sensitivity to venetoclax may vary among different AML cell types, these results suggest that PIK3CG degradation is more likely to enhance the efficacy of venetoclax. This potentiation of venetoclax’s effect was also shown in primary patient cells. ARM165 alone induced a marked decrease in colony formation when compared to parental AZ2, and this effect was further enhanced when combined with venetoclax (Figure 6D). Using an excess over Bliss analysis, we determined that ARM165 and venetoclax act synergistically to impair AML cell viability (Figure 6E). Notably, this synergistic activity of ARM165 with venetoclax was significantly greater than that of AZ2 with venetoclax across nine primary patient samples (Figure 6F).

Figure 6. PIK3CG Degradation Potentiates the Effect of Venetoclax in Multiple AML Models.

A. Venetoclax sensitization effect of AZ2, or ARM165 across indicated human AML cell lines. Error bars represent mean ± SD of three biological replicates composed of seven technical repeats after three days of seeding.

B. Colony formation from five human AML cell lines treated with AZ2 or ARM165 in combination with venetoclax (MOLM-14: 0.75μM AZ2 or ARM165, and 375nM venetoclax; MV4–11: 0.5μM AZ2 or ARM, and 75nM venetoclax; OCI-AML2: 1μM AZ2 or ARM165, and 500nM venetoclax; NOMO-1: 0.625μM ARM165 or AZ2, and 10μM venetoclax; HL60: 1μM ARM or AZ2, and 2.5μM venetoclax). Error bars represent mean ± SD of three biological replicates composed of four technical repeats after seven days of seeding.

C. Growth inhibition curves (left panel), IC50 values, and AUCs (right panel) of OCI-AML2 cells treated with DMSO or 500nM ARM165 in combination with increasing doses of either venetoclax, an MCL1 inhibitor, S63845, or two BCL-XL inhibitors, WEHI-539 and A-1331852. Error bars represent mean ± SD of three biological replicates composed of seven technical repeats after three days of seeding.

D. Colony formation from two human AML primary samples (n=4, mean ± SD) treated with indicated AZ2 or ARM165 concentrations in combination with 375nM venetoclax after seven days of seeding.

A-D. P-values calculated using one-way ANOVA and reported on the figure panel. Different shapes of dots indicate distinct biological replicates.

E. Bliss synergy plots for two primary patient samples with AML and treated with ARM165 and venetoclax across a drug-dilution matrix for five days. Delta scores from high synergy (lighter blue) to no synergy (dark blue).

F. Bliss synergy scores for 9 primary patient samples with AML and treated with a drug-dilution matrix of AZ2 and venetoclax, or ARM165 and venetoclax. Bars represent median and violin plots represent range. P-values calculated using two-tailed Mann-Whitney test and reported on the figure panel.

G. In vivo limiting dilution assay performed with primary murine Cbfb-MYH11 leukemic cells treated for 24h with 0.5μM AZ2 or ARM165 and injected into sublethally-irradiated recipient animals at decreasing cell concentrations. Determination of Leukemia-Initiating Cell (LIC) frequency with a 95% confidence interval in each group using Extreme Limiting Dilution Analysis (ELDA). Two-sided chi-squared test used for statistics and reported on the figure panel.

H. Proportion of Cbfb-MYH11-driven GFP-positive leukemic blasts in bone marrow of euthanized animals treated by intravenous injection with vehicle (Veh), 0.051mg/kg AZ2, or 0.051mg/kg ARM165 for five consecutive days (n=9 mice per group, mean ± SD).

I. Proportion of Cbfb-MYH11-driven GFP-positive leukemic blasts in spleen (left panel) and corresponding spleen weight (right panel) of euthanized animals treated by intravenous injection with vehicle (Veh) or 0.051mg/kg ARM165 for five consecutive days (n=11 mice in vehicle and n=10 mice in ARM165 group, mean ± SD).

J. Left panel Proportion of blood circulating Cbfb-MYH11-driven GFP-positive blasts in animals treated by intravenous injection with vehicle (Veh), 0.051mg/kg AZ2, or 0.051mg/kg ARM165 in combination with 100mg/kg venetoclax for five consecutive days (n=5 mice per group, mean ± SD). Right panel. Proportion of blood circulating CD45-positive human AML primary blasts in animals transplanted with a PDX for a month and then treated by intravenous injection with vehicle (Veh) or 0.051mg/kg ARM165 in combination with 100mg/kg venetoclax for consecutive days (n=5 mice per group, mean ± SD).

H-J. P-values calculated using two-tailed Mann-Whitney test and reported on the figure panel.

Given reports that durable response to therapy in AML necessitates an abrogation of leukemia-initiating cells (LICs), we investigated the impact of ARM165 on the LIC fraction of a Cbfb-MYH11-driven mouse model of AML. Cbfb-MYH11-positive cells were treated ex vivo with either ARM165 or AZ2 prior to reinjection into sublethally-irradiated secondary recipient mice at various cell concentrations. Using extreme limiting dilution analysis, we observed a 4.9-fold decrease in LIC frequency in secondary recipients injected with blasts treated with ARM165 compared to those engrafted with blasts exposed to AZ2 (Figure 6G). These promising results led us to evaluate the impact of ARM165 on mice injected with Cbfb-MYH11-driven leukemic cells. The large molecular weight of PROTACs and other heterobifunctional molecules is well-known for presenting challenges in terms of solubility and bioavailability for in vivo testing. IV injection of ARM165, over four consecutive days, of mice injected with Cbfb-MYH11-driven AML cells resulted in a significant reduction in leukemia burden in bone marrow when compared to animals treated with either vehicle or AZ2 (Figure 6H). The assessment of leukemic burden of the spleen as a secondary site of leukemic infiltration confirmed the anti-leukemia effect of ARM165, mirroring the results observed in the bone marrow (Figure 6I). This treatment regimen did not significantly alter the weight of naive mice (Extended Data Fig. 9A). It induced a slight yet significant decrease in the hematopoietic stem and myeloid progenitor cell fraction in the bone marrow and spleen monocytes, and an increase in blood hematocrit and hemoglobin levels (Extended Data Fig. 9B).

Finally, we investigated the anti-leukemic effect of ARM165 in combination with venetoclax in i) syngeneic mice transplanted with Cbfb-MYH11-driven leukemic cells and ii) NSG-S mice xenografted with AML primary patient cell material. In both mouse models, animals treated with the combination of these two drugs displayed a significantly reduced leukemic burden compared to those treated with ARM165 or venetoclax alone (Figure 6J). These findings confirm the promise of PROTAC-based PIK3CG destabilization, whether used alone or in combination with venetoclax, as a therapeutic approach with potential advantages over existing PIK3CG-targeting small-molecule inhibitors.

