Identification of a Selective G1-Phase Benzimidazolone Inhibitor by a Senescence-Targeted Virtual Screen Using Artificial Neural Networks12

Cellular senescence is a barrier to tumorigenesis in normal cells, and tumor cells undergo senescence responses to genotoxic stimuli, which is a potential target phenotype for cancer therapy. However, in this setting, mixedmode responses are common with apoptosis the dominant effect. Hence, more selective senescence inducers are required. Here we report a machine learning–based in silico screen to identify potential senescence agonists. We built profiles of differentially affected biological process networks from expression data obtained under induced telomere dysfunction conditions in colorectal cancer cells and matched these to a panel of 17 protein targets with confirmatory screening data in PubChem. We trained a neural network using 3517 compounds identified as active or inactive against these targets. The resulting classification model was used to screen a virtual library of ~2M lead-like compounds. One hundred and forty-seven virtual hits were acquired for validation in growth inhibition and senescence-associated β-galactosidase assays. Among the found hits, a benzimidazolone compound, CB-20903630, had low micromolar IC50 for growth inhibition of HCT116 cells and selectively induced senescence-associated β-galactosidase activity in the entire treated cell population without cytotoxicity or apoptosis induction. Growth suppression was mediated by G1 blockade involving increased p21 expression and suppressed cyclin B1, CDK1, and CDC25C. In addition, the compound inhibited growth of multicellular spheroids and caused severe retardation of population kinetics in long-term treatments. Preliminary structure-activity and structure clustering analyses are reported, and expression analysis of CB-20903630 against other cell cycle suppressor compounds suggested a PI3K/AKT-inhibitor–like profile in normal cells, with different pathways affected in cancer cells. Neoplasia (2015) 17, 704–715 UK Cambridge Institute Core Grant, and Hutchison Whampoa (M.N.). Conflict of Interest Statement: The authors have nothing to disclose. Received 2 April 2015; Revised 28 August 2015; Accepted 31 August 2015 ©2015TheAuthors. PublishedbyElsevier Inc. onbehalf ofNeoplasiaPress, Inc.This is anopen access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/). 1476-5586 http://dx.doi.org/10.1016/j.neo.2015.08.009 Neoplasia Vol. 17, No. 9, 2015 A Selective G1-Phase Benzimidazolone Inhibitor Bilsland et al. 705 Introduction Cellular senescence in normal cells is an irreversible cell cycle arrest which is involved in cellular aging and tissue maintenance, and which is induced by critically shortened telomeres at the end of replicative lifespan. Oxidative damage and oncogene activation accelerate both telomere shortening and senescence induction [1]. Therefore, senescence is considered to be a barrier to tumorigenesis which cancer cells must bypass to acquire a transformed phenotype [2,3]. Many cancer cells retain the capacity to undergo senescence-like growth arrest in response to agents including chemotherapeutics and ionizing radiation in addition to many targeted agents [4]. Hence, despite inactivation of some key pathways, many tumor cells retain the ability to exit the cell cycle under appropriate treatments. Thus, latent senescence signaling may persist in tumors [5]. There is substantial interest in senescence induction as a therapeutic outcome in cancer. However, senescence involves multiple processes including telomere homeostasis, DNA damage and inflammatory signaling, chromatin regulation, and metabolism [6,7]. Interaction of these with the diverse mutational backgrounds of cancer cells adds further complexity in attempting to define the best targets for therapeutic intervention. It seems likely that a spectrum of senescence-like responses is possible in cancer cells depending on induction agent and signaling environment [8,9]. Given limitations in current knowledge, phenotypic screening is attractive both for compound and pathway discovery focused on senescence [10–12]. Suitable phenotypic markers for assay development include p21 and p16 levels, the senescence-associated secretory phenotype, senescence-associated β-galactosidase (SA-β-gal) staining, senescence-associated heterochromatin foci, and altered morphology [1]. However, although many agents elicit senescence, responses obtained are often restricted to subsets of cells, with apoptotic cell death dominant [13]. To evaluate senescence induction as an anticancer modality will require identification of senescence agonists which are substantially more selective than currently available tools [14]. Without detailed knowledge of targets, the screening challenge is not simply identification of compounds which can cause senescence; rather, stratification of the most selective compounds among many expected partial actives is critical. Identification of enriched libraries would be beneficial before initiating a screening campaign. We reasoned that virtual screening might identify such an enriched set. Ligand-based virtual screening is of increasing interest in the construction of activity models, ranging from well-defined target binding studies [15] to more complex scenarios such as modeling of experimental microsomal stability results [16], and a wide variety of platforms and datasets are now available [17]. Another major goal is to identify new compounds with activity against a given target based on feature recognition [18]. In either case, abstraction of chemical structure information into a set of numerical descriptors is critical. These must provide detailed representation of the chemical and property space for a given compound set [19]. An assumption is that a relation can be made between these “fingerprints” and a classifier (active/inactive) or known quantity such as IC50. Machine learning methods such as neural networks [18,20] or support vector machines [21,22] provide a powerful approach. Feature recognition rules are learned from a training set with known activity; trained models are then simulated against a new compound set of unknown activity. Here we report a virtual screen using an artificial neural network ensemble trained by the scaled conjugate gradient descent method [23] using compounds identified from pooled PubChem screens [24,25] against a panel of senescence-related targets. Targets were selected by matching available screens to cellular “process networks profiles” obtained by functional enrichment analysis of expression data in colorectal cancer cells with induced telomere dysfunction. The trained ensemble was used to classify a library of around 2M lead-like compounds, leading to identification of a benzimidazolone compound with lowmicromolar IC50 which selectively induces G1 blockade and SA-β-gal without causing apoptosis. Preliminary structure/activity relationships (SARs) and clustering studies are reported.