Discovering novel SNPs that are correlated with patient outcome in a Singaporean cancer patient cohort treated with gemcitabine-based chemotherapy

Gemcitabine (2′-2′ difluorodeoxycytidine) is a deoxycytidine analogue with antitumor activity against a variety of solid tumors such as non-small cell lung cancer (NSCLC), breast cancer [1] and pancreatic cancer [2]. Gemcitabine requires phosphorylation to mono-, di-, and triphosphates (dFdCTP) to be active. This mechanism results in a unique pattern of self-potentiation of the drug and when this drug is incorporated into the DNA during replication, it causes chain termination. Gemcitabine also has multiple intracellular targets. Up- or downregulation of these targets may confer resistance to this drug.

Wider availability and lower costs of genome and expression profile sequencing made application of those techniques in clinical practice feasible; thus, the scientific question of how patient-specific mutations and chromosomal aberrations influence personal clinical outcomes via biomolecular mechanisms has become acute [3, 4]. For example, pharmacogenetics studies published in the last decades have provided evidence that Single Nucleotide Polymorphisms (SNPs) can causally influence patient outcome such as drug response and toxicity after drug intervention [5]. Most SNPs associated with patient outcome have been found in genes involved in the drug pharmacology i.e. affecting drug transport, metabolism and/or activity with drugs. Soo et al. tested 26 SNPs from nine genes that are already known to be directly associated with gemcitabine transport, metabolism and activity [6]. They found several SNPs that were associated with patient outcome in Singaporean NSCLC patients treated with gemcitabine [6]. However, a systematic approach to investigate the relationship between gene variants and patient outcome is still lacking. Therefore, the purpose of this study is to develop a systematic pathway approach to accurately and efficiently predict novel non-synonymous SNPs (nsSNPs) that could be causative to gemcitabine-based chemotherapy treatment outcome in Singaporean NSCLC patients. After detailed SNP analysis, we prioritized the SNPs based on their importance in protein function and molecular mechanism. From the top-ranking SNPs, we specifically selected six final candidate SNPs for clinical validation. We genotyped these six SNPs in a Singaporean patient cohort and have found that three out of the six SNPs correlated with patient outcome.

In this study, three out of the six candidate SNPs were confirmed to be associated with NSCLC patient outcome i.e. OS, PFS and side effect. To the best of our knowledge, this is the first study showing association of ABCG2 Q141K (rs2231142), SLC29A3 S158F (rs780668) and POLR2A N764K (rs2228130) with NSCLC patient outcome treated with gemcitabine-based chemotherapy. ABCG2 belongs to the ABCG subfamily and ABC transporter superfamily. The ABCG family has five members i.e. ABCG2, ABCG1, ABCG4, ABCG5 and ABCG8. ABCG2 consists of a nucleotide-binding domain (NBD) in the amino terminus followed by six putative transmembrane domains (Fig. 3a). The ABCG2 Q141K SNP is located at the NBD in the cytoplasmic part of the protein. The c.421A allele frequency of ABCG2 Q141K is known as one of the common SNPs in Asian people (about 26-35%) [32]. Moreover, this SNP has been shown to be associated with increased risk of gout [33]. When we created our own detailed multiple sequence alignment using all members in the ABCG family, we found that glutamine in this position is well conserved in ABCG2 orthologs but not in other members in the family, therefore Q141 can be considered as an ABCG2-subfamily specific conserved residue (Fig. 3b). ABCG2 is the only member in this family that is not involved in cholesterol efflux but it mediates the efflux of a wide range of xenobiotics including gemcitabine, using ATP as an energy source [34]. There is in vitro evidence that ABCG2 Q141K decreases efflux activity and increases intracellular gemcitabine levels and it has been known to be associated with impaired ABCG2 activity by lowering protein expression level or decreasing ATPase activity [35]. The study supports the observation that ABCG2 itself plays a role in decreasing intracellular concentration of gemcitabine. Another in vitro study demonstrates significantly worse overall survival for carriers of the ABCG2 421A-allele treated with platinum-based drugs [36]. Mizuarai et al. described that the ATPase activity of the Q141K variant was reduced approximately 1.3-fold compared to the activity of the wild type ABCG2 in polarized LLC-PK1 cell lines, resulting in increased drug accumulation and decreased drug efflux in the variant ABCG2-expressing cells [36]. According to BLAST against PDB, a crystal structure of Malk, the ATP subunit of the maltose transporter from E.coli (PDB:1Q12 chain A) [37] was the top hit with a E-value of 2.0E-17. We used this template to do homology modeling of the NBD region (position 41-299) of ABCG2 using MODELLER. The ABCG2 model is shown in Fig. 3c. We used this model to calculate the stability change upon mutation by FoldX and found the average free energy changes (ddG) when mutating Q to K at position 141 of ABCG2 to be 1.93 kcal/mol with a standard deviation (SD) of 0.10 kcal/mol. This suggests that the SNP has a destabilizing effect on the protein structure which is in agreement with a recent finding that Q141 causes instability in the NBD [38]. The SNP is located in the loop region which is relatively near the ATP binding site of the dimer and changing the neutral side-chain glutamine to positively-charged side-chain lysine may affect the scaffold of the neighboring ATP binding site formed by the homodimer (Fig. 3c). Therefore, it can be proposed that if a patient has this variant and is treated with gemcitabine, efflux of gemcitabine can be diminished resulting in an increase in the intracellular concentration of gemcitabine in cancer cells and it is thus more effective at killing cancer cells. However, since normal cells also have this SNP which causes accumulation of the drug and other substrates exported by this protein, this SNP is also linked to increased toxicity in normal cells (Fig. 4).