DISCUSSION

The success of targeted therapies in cancer is linked to the capacity to identify molecular subtypes within tumors and discern the dependencies responsible for their onset and progression. In this context, the concept of lineage addiction emerged as the idea that certain cancer types are reliant on the abnormal activation of genes or pathways generally essential for the development and function of their lineage of origin. Here, we confirmed that the expression of PIK3CG and its regulatory subunit, PIK3R5, is predominantly restricted to the myeloid compartment, thereby completing previous observation that PIK3CG is highly expressed in myeloid cells^20–22. Using a variety of genetic tools, we showed that the suppression of these genes has a significant impact on AKT signaling and the growth of AML cells, contrastingly, not markedly observed upon knockdown of other PI3K isoforms. Despite the expression of other PI3K isoforms in AML cells, our findings, and those of a corroborating study published during the review of this manuscript²⁵, suggest that PIK3CG and PIK3R5 constitute AML-specific vulnerabilities, primarily because of their prominent role in regulating AKT signaling in comparison to the other PI3K isoforms. Targeting this ‘lineage-specific signaling addiction’ may enable therapy to be directed toward the tissue compartment where malignancy is present, creating the potential for a therapeutic window that can spare non-malignant tissues.

We demonstrated that targeting PIK3CG can enhance the effectiveness of the BCL2 inhibitor venetoclax, which is a newly approved standard-of-care therapy for elderly patients with AML⁴⁵. This increased sensitivity to venetoclax is likely a direct result of AKT inhibition. We established this through a whole-genome CRISPR screen, which ranked BCL2 as one of the top sensitizers to the AKT inhibitor MK-2206. The underlying mechanism of this interrelationship between AKT and BCL2 may lie in the regulation of the anti-apoptotic protein MCL1. Inhibition of AKT and mTOR signaling has been shown to decrease MCL1 protein levels in AML cells⁴⁶, and in multiple cancer models, MCL1 expression has been found to be anti-correlated with the response to venetoclax⁴⁷. The decreased expression of MCL1 upon PIK3CG and PIK3R5 signaling inhibition may thus play a substantial role in potentiating the effect of venetoclax, though additional studies are required to fully elucidate the mechanism underlying venetoclax sensitization.

The small-molecule inhibitor IPI-549 was developed to target the mechanism of immune suppression in tumorigenesis^20,23,48, and subsequently, other PIK3CG-targeting compounds including AZ2 were developed⁴⁹. Despite these developments, IPI-549 remains the only PIK3CG-targeting molecule under clinical investigation to date²⁴. Considering the evidence that merely inhibiting PIK3CG-mediated AKT activation with these small-molecule kinase inhibitors was unable to achieve a durable elimination of AML cells, we developed a comprehensive targeting strategy using a CRBN-based PROTAC system to degrade PIK3CG. This approach demonstrated a significant anti-leukemia effect. Notably, several studies speak to the dual scaffolding and catalytic role of PIK3CG in promoting downstream signaling events. The contrasting phenotypes between PIK3CG-deficient versus PIK3CG kinase-dead mice revealed that this enzyme serves as a molecular scaffold orchestrating cellular signaling complexes independently of its lipid kinase activity^50–53. These scaffolding functions are notably selective, and the absence of a specific PI3K isoform is unlikely to be compensated for by others. Interestingly, these functions have been well-documented for GPCR-dependent PI3-kinases like PIK3CG, and they may act in concert or independently from the kinase activity. For example, PIK3CG’s kinase-independent scaffold functions are intertwined with its kinase activity in diet-induced obesity, while PIK3CG plays a role in insulin signaling independently of its kinase activity^52,54–56.

The findings that PI3K isoform scaffolding and kinase activity functions can be disentangled to affect a broad array of cellular functions highlights how various PROTAC-based degradation approaches can result in more profound functional outcomes compared to the targeting of kinase function alone. Similar to our findings with PIK3CG degradation, targeted protein degradation can often outperform small molecule inhibitors initially designed for the same protein target. For example, dBET1 exhibited a greater apoptotic response in primary AML cells compared to its precursor, JQ1, emphasizing the potential superiority of BET degradation over bromodomain inhibition⁵⁷. PROTAC degraders targeting STAT3 and STAT5 have also demonstrated significantly greater potency than small molecule inhibitors, which were reported to lack adequate potency and selectivity^58,59. Similarly, the development of a selective FAK kinase degrader exhibited improved activity in downstream signaling, cancer cell viability, and migration compared to FAK kinase inhibitors⁶⁰.

The results presented here argue for cancer therapeutics specifically designed to target dependencies driven by proteins with tissue- and/or tumor-restricted expression. Our data also provide a proof-of-concept that the degradation, rather than inhibition, of signaling molecules, such as PIK3CG, could offer improved therapeutic benefits for patients, and more sustained response to adjuvant therapy regimens. Further efforts are necessary to enhance the solubility and in vivo bioavailability of our degrader, thereby creating a pharmacologically more efficient preclinical-grade version which could open the path for promising clinical development.

METHODS

Cell Culture and Reagents

U937, KG1α, HL60, and OCI-AML3 cell lines were purchased from the American Type Culture Collection (ATCC), OCI-AML2 cells were provided by the Duke University Cell Culture Facility (CCF), and MOLM-14 and MV4–11 cells were provided by Dr. Scott Armstrong (Dana-Farber Cancer Institute, Boston, MA). MCF7, HCT116, HT29 cells were provided by Dr. Claude Gazin (IFJ, Paris, France), and NOMO-1 cells were provided by Dr. Ross Levine (MSKCC, New-York, NY). Identity of all cell lines was confirmed by short tandem repeat loci profiling. All cell lines were tested negative for Mycoplasma using MycoAlert PLUS Mycoplasma Detection Kit (Lonza #LT07–705). All cell lines were maintained in RPMI 1640 (Sigma #R2405) supplemented with 1% penicillin–streptomycin and 10% FBS (Sigma # F2442) in a humidified incubator at 37°C with 5% CO2. HEK293T cells were maintained in DMEM (Sigma #D6429) supplemented with 10% FBS and 100U/mL penicillin–streptomycin (Sigma #P4333). Cells infected with shRNA and sgRNA constructs were maintained in culture with 1μg/ml Puromycin and 1μg/ml Doxycycline three days prior to fluorescence activated cell sorting and thereafter. Venetoclax (HY-15531), IPI-549 (HY-100716), MK-2206 (HY-10358), AZ2 (HY-111570), S63845 (HY-100741), WEHI-539 hydrochloride (HY-15607A), A-1331852 (HY-19741), Quizartinib (HY-13001), Sorafenib (HY-10201) and Gilteritinib (HY-12432) were purchased from MedChemExpress. Telatinib (S2231) and Amunatinib (S1244) were purchased from SelleckChem.

