Virtual high-throughput screens identifying hPK-M2 inhibitors: Exploration of model extrapolation

Highlights

• Trained predictive SVM classification/regression models with available hPK-M2 data.

• Screened PubChem Compound Database, ∼72 million compounds, for lead candidates.

• Select candidates identified and purchased for activity validation experiments.

• Candidate screening hit-rate: 1st round – 1/12 (8.13%), 2nd round – 6/11 (54.5%).

• 2nd round result analysis for extrapolation effects: stepwise extrapolation advised.

Abstract

Glycolysis with PK-M2 occurs typically in anaerobic conditions and atypically in aerobic conditions, which is known as the Warburg effect. The Warburg effect is found in many oncogenic situations and is believed to provide energy and biomass for oncogenesis to persist. The work presented targets human PK-M2 (hPK-M2) in a virtual high-throughput screen to identify new inhibitors and leads for further study. In the initial screen, one of the 12 candidates selected for experimental validation showed biological activity (hit-rate = 8.13%). In the second screen with retrained models, six of 11 candidates selected for experimental validation showed biological activity (hit-rate: 54.5%). Additionally, four different scaffolds were identified for further analysis when examining the tested candidates and compounds in the training data. Finally, extrapolation was necessary to identify a sufficient number of candidates to test in the second screen. Examination of the results suggested stepwise extrapolation to maximize efficiency.

Introduction

Pyruvate kinase (PK; EC 2.7.1.40) is one of the final enzymes participating in glycolysis: it transfers phosphate groups from phosphoenolpyruvate (PEP) to adenosine diphosphate (ADP) to produce pyruvate and adenosine triphosphate (ATP). Two genes encode pyruvate kinase, each producing different isoenzymes: the PKL gene produces tissue-specific PK in the liver (PKL) and erythrocytes (PKR) (Noguchi et al., 1987), while the PKM gene produces the PK-M1 and PK-M2 isoform depending on the exon splicing, 9 or 10 respectively (Noguchi et al., 1986). Furthermore, PK-M2 can be an active tetramer, allosterically activated by fructose-1,6-bisphosphate (Ashizawa et al., 1991), or a nearly-inactive dimer (Eigenbrodt et al., 1992). Glycolysis with PK-M1 occurs in aerobic conditions, yielding much more energy (net production of 36 ATP/glucose), in mature differentiated muscle and brain cells (Nelson and Cox, 2012, Takenaka et al., 1991). Glycolysis with tetrameric PK-M2 occurs in anaerobic environments, yielding less energy (net production of 2 ATP/glucose), predominantly in dividing and embryonic/fetal tissue (Nelson and Cox, 2012, Takenaka et al., 1991, Eigenbrodt et al., 1992). During normal development, the M2 isoenzyme is highly-expressed initially and gradually replaced by the other three isoenzymes (Dombrauckas et al., 2005, Ao et al., 2017).

In tumors and oncogenic environments, PK-M2 has two different roles. The first is ATP production by the active tetramer in oxygen-rich environments (Elbers et al., 1991, Hacker et al., 1998), known as the Warburg effect (Warburg, 1926). The second is biomass accumulation as PK-M2 tetramers breakdown to less active dimers in hypoxic and acidic environments (Kumar et al., 2010). The accumulation of byproducts shunts biomass for use in processes critical for cell proliferation and development, such as the production of nucleic acids, proteins, and lipids (Mazurek et al., 1997, Mazurek et al., 2001, Vander Heiden et al., 2009; Li et al., 2014). Furthermore, PK-M2 may play a role in anti-cancer drug resistance (Yoo et al., 2004, Martinez-Balibrea et al., 2009). When PK-M2 is replaced with PK-M1 (Trialists et al., 2005, Wang et al., 2012) or absent (Ao et al., 2017), cancer cells cannot sustain growth or start apoptosis (Ao et al., 2017, Wang et al., 2012).

Given its presence and functionality in different cancer types such as gastrointestinal (Schneider and Schulze, 2003, Hardt et al., 2000, Oremek et al., 1997), kidney (Brinck et al., 1994), ovarian (Chao et al., 2017), lung (Schneider et al., 2003), and breast (Ibsen et al., 1982) cancer, it is increasingly considered a target for cancer treatments. Exploration of possible PK-M2 treatments include: small-molecule activators (Auld et al., 2018, Boxer et al., 2011, Kung et al., 2012), small-molecule inhibitors (Spoden et al., 2008, Auld et al., 2018, Vander Heiden et al., 2010, Chen et al., 2011, Anastasiou et al., 2012, Yacovan et al., 2012, Guo et al., 2013, Parnell et al., 2013, Xu et al., 2014), down-regulation (Pandita et al., 2014, Wong et al., 2014, Wang et al., 2012), and switching the expressed isoenzyme form to PK-M1 (Liu et al., 2014).

There are two relatively common methods to identify hPK-M2 treatment candidates: high-throughput screening (HTS) and computational analysis. HTS was often used to identify drug leads in compound libraries (Auld et al., 2018, Boxer et al., 2009, Kung et al., 2012, Xu et al., 2014, Guo et al., 2013, Vander Heiden et al., 2010) while computational analysis was used to probe structure and interaction (Kalyaanamoorthy and Chen, 2011, Anastasiou et al., 2012, Guo et al., 2013). Eventually, HTS and computational analysis were used complementarily (Yacovan et al., 2012): computational analysis via virtual HTS (vHTS) identified candidates to focus HTS resources on while HTS-confirmed leads are computationally analyzed identify ways to enhance activity via structural modification. However, much of the work is proprietary (Yacovan et al., 2012) and unavailable for review in the open literature. The presented work uses open-sourced, published software and methodologies that take advantage of the technological advancements since 2012 to screen a more extensive database (PubChem Compound: 72 million).

This work is one in a series to characterize and test the effectiveness of the authors’ previously presented approach (Chen and Visco, 2016, Chen and Visco, 2017, Chen et al., 2018) with different protein/ligand systems and datasets of varying sizes, activity classification distributions, and quality. Previously studied targets include Cathepsin L (Chen and Visco, 2016), clotting factor XIIa (Chen and Visco, 2017), and complement factor C1s (Chen et al., 2018). PK-M2 was chosen for this work because of its prevalence and role in tumors and cancers metabolism (Vander Heiden et al., 2009). Additionally, this is the first instance where activators, instead of inhibitors, are identified by the approach.

The approach uses structural features in compounds to create classification- and activity-predicting models. This does not mean the approach identifies structural causes of biological activity. However, inferences may be drawn by examining the structural features used in the models (Kayello et al., 2014) and found in the experimentally tested candidates (Chen and Visco, 2016, Chen and Visco, 2017).