(M1130-02-09) Deciphering Response Patterns to AKT Inhibitor MK2206 in Breast Cancer Patients: Unveiling Signature Profiles

Purpose: The AKT/mTOR/PI3K signaling pathway is central to tumor cell development, survival, and proliferation, rendering it an appealing target for drug therapy [1, 2, 3]. A clinically advanced AKT inhibitor (MK2206) was previously evaluated in I-SPY 2, a Phase 2 adaptive trial for women with early stage, locally advanced, aggressive breast cancer. The trial was adaptive in that patients were assigned preferentially to the arm where their cancer subtype would be most likely to respond. Subtype definitions were: HER2+, HR+/HER2-, and triple negative (HR-/HER2-) [4]. Of 96 patients enrolled in the I-SPY 2 MK2206 arm, 35 were treatment-responsive. An analysis examining a panel of candidate genes found that none specifically associated with drug sensitivity after adjusting for HR and HER2 status [4]. We assessed the ability of our graph-based signature extraction approach to accurately classify patients from publicly available transcriptomic data (Fig. 1) and compared the performance of our approach vs classic single-protein classification in distinguishing responders from non-responders. Our method using a higher-order signature extraction algorithm resulted in good classification performance AUC > 0.7 (Fig. 2A).

Methods: The dataset GSE149322 contains gene expression microarray data from 96 patients enrolled in the I-SPY 2 MK2206 arm [4]. We binned each expression into 5 levels and generated a protein network graph for each patient. We performed subgraph extraction to recurrent signatures, sets of three network-linked nodes at specific expression levels, within responder and non-responder subsets. The composite presence or absence of these signatures in each patient graph represents their signature profile. Signatures with a high differential frequency between the two groups were tested, same as the protein features previous [4], as predictors of MK2206 sensitivity in logistic regression models.

Results: Signatures were more predictive of drug response compared to single gene features in decision tree classification (Fig. 2A). Two signatures were significantly associated with drug resistance (negative odds-ratios) in logistic models even after adjusting for HR/HER2 status (Fig 2B). This is an improvement over the individual gene features where, although some were significant in subtype-agnostic models, they all lost their significance models adjusting for HR/HER2 status. All genes of the significant signatures are connected to disease endpoints by at most two mediators in the interaction network (Fig 3), supporting the relevancy of these signatures in breast cancer. Signature 1: MAP2K5-very.low, RB1CC1-low, SIRT1-low. Signature 2: THAP3-low, THAP7-low, UTS2-very.low

Conclusion: Modelling gene expression within the physiological interaction networks in which proteins function enables the extraction of higher-order signatures, better identifying biomarkers indicative of disease response and treatment efficacy. In this case, we were able to identify biomarkers which were robust across breast cancer subtypes when traditional single-gene features failed to do so. This method has the potential to identify patient subtypes at a high granularity to facilitate precise, personalised drug intervention. We identified patterns that simultaneously depict causal outcomes by capturing interacting proteins specific to breast cancer.

References: 1. Brown, J. S. & Banerji, U. Maximising the potential of AKT inhibitors as anti-cancer treatments. Pharm. Ther. 172, 101–115 (2017).
2. Liao, Y. & Hung, M. C. Physiological regulation of Akt activity and stability. Am. J. Transl. Res. 2, 19–42 (2010).
3. Scheid, M. P. & Woodgett, J. R. Unravelling the activation mechanisms of protein kinase B/Akt. FEBS Lett. 546, 108–112 (2003).
4. Wolf DM, Yau C, Wulfkuhle J, Brown-Swigart L et al. Mechanism of action biomarkers predicting response to AKT inhibition in the I-SPY 2 breast cancer trial. NPJ Breast Cancer (2020).


Fig. 1 A Diagram describing the process of signature extraction from patients gene expression profiles.


Fig. 2 A) ROC curves for classifier performance using individual gene expressions as input features (yellow) and signatures as input features (blue). B) Signatures with increased odds of drug resistant (logistic regression).


Fig. 3 Proteins encoded by genes in significant signatures (blue) are either directly linked, or separated by one intermediary protein (black), to breast cancer endpoints (red) in the physical interaction network.