Patient Profiling and Primary Patient Sample Preparation

Patient samples were collected from AML patients, from whom informed consent had been given (including for sex collection) as part of an ongoing clinical registry at St Louis Hospital (THEMA, IRB approval #IDRCB2021-A00940–41). Samples were anonymized and stored at the St Louis Hospital Tumor biobank, as declared to the ministry of Higher Education, Research and Innovation. The use of these primary patient cells for experimental procedures derived from clinical practice was approved by the INSERM IRB. In accordance with the declaration of Helsinki and French protection of personal data law, only anonymized clinical data were made available to research teams. Sex of patients has been collected on medical charts and is reported in Supplementary Table 1. No race, ethnicity, or socially relevant data is reported in this study. Cytogenetic analyses required karyotyping and fluorescence in situ hybridization studies guided by karyotype. Genetic profiling consisted of fragment analysis for NPM1, FLT3 and IDH1/2 mutational status.

Mononuclear cells from patients with AML were isolated using Ficoll-Paque PLUS (GE Healthcare #17–1440-02) and red blood cells were lysed (Sigma # R7757). For methylcellulose and synergy assays, cells were maintained in RPMI 1640 medium supplemented with 20% FBS, 20ng/mL IL3 (Peprotech, #200–03), 20ng/mL IL6 (Peprotech, #200–06), 20ng/mL GM-CSF (Peprotech, #300–03), 10ng/ml G-CSF (PeproTech, #300–23), 10ng/mL EPO (PeproTech, #100–64), 50ng/mL TPO (PeproTech, #300–18), 100ng/mL FLT3-Ligand (PeproTech, #300–19), and 100ng/mL SCF (PeproTech, #300–07).

AML patients H3K27ac ChIP-seq Analysis

H3K SRA database was accessible under accession number SRP103200. Sequencing reads were aligned to the human hg19 version of the genome using Bowtie2⁶¹ and duplicate reads were marked using Picard tools MarkDuplicates. Normalized bigwig files for gene track representations were generated using Deeptools⁶² with the --normalizeUsing RPKM --extendReads 200 --smoothLength 150 --ignoreDuplicates options. Normalized density signals from PIK3CG and PIK3R5 gene were extracted using bwtool⁶³ and density heatmaps with average signal bar plots were generated in R using iheatmatpr package.

In vitro Drug Sensitivity Assays

Cells were seeded in 384-well plates, with between 5 and 7 replicates (Corning, #3570). A range of 10 descending concentrations plus a control vehicle-treated well were plated per drug tested. ATP content was measured using CellTiter-Glo^® (Promega, #G7573) per manufacturer’s instructions. Relative cell viability was determined by normalizing raw luminescence values to either DMSO or the indicated background drug. IC50 and AUC values were approximated from dose-response curves plotted using the GraphPad/Prism 8 software.

In vitro Synergy Assays

Cells were seeded in 384-well plates, with 4 replicates per concentration combination of venetoclax, and AZ2 or ARM165. Drugs were diluted 1:2 across, with 10 descending concentrations of venetoclax and 7 descending concentrations of ARM165. Maximum applied concentration of venetoclax and ARM165 were 5μM and 5μM, respectively. Synergy matrix was determined using Bliss additive synergy analysis using the following formula: C=A+B-A*B. (A= effect of agent 1. B= effect of agent 2. C= an expected effect of the combined response).

Plasmids and sgRNA Constructs

sgRNA constructs directed against human and murine PIK3CG, PIK3R5, PIK3CA, PIK3CB, PIK3CD, and CRBN (Supplementary Table S2) were cloned into TCLV2 vector (addgene #87360). The phosphorylated and digested BsmBI sgRNA were ligated into BsmBI digested TCLV2 plasmid in a 10μl reaction containing 1X Quick ligase buffer (NEB), diluted oligo duplex and Quick ligase (NEB M2200S) for 10 minutes at room temperature. For the inducible ectopic expression of PIK3R5, a codon-optimized 3xFlag-HA-PIK3R5-encoding gBlock (IDT) was PCR-amplified, digested using NheI and AgeI enzymes prior to its cloning into a neomycin-selectable doxycycline-inducible pCW57-Crimson vector. This vector enabled flow cytometry-based sorting of cells overexpressing 3xFlag-HA-PIK3R5 using the Crimson fluorescence marker. For the overexpression of AKT, cells were infected with the lentiviral plasmid carrying Myristoylated AKT (addgene #64606) or the control (addgene #64648) as reported before⁶⁴. PIK3CG lysine into arginine mutants was generated by synthesizing a gBlock fragment in which lysines at position 437, 444, 455. 457, 490, 510, 531, 553, 572, 584, 587, and 597 were converted into arginines. This gBlock was subsequently cloned in a pLVX-puromycin vector.

Western Immunoblotting

Western immunoblotting was performed according to Su et al.⁶⁵ using cell lysates normalized for total protein content; cells were resuspended in lysis buffer (Cell Signaling Technologies, #9803S) supplemented with Halt protease and Phosphatase Inhibitor cocktail, EDTA-free (Thermo Fisher Scientific, #78443). Membranes were probed for 16 hours with the primary antibodies depicted in Supplementary Table 3.

CXCL12 Cytokine Experiment

Cells were incubated in serum free RPMI-1640 medium for one hour prior to spiking with 200ng/ml CXCL12/SDF-1a (PeproTech, #300–28A). After 30 minutes, cells were harvested, washed twice, and pellets produced for western blotting.

Generation of Doxycycline-Inducible shRNA Cell Lines

Inducible expression of shRNAs was achieved as previously described using a doxycycline-inducible pLKO-Tet-On lentiviral system⁶⁶. Lentivirus was produced and cells were transduced as described by Su et al.⁶⁵. Following selection with puromycin, shRNA-transduced cells were treated with 75ng/ml doxycycline for 72 hours prior to analysis or experimentation. shRNA target sequences are depicted in Supplementary Table 4.

Generation of Doxycycline-Inducible sgRNA Cell Lines

Following selection with 1ug/mL puromycin for a minimum of seven days, cells were treated with 1μg/mL doxycycline for 72 hours prior to flow cytometry-based sorting of the GFP⁺ population. The sorted fraction containing a bulk population of sgRNA-transduced cells was then reintroduced into culture supplemented in media containing 1μg/ml puromycin and 1μg/ml doxycycline for a minimum of 24 hours and a maximum of seven days, during which time cells were used for experimentation. To generate the CRBN knockout cells, OCI-AML2 Cas9 cells were infected with CRBN-targeting or non-targeting control guides cloned into pLentiGuide-mTagBFP. Nine days after infection, cells were treated with ARM-165 for 72 hours and analyzed for BFP-positive cells by BD Bioscience Symphony flow cytometer.

RT-qPCR

RNA was isolated from cells using QIAshredder Homogenizers and the RNEasy Mini kit (Qiagen) and reverse transcribed to cDNA using the iScript cDNA Synthesis Kit (BioRad) with 1μg of RNA template. qRT-PCR was then performed using iQ SYBR Green Supermix run on a CFX384 Touch Real-Time PCR Detection System. To quantify fold expression change, the ΔΔCq method was used to quantify fold expression change by normalizing cycle threshold (Cq) values to housekeeping gene (ACTB) and normalized to control sample (no doxycycline).

Tandem Affinity Purification of PIK3R5/p101 and Proteomics

OCI-AML2, MV4–11, and NOMO-1 AML cells were transduced with a doxycycline-inducible construct enabling ectopic expression of 3xFlag-HA-PIK3R5. After harvesting, cell pellets were lysed in lysis buffer (100mM KCl, 5mM MgCl2, 20mM Tris-HCl pH 8, 0.1% Tween 20, 0.1% NP40, 10% glycerol + protease inhibitors) and 1mg of lysate was used to perform tandem affinity purification. Lysates were first incubated with anti-flag agarose beads (Sigma-Aldrich, #A2220) for four hours, washed four times with lysis buffer and eluted with 3xflag peptides (Sigma-Aldrich, #F4799). Eluates were incubated overnight with anti-HA agarose beads and washed with lysis buffer then water before digestion for four hours with 1μg of trypsin prior desalting and MS acquisition. The top-scoring interactors of PIK3R5 were identified based on: i) a significant log₂ fold change enrichment compared to control above 1, ii) specific enrichment in test conditions (no peptides identified in controls), with the number of peptides identified significantly above the mean as defined by z-test, and iii) being identified in at least two out of three cell lines. These PIK3R5 interactors were used to construct a STRINGdb physical interaction network (score > 0.5) to visually represent known interactions.

Whole-Genome CRISPR/Cas9 Screen

shRNA library was amplified and prepared as described by Lin et al.⁶⁶ using the Toronto Knockout CRISPR Library – Version 3 (TKOv3) obtained from Addgene (Pooled Libraries #90294, #125517). OCI-AML2 cells were transduced at 1000X coverage of the library (72 × 10⁶ cells transduced) and cultured for a minimum of 1000X coverage for the duration of the screen. After 7 days of puromycin selection, cells were divided into three treatment arms – cells treated with DMSO or MK-2206 (1μM) – for two weeks. Genomic DNA was extracted using the DNeasy Blood & Tissue Kit (Qiagen). Amplification of sgRNA barcodes and indexing of each sample was performed via 2-step PCR⁶⁶. Identification of sensitizing or resistor genes was performed as previously described by comparing the final drug treated populations to DMSO treated⁶⁶.

Tracking of Indels by Decomposition (TIDE) Assay

DNA was extracted from sgRNA infected cells using the Zymoclean^™ Gel DNA Recovery Kit (Zymoclean #D4008). Regions of interest, encompassing the CRISPR-Cas9-edited region, were amplified from the genomic PIK3CG and PIK3R5 DNA regions and prepared for Sanger sequencing. Non-targeting control sequences were aligned to sgPIK3CG/sgPIK3R5 sequences to quantify the efficiency of the sgRNA using the TIDE online tool⁶⁷.

Methylcellulose Assays

AML cell lines were either treated with indicated small-molecule inhibitors or infected with a control, or two PIK3R5- and PIK3CG-directed sgRNAs. Infected cells were selected with puromycin two days after infection and Cas9 expression was then induced with 1μg/mL doxycycline for three days before the sorting of GFP-positive cells. 1×10⁴ cells were plated in methylcellulose (Clona-Cell-TCS Medium #03814) with 1μg/ml doxycycline, 1μg/ml puromycin and treated with indicated drug concentrations. Primary viable cells (1×10⁵ cells/ml) were plated onto semi-solid methylcellulose medium (MethoCult; StemCell Technologies, #04435) and treated with indicated drug concentrations. Cell colonies were evaluated after a minimum of 10 days after plating.

Flow Cytometry

For the detection of GFP-positive Cbfb-MYH11-driven or luciferase-expressing OCI-AML2 leukemic cells, bone marrow and spleen were crushed and smashed, respectively, prior to be washed and resuspended in PBS 2mM EDTA, and analyzed on a BD FACSCanto II Instrument (BD Biosciences). For the detection of primary Patient-Derived Xenograft (PDX) cells, peripheral blood was collected from NOG-EXL mice prior to lysis of red blood cells for 10 minutes (Sigma-Aldrich, #R7757), washing, and resuspension of the leukemic cells into PBS 0.1% BSA 2mM EDTA. PDX cells were then stained for 25 minutes at four degrees with PE-Vio770-coupled anti-hCD45 antibody (Miltenyi Biotec, # 130–110-634). Cells were then washed twice with PBS 0.1% BSA 2mM EDTA, and analyzed on a BD FACSCanto II Instrument (BD Biosciences). The cell gating strategy is shown in Supplementary Figure 1.

Annexin V5 Detection

Annexin V staining was performed using the eBioscience^™ Annexin V Apoptosis Detection Kits per the manufacturer’s protocol (Thermofisher #88–8007-74). Treated cell lines were diluted with 1X binding buffer to 1 million cells per mL before they were stained with Annexin V APC (1:20 dilution) for 10 −15 minutes at room temperature. Cells were washed with 1X of binding buffer. Cells were resuspended with 200uL of 1X binding buffer and immediately analyzed on BD FACScanto II (BD Biosciences).

Animal Studies

Our research complies with all relevant ethical regulations; the French National Ethics Committee on Animal Care reviewed and approved all mouse experiments described in this study. Authorization number: APAFIS #8909–2017021413452743 v1. Animals were housed under a dark-light cycle under ambient temperature (65–75°F (~18–23°C)) with humidity of 40–60%. Animals were excluded of the study if any signs of distress are observed without clinical signs of leukemia: absence of leukemic blasts in bone marrow, spleen, and blood.

Generation of Luciferase-Expressing OCI-AML2 Cells

OCI-AML2 cells were transduced with pMMP-LucNeo retrovirus and selected with 1mg/mL neomycin. These cells were then transduced with lentiviral particles encoding either a non-targeting control or PIK3CG-directed sgRNAs, and then selected for seven days with 1μg/mL puromycin and then used for subsequent in vivo studies.

In Vivo Imaging

After a minimum of six days post-injection of bioluminescent cells, mice were injected with 150mg/kg luciferin (Xenolight, #12799) 15 minutes prior to imaging, anesthetized with 2–4% isoflurane, and imaged, using the IVIS Lumina III according to the manufacturer’s protocol. Bioluminescent signal was quantified in ph. s^-1. cm^-2. sr⁻¹ using a standardized region of interest (ROI) encompassing the entire mouse.

In Vivo Genetic Studies

Luciferase-expressing OCI-AML2 cells were infected with non-targeting control or PIK3CG-directed sgRNAs, selected with puromycin; the GFP-positive cell fraction was sorted as described above. Following sorting, cells were allowed to recover in media supplemented with 1μg/mL doxycycline and 1μg/m puromycin for four days. 6-week-old male NSG mice (NOD.Cg.Prkdc <<scid> II2rg<tm1Sug>Tg(SV40/HTLV-IL-3, CSF2)10–7jic Taconic) were then irradiated at 1.25 Gy and injected with 1 ×10⁶ cells per mouse by intravenous injection. These mice were supplemented with doxycycline-containing food (SAFE nutrition service, #E8200P01R) for the whole duration of the experiment. Imaging of the mice was performed every three days to follow the disease progression. Peripheral blood was collected by submandibular bleeding and bone marrow biopsies were performed on mice femurs to collect bone marrow mononuclear cells. Either, at the end point, mice were euthanized, and spleen and bone marrow harvested. The GFP-positive AML cell fraction was quantified in both bone marrow and spleen on a BD FACSCanto II Instrument (BD Biosciences).

For survival studies, mice were monitored for signs of distress, such as weakness, weight loss, ruffled coat, lethargy or bruising. Observance of humane endpoints as described in our protocol—loss of body weight > 15% free-feeding body weight, inability to rise or ambulate, presence of labored respiration, wound ulceration, or other signs of active infection—resulted in euthanization. As this study pertains to leukemia, a maximal tumor size was not enforced. Drug treatments and routine monitoring continued for 52 days or until a humane endpoint was reached. Maximal disease burden permitted by the ethics committee was not exceeded.

In vivo Drug Treatment Studies

Cbfb-MYH11 Syngeneic Mouse Model:

The Cbfb-MYH11-driven leukemic cells were kindly provided by Dr. Lucio U. Castilla’s team. Those were isolated from the offspring of floxed Cbfb-MYH11 knock-in mice which were crossed with the Mx1-Cre transgenic mice⁶⁸. 0.5 × 10⁶ Cbfb-MYH11 cells were injected into 5- to 8- week-old male C57BL/6J mice (Envigo). Seven days after injection, a biopsy of the bone marrow was performed to confirm disease engraftment and mice were randomized to ensure homogenous disease onset across animals prior to the start of the treatment.

Patient-Derived Xenograft (PDX) Mouse Model:

The PDX sample was derived from a 69-year-old female who was diagnosed with secondary AML with MDS-related changes; the genetic profiling of this patient revealed mutations in CEBPA/ ASXL1/ RUNX1/ EZH2/ JAK2/ TET2; patient karyotype is 46, XX, t(6;7)(q23;q11.2)[1]/46,XX[cp19]. 1×10⁶ mononucleated cells were injected via the tail vein into sublethally-irradiated (1.25Gy) 8-week-old NOG-EXL mice (Taconic Biosciences). Blast engraftment was confirmed by flow cytometry one month after injection using the PE-Vio770-coupled CD45 (hCD45)-based staining protocol detailed in the flow cytometry section to confirm disease engraftment and mice were randomized accordingly.

Venetoclax was administered daily at a dosage of 100mg/kg by oral gavage (in 5% DMSO; 40% PEG300, 5% TWEEN80, 50% Saline), AZ2 and ARM165 were administered daily until mice demise at 0.051mg/kg by IV injection (in 20% N-methylpyrrolidone EMPLU, #8060721000, Sigma-Aldrich; 16% PEG400, 64% saline solution). At indicated time points response to treatment was tracked using either peripheral mandibular bleeding and bone marrow biopsy. 14 days after cell injection, mice were euthanized, bone marrow and spleen were harvested, weighted and the GFP-positive leukemic cell fraction was quantified using a BD FACSCanto II Instrument (BD Biosciences).

In vivo Toxicity Studies

Naïve C57BL/6 mice were treated with ARM165 at 0.051mg/kg by IV injection (in 20% N-methylpyrrolidone EMPLU, # 8060721000, Sigma-Aldrich; 16% PEG400, 64% saline solution) or Vehicle alone for 5 days and toxicity was evaluated via weight measurement and characterization of hematopoietic cell compartment. Mice were monitored for signs of toxicity through measurement of body weight over 7 consecutive days or on day 5 by FACs analysis of the hematopoietic cell compartment. Blood was collected and hematopoietic populations were characterized using a MS9–5 machine as per the manufacturer’s protocol. Bone marrow and spleen were collected and processed with antibodies (listed in Supplementary Table 5) to analyze each hematopoietic stem, progenitor, or mature cell fractions in bone marrow, and granulomonocytic and B- and T-cell fractions in spleen.

RNA-sequencing Analysis

For RNA-sequencing analysis, total RNA was extracted from MOLM-14 and OCI-AML2 cells treated with either DMSO, 0.5μM or 1μM AZ2 or ARM165, respectively, for 24 hours using RNAeasy plus mini kit following manufacturer’s instruction. Sequencing libraries were prepared using TruSeq Stranded mRNA LT Sample Prep Kit (Illumina) and 151bp paired-end sequencing on a NovaSeq6000 (Illumina) at Macrogen Europe (Amsterdam, The Netherlands). The total number of reads for individual samples ranged from 134 million to 217 million with at least 94% of reads >Q30. Quality-control tests for the unmapped reads were performed using the FASTQC software (http://www.bioinformatics.babraham.ac.uk/projects/fastqc/). Sequencing reads were pseudo aligned to the ENSEMBL GRCh38 version of the human genome and abundance quantified using Kallisto (https://pachterlab.github.io/kallisto/). Differential expression analysis was done using the Sleuth R package with Wald test between comparisons⁶⁹. q-values were determined using a Benjamini-Hochberg correction of p-values. PCA plots and heatmap were plotted using R packages ggplot2 and pheatmap.

Overrepresentation Analysis

A list of the top 2928 commonly up-regulated and 2756 commonly down-regulated genes (based q-value < 0.001 by Sleuth analysis) between the control and ARM165-treated conditions, was selected across the two AML cell lines for over-representation analysis. Overrepresentation analysis was performed in R v4.2.2 using genekitr package v1.2.2 using MSigDB HALLMARK and C2 geneset libraries.

Single Sample Gene Set Enrichment (ssGSEA) Analysis

Single-sample GSEA (ssGSEA) was used to calculate separate enrichment scores for each sample whose transcriptomic data was available from BEAT-AML (n=451 patients) and gene sets (queried from the MSigDB Hallmark and C2 and ARM165 downregulated gene signature). The top 149 common downregulated genes in the two AML cell lines treated with ARM165 was retained for ssGSEA analysis. ssGSEA was computed in R with GSVA v1.48.3 package using the following parameters: method=“zscore”, abs.ranking=FALSE, min.sz=2, max.sz=Inf, parallel.sz=18, mx.diff=TRUE, tau=switch(method, gsva=1, ssgsea=0.25, NA), ssgsea.norm=TRUE.

PIK3CG expression patient stratification: Gene expression data of AML patients whose transcriptomic profiling was available from the BEAT AML cohort (n=451)²⁶, were z-score normalized and high versus low PIK3CG levels were evaluated based on the absolute z-score cut-off of 0.7. On versus off ARM165 signatures were assigned for each patient based on a normalized enrichment score below or above the median score across the whole cohort. The significance of the differences between the proportions of each subgroup of patients was evaluated by applying the two tailed Fisher’s Exact Test.

Associations Between Genetic/Clinical Features and PIK3CG/PIK3R5 Levels

For BEAT-AML RNAseq analysis, we used STAR counts and clinical annotation from BEAT AML 1.0 cohort²⁶. Gene expression from AML diagnosis and relapse samples were normalized using the median of ration method, using DESeq2 package. Single-sample cell type signatures^27,28 were calculated using the gsva package⁷⁰. Differences between PIK3CG/PIK3R5 gene expression and cytologic/genetic groups were tested using a Wilcoxon test, and correlations with transcriptomic signatures and drug sensitivity AUC were tested using Spearman correlation.

Protein Abundance Analysis

Pellets from HEK293T expressing PIK3CG with or without treatment of 2μM of ARM165 for 24 hours were lysed in 8M urea + 25mM ammonium bicarbonate (ABC) followed by reduction (5mM DTT for 1hr at 37C), alkylation (10mM iodoacetamide for 30min at RT in the dark), and digestion overnight with 2ug of trypsin (Promega). Peptide samples were applied to activated columns, and the columns were washed three times with 200 μL of 0.1% TFA. Peptides were eluted with 140 μL of 50% ACN and 0.25% formic acid and lyophilized. Samples were resuspended in 0.1% formic acid and separated by reverse phase using a Vanquish Neo HPLC (Thermo Scientific) using 15 cm long PepSep column with a 150 μm inner diameter packed with 1.5um Reprosil Saphir C18 particles (Bruker). Samples were acquired by DDA (2 untreated and 2 ARM165 samples to build spectral library) and DIA (all 4 replicates per conditions) methods. Mobile phase A consisted of 0.1% FA in water and mobile phase B consisted of 80% acetonitrile (ACN)/0.1% FA. Peptides were separated at a flow rate of 500 nL/min by the following 80min gradient: 3%–24% B over 60 min; 24%–40% B over 10 min; and 10 min at 95% B. Eluting peptides were analyzed by an Orbitrap Exploris Mass Spectrometer (Thermo Scientific). Data were collected in positive ion mode with MS1 detection in profile mode in the orbitrap using 120,000 resolution, 350–1,150 m/z scan range, auto maximum injection, and a normalized AGC target of 300 (%). For DDA acquisitions, MS2 fragmentation was performed on charge states from 2 to 6, MIPS mode = peptide, with a 30-s dynamic exclusion after a single selection, and 10 ppm ± mass tolerance. All raw MS data were searched using Spectronaut (v 18.4.231017.55695) against the human proteome (Uniprot canonical protein sequences downloaded September, 2022) using default settings for spectral library generation and DIA analysis.

Chemistry

General considerations:

All organic solvents and reagents were purchased from commercial sources and used as received. ¹H and ¹³C{¹H} NMR spectra were recorded on a Bruker 400 or 500 MHz spectrometer; chemical shifts are given in ppm and referenced to the solvent residual peak (CDCl₃: 7.16 ppm for ¹H and 77.16 for ¹³C; DMSO-d6: 2.50 ppm for ¹H and 39.52 for ¹³C), coupling constants are given in Hertz (Hz). The purity of the final compounds (ARM165 and ARM204) was verified to be ≥ 95 % purity by UPLC analysis at 214 nm on a Waters Acquity H-Class system equipped with a UV detector, SQD2 mass spectrometer detector and an acquity HSS T3 column (100Å, 1.8 μm, 2.1 mm × 50 mm; gradient water/acetonitrile (1‰ formic acid): 100/0 to 0/100 in 5 min; flow: 0.8 mL/min). Preparative HPLC were performed either on Gilson PLC2250 using a C18-reverse phase (DeltaPack Waters, 15 μm, 100Å, 100 × 40 mm) or a GILSON HPLC system – SKID LC 009SK equipped with a C18 reversed-phase Luna column (Phenomenex 15 μm, 100Å, 250 × 50 mm). Docking of AZ2 in PIK3CG was performed on the Alphafold structure of human PIK3CG (Uniprot P48736) using Autodock Vina and the following grid parameters on a rigid receptor: centre x, y, z: 18.701, 19.536, −3.725 and size x, y, z: 21.75, 21.75, 21.75⁷¹.

Synthesis of N¹-(5-(2-((S)-1-cyclopropylethyl)-7-methyl-1-oxoisoindolin-5-yl)-4-methylthiazol-2-yl)-N⁵-(8-((2-(2,6-dioxopiperidin-3-yl)-1,3-dioxoisoindolin-4-yl)amino)-8-oxooctyl)glutaramide (ARM165, Extended Data Fig. 6) and N¹-(5-(2-((S)-1-cyclopropylethyl)-7-methyl-1-oxoisoindolin-5-yl)-4-methylthiazol-2-yl)-N⁵-(8-((2-(1-methyl-2,6-dioxopiperidin-3-yl)-1,3-dioxoisoindolin-4-yl)amino)-8-oxooctyl)glutaramide (ARM204, Extended Data Fig. 7A):

1- Preparation of Compound I ((S)-5-(2-amino-4-methylthiazol-5-yl)-2-(1-cyclopropylethyl)-7-methylisoindolin-1-one)

Synthesis: To a solution AZ2 (cas #: 2231760–33-9; 1.0 g, 2.71 mmol, 1.0 equiv.), in methanol (10 mL) at room temperature, was added a solution of lithium hydroxide monohydrate (1.14 g, 27.1 mmol, 10 equiv.) in water (10 mL). Thereafter, the reaction mixture was stirred at 60°C for 40 h. The reaction mixture was then concentrated to dryness and diluted with ethyl acetate (100 mL). This organic layer was successively washed with a saturated aqueous solutions of sodium bicarbonate and brine, before being dried over magnesium sulphate, filtered, and concentrated to dryness. Compound I was obtained as a beige solid in a 94% yield (834 mg) without further purification.

2- Preparation of Compound IV (5-((8-((2-(2,6-dioxopiperidin-3-yl)-1,3-dioxoisoindolin-4-yl)amino)-8-oxooctyl)amino)-5-oxopentanoic acid):

Synthesis: A) To a solution of pomalidomide (cas #: 19171–19-8, 1.0 g, 3.65 mmol, 1.0 equiv.), and Boc-Aoc-OH (8-[(tert-butoxycarbonyl)amino]octanoic acid, cas #: 30100–16-4, 0.95 g, 3.65 mmol, 1.0 equiv.) in N,N-dimethylformamide (DMF, 25 mL) and pyridine (3.0 mL, 36.5 mmol, 10 equiv.) at room temperature, was added a solution of T3P (2,4,6-tripropyl-1,3,5,2,4,6-trioxatriphosphorinane) 50 wt.% in ethyl acetate (cas # 68957–94-8, 10.8 mL, 18.25 mmol, 5.0 equiv.). the reaction mixture was allowed to stir at 80°C for 16 h. After concentration, the oily residue was dissolved in 100 mL of ethyl acetate and successively washed with a saturated aqueous solutions of sodium bicarbonate and brine, before being dried over magnesium sulphate, filtered, and concentrated to dryness. The crude Compound II was directly engaged in the subsequent step.

B) Crude Compound II (1.46 mmol) was dissolved in 5 mL of anhydrous DMF, at room temperature, and reacted with methyl iodide (120 μL,1.90 mmol, 1.3 equiv.) in the presence of potassium carbonate (260 mg, 1.90 mmol, 1.3 equiv.) during 16 h. After concentration to dryness, the residue was dissolved in 100 mL of ethyl acetate and successively washed with KHSO₄ aq. (1 M, twice), a saturated aqueous solution of sodium bicarbonate and brine. The organic layer was dried over magnesium sulphate, filtered, and concentrated to dryness. The crude Compound II^Me was directly engaged in the subsequent step.

C) This step was separately performed on Crude Compound II (3.65 mmol) and Crude Comopound II^Me (1.46 mmol) as the starting materials. The starting material was dissolved in 15 mL of dichloromethane and treated with trifluoroacetic acid (TFA, 8 mL) at room temperature during 2 h. After concentration to dryness and lyophilisation from a 1/1 mixture of acetonitrile/water, the corresponding Compound III or Compound III^Me was obtained as a trifluoroacetate salt in a light orange solid form and directly engaged in the next step.

D) This step was separately performed on Crude Compound III (3.65 mmol) and Crude Comopound III^Me (1.46 mmol) as the starting materials. The starting material (1.0 equiv.) was dissolved in 20 mL of DMF and 20 mL of toluene. To this mixture, N,N-diisopropylethylamine (DIEA, 3.0 equiv.) and glutaric anhydride (1.1 equiv.) were successively added and the media was allowed to stir at 110°C for 2 h. Thereafter, the volatiles were evaporated off and the residue freeze-dried from a 1/1 mixture of acetonitrile/water. Purification of Compound IV: the crude product (2.6 g) was solubilized in water/acetonitrile 35/65 (v/v) mixture with 1‰ Formic acid (FA) (20 ml), filtered on a 0.45 μm filter and purified by RP-preparative HPLC. The purification was performed on a GILSON HPLC system – SKID LC 009SK - equipped with a C18 reversed-phase Luna column (Phenomenex 15 μm, 100Å, 250 × 50 mm) with a flow rate of 120 mL/min. Eluents were water 1‰ FA (A) and acetonitrile 1‰ FA (B). Purification gradient: 0% to 20% of B in 5 min., then 20% to 30% of B in 5 min., and 30% to 45% of B in 15 min.. The compound is eluted in 19 min. UV detection was performed at 214 nm. Pure Compound IV was obtained as a colourless solid in a 62% (1.2 g) overall yield (3 steps).

Purification of Compound IV^Me: the crude product (0.8 g) was solubilized in water/acetonitrile 70/30 (v/v) mixture with 1‰ TFA (25 ml), filtered on a 0.45 μm filter and purified by RP-preparative HPLC on a PLC2250 Gilson system equipped with a DeltaPack Waters column (15 μm, 100Å, 100 × 40 mm) at 50 mL/min. flow rate. Eluents were water 1‰ TFA (A) and acetonitrile 1‰ TFA (B). Purification gradient: 0% B to 15% B in 3 min., 15% to 25% in 5 min., and 25% to 55% in 30 min. UV detection was done at 214 nm. The product is eluted at 45% of B. Pure compound IV^Me was obtained as a colourless solid in a 54% (432 mg) overall yield (4 steps).

3- Preparation of ARM165 N¹-(5-(2-((S)-1-cyclopropylethyl)-7-methyl-1-oxoisoindolin-5-yl)-4-methylthiazol-2-yl)-N⁵-(8-((2-(2,6-dioxopiperidin-3-yl)-1,3-dioxoisoindolin-4-yl)amino)-8-oxooctyl)glutaramide and ARM204 N¹-(5-(2-((S)-1-cyclopropylethyl)-7-methyl-1-oxoisoindolin-5-yl)-4-methylthiazol-2-yl)-N⁵-(8-((2-(1-methyl-2,6-dioxopiperidin-3-yl)-1,3-dioxoisoindolin-4-yl)amino)-8-oxooctyl)glutaramide (Extended Data Figs. 6 and 7A)

Synthesis: Compound I, (1.0 equiv.), and Compound IV (1.0 equiv.) for the preparation of ARM165, or Compound IV^Me (1.0 equiv.) for the preparation of ARM204, were dissolved in 1 mL of DMF and DIEA (5.0 equiv.). Then, HATU (Hexafluorophosphate Azabenzotriazole Tetramethyl Uronium, 1.0 equiv.) was added and the mixture stirred at room temperature for 6 h before an additional portion of 0.2 equiv. of HATU was introduced. After 16 h, without any treatment, the reaction media was purified on reverse phase preparative HPLC – PLC2250 Gilson (DeltaPack Waters, 15 μm, 100Å, 100 × 40 mm). Of note, ARM165 and ARM204 were prepared on a 0.37 mmol and 73 μmol scale, respectively.

Purification: in both cases, eluents were acetonitrile 1‰ TFA (B) and water 1‰ TFA (A) and elution gradient (50 mL/min.) was: 0% B during 3 min., 0% to 35% B in 7 min., 35% to 45% in 5 min., and 45% to 65% in 20 min. The product is eluted at 55% of B (ARM165), or 58% of B (ARM204). Compounds ARM165 and ARM204 were obtained as colourless solids in a 55% and 29% yield, respectively.

Statistics and Reproducibility

Sample sizes ranged from n=3 to n=7 replicates per condition for in vitro studies, n=5 to n=11 mice per condition for in vivo studies, and n=9 samples for analyses performed with primary patient samples. Sample sizes were sufficient to identify significant changes as indicated in each figure and for in vivo studies, and measurements were taken from distinct samples. Sample size calculations were not performed because sample size was chosen in light of the fact that these our in vivo models were historically highly penetrant, aggressive, and consistent with previous reports^65,66. No data were excluded from the analysis. All mice were randomized across treatment conditions. The Investigators were not blinded to allocation during experiments and outcome assessment.

Determination of statistical significance using Microsoft Excel and Prism 8.0.1 (GraphPad); unpaired student’s T-test was applied where data was normally distributed. Where data variance was high, we applied Welch’s correction to analyze significance, and pairwise comparison of data which was not normally distributed was analyzed using non-parametric Mann-Whitney test. Multiple comparison analysis testing was achieved using the one-way ANOVA test and the mean of each condition was compared with the mean of every other column. Bioluminescence imaging testing was achieved using two-way ANOVA test across the different assessed time points. Associations between PIK3CG/PIK3R5 gene expression and cytologic/genetic groups were tested using Wilcoxon test, and correlation studies were carried out using Spearman correlation. The level of significance was set at 0.05. Figures were generated using Adobe Illustrator CS6 Version 16.0.0. We verified that experiments were reproducible; where possible, experiments were performed at least twice with similar results and often in multiple cell lines. Data collection and analysis were not performed blind to the conditions of the experiments. Further information on research design is available in the Nature Research Reporting Summary linked to this article.

Extended Data

Extended Data Fig. 1:

A. Comparison of the expression levels of PIK3CG and PIK3R5 across AML and healthy tissues. P-values calculated using one-way ANOVA. Error bars represent mean + SD. B-D. Expression levels of PIK3CG and PIK3R5 in AML patients from various FAB (B) and genetic (C) subcategories, or at diagnosis versus relapse. Standard boxplot representation; median with lower and upper hinges corresponding to the first and third quartiles. The lower and upper whiskers represent the lowest and largest value within 1.5 times the lower and upper interquartiles, respectively. The points represent individual values. (D). P-values calculated using two-sided Wilcoxon test. E and F. Spearman correlation between AML cell differentiation state and the expression of PIK3CG and PIK3R5. B-F. Data generated using the BEAT-AML cohort. F. The smooth area corresponds to the 95% confidence interval of the linear regression model. Correlation between PI3K3CG or PIK3R5 expression and ssGSVA score was tested using Spearman’s rank correlation. Tests were two-sided, and p-values were corrected for multi-testing using the Benjamini & Hochberg method.

Extended Data Fig. 2:

A. DNA sanger sequencing of the PIK3CG and PIK3R5 genomic regions targeted by the CRISPR-Cas9 PIK3CG- and PIK3R5-directed guides in OCI-AML2 cells. Sequences were aligned using TIDE online tool to determine the relative efficiency of each sgRNA. B. Bioluminescence pictures of three representative mice from Figure 1J injected with OCI-AML2 infected with either a non-targeting control or a PIK3CG-directed sgRNA. Median bioluminescence is depicted in radiance on days 7, 13, 18 and 22.

Extended Data Fig. 3:

A and B. Representative growth inhibition curves, and corresponding IC50 and AUC values from Figure 2C of OCI-AML2 cells transduced with either a non-targeting control, two PIK3CG-directed, or two PIK3R5-directed sgRNAs and treated with increasing concentrations of the indicated targeted therapies or chemotherapy drugs (A), FLT3 inhibitors, gilteritinib and sorafenib, or KIT inhibitors, amuvanib and telatinib (B). Error bars represent mean ± SD of four (A) and seven (B) replicates after three days of seeding. Av. = Averaged. C. Bioluminescence pictures of three representative mice from Figure 2I injected with OCI-AML2 cells harboring a non-targeting control or PIK3CG-directed sgRNA. Median bioluminescence is depicted in radiance on days 17, 24, 27 and 31.

Extended Data Fig. 4:

A. Volcano plots of the PIK3R5-interacting protein pulled down in OCI-AML2, MV4–11, and NOMO-1 cells compared to control. B. Ranking of proteins with positive infinite ratios (only identified in PIK3R5 pull down) according to the number of peptides identified. Only proteins with a z-test score > 0.95 were included for further analysis. C. Network corresponding to Figure 3A depicting, in two colors, the PIK3R5-interacting proteins identified in two cell lines (in yellow) or all three cell lines (in orange).

Extended Data Fig. 5:

A. Correlation between gene dependency scores of AKT1 and AKT2 versus PIK3CG in a panel of non-AML cell lines. Data from DepMap CRISPR/Cas9 dependency profiling. Pearson correlation coefficient (ρ) provided to demonstrate no correlation. B. MK-2206 sensitizer and resistor topoisomerase-encoding genes, polymerase-encoding genes, and anti-apoptotic protein-encoding genes identified from a pooled drug-modifier screen conducted in OCI-AML2 cells. Gene-level scores were obtained by averaging sgRNA-level comparisons. Red dot denotes BCL2 identified as a sensitizer gene. All other genes included in the analysis are depicted as blue dots. C. Chemokine/receptor expression heatmap indicating upregulation of both CXCL12/CXCR4 in AML relative to normal tissue. Tumor expression data from TCGA database; normal expression data from GeTex expression database. Data accessed through Gepia gene expression portal.

Extended Data Fig. 6:

Synthetic scheme for the synthesis of the PIK3CG degrader, ARM165. Reagents and conditions: i) LiOH monohydrate, MeOH, H₂O, 60°C, 40 h (94%) ; ii) Boc-AOc-OH, T3P 50% in ethyl acetate, pyridine, N,N-dimethylformamide, 80°C, 16 h ; iii) TFA, CH₂Cl₂, r.t. 2 h ; iv) Glutaric anhydride, N,N-diisopropylethylamine, toluene, N,N-dimethylformamide, 110°C, 2 h (62% over 3 steps, ii-iv) ; v) Compound I, HATU, N,N-diisopropylethylamine, N,N-dimethylformamide, r.t. 16 h (55%).

Extended Data Fig. 7:

A. Representative growth inhibition curves, with IC50s and AUCs, corresponding to the Figure 5B for AML and non-AML cells treated with increasing doses of ARM204, AZ2, and ARM165. Error bars represent mean ± SD of seven replicates after three days of seeding. B. Synthetic scheme for the synthesis of non-PIK3CG-targeting control compound for ARM165, ARM204. Reagents and conditions are provided below the synthesis scheme.

Extended Data Fig. 8:

A. Representative growth inhibition curves, IC50s, and AUC values, corresponding to Figure 6A of indicated AML cells treated with increasing doses of venetoclax in presence of AZ2 or ARM165. Error bars represent mean ± SD of seven replicates after three days of seeding. B-C. Growth inhibition curves, IC50s, and AUC values of OCI-AML2 cells treated with increasing doses of cytarabine, daunorubicin, or venetoclax in combination with 500nM ARM165. Error bars represent mean ± SD of seven technical replicates after three days of seeding in three biological repeats. P-values calculated using one-way ANOVA.

Extended Data Fig. 9:

A-B. Toxicity profile of ARM165 treatment in naive mice. Mice were treated with IV injection of 0.051mg/kg ARM165 for seven consecutive days. Individual mouse weight was measured daily (n=5 mice per group) (A) and the proportion of each indicated hematopoietic cell fraction (B) was established in blood using an MS9 instrument, and in bone marrow (BM) and spleen (SP) by flow cytometry (n=10 mice per group). Error bars represent mean ± SD. P-values calculated using Mann-Whitney.