Our study also showed for the first time that patients who were carrying either the CT or TT of SLC29A3 473 C > T (rs780668) were associated with increased OS. SLC29A3 belongs to the equilibrative nucleoside transporter (ENT) family, responsible for passive nucleoside transport and has 11 transmembrane helices (TMs) within the nucleoside transporter domain (Pfam:PF01733) (Fig. 5a). SLC29A3 S158 is likely to be a subfamily specific residue since serine at this position is fully conserved from human to fish among SLC29A3 orthologs but cannot be seen in other members in the family (Fig. 5b). This residue may relate to its unique function compared to other members in that it seems to function in the inner membrane of mitochondria and/or in the lysosome which requires an acidic pH environment and the position of the SNP seems to localize outside of the inner membrane of mitochondria [39]. For homology modeling of SLC29A3, we had a problem retrieving correct TMs models using MODELLER with loop refinement. The template used was a crystal structure of the glycerol-3-phosphate transporter from E.Coli (PDB:1PW4) chain A which has only 12% identity to our query. This problem is common when we used MODELLER which is more suitable for modeling soluble proteins than membrane proteins. Therefore, we used another software called Memoir [40] which is a homology modelling algorithm designed specifically for membrane proteins. A homology model from this software retrieved 11 TMs with long N-terminus and long loop regions between TM6 and 7 and the correct SNP’s position on the 3D structure (Fig. 5c). According to FoldX, SLC29A3 S158F was predicted to have a significant destabilizing effect with average ddG of 2.18 kcal/mol which could be explained by the strong change from the polar side chain serine to the larger and more hydrophobic side chain phenylalanine (Fig. 5d). Besides potential involvement of the conserved wildtype serine in the transport process, increasing the hydrophobicity through the mutation at the outside interface with the membrane may result in deeper insertion of the affected helix in the membrane. SLC29A3 can transport gemcitabine into organelles, e.g., mitochondria [39]. Moreover, SLC29A3 could be involved in the mitochondrial toxicity of nucleoside drugs [8]. Since SLC29A3 S158F is most likely to affect protein function, it may have a significant impact on transporting gemcitabine into mitochondria.

POLR2A (DNA-directed RNA polymerase II subunit RPB1) encodes the largest subunit (out of 12 subunits) of RNA polymerase II (Pol II) which catalyzes the RNA synthesis from DNA. POLR2A contains a carboxy terminal domain (CTD) which is composed of 52 heptapeptide repeats that are necessary for the polymerase activity (Fig. 6a). POLR2A N764K is located in the RNA polymerase’s domain 4 (Pfam:PF05000) which is also known as the funnel domain. The N764 is highly conserved among orthologs (Fig. 6b) and both PolyPhen-2 and SIFT analyses predicted that the SNP affects protein function (Additional file 1: Table S2). We created a homology model of POLR2A without the CTD using a crystal structure of yeast RNA polymerase II (PDB:1I3Q chain A) as a template (%identity = 50.3%) (Fig. 6c). The SNP is in the loop region and it was predicted to have a destabilizing effect by FoldX (average ddG 1.13 kcal/mol) (Fig. 6d). This could be due to the longer and charged lysine side chain causing a change in the conformation of the local loop structure. In our study, interestingly, wild type (CC) is found to be associated with higher grade 3 or 4 thrombocytopenia when compared to the CT variant (Table 3). Further analysis is needed to understand the mechanism for this. There is an evidence that dFdCTP is incorporated into RNA which is concentration- and time-dependent, resulting in inhibition of RNA synthesis [41]. In human parental NSCLC cells with a different inherent gemcitabine resistance, sensitivity to gemcitabine was related to differences in RNA corporation [42]. Since the SNP is found to be strongly deleterious from our analyses, it would be interesting to investigate whether POLR2A itself plays a role in the gemcitabine pharmacologic pathway and whether POLR2A N764K SNP has any implications on RNA synthesis.

NT5C2 (5′-nucleotidase, cytosolic II) encodes a hydrolase that serves a crucial role in cellular purine metabolism by acting primarily on inosine 5′-monophosphate (IMP) or guanosine monophosphate (GMP). The 5′-nucleotidase is a huge family of enzymes that catalyze the dephosphorylation of deoxy- and ribonucleoside monosphosphates to nucleoside analogues and inorganic phosphates [43, 44]. NT5C2 is known to dephosphorylate monophosphorylated gemcitabine. The NT5C2 D549 is the first charged amino acid in the last 13 acidic residues on the C-terminus of the protein (Fig. 7a). A study in 1999 proved the importance of the highly acidic C-terminus in NT5C2 protein using cDNA constructs encoding proteins lacking either N- or C-terminus and obtained the kinetic and molecular characteristics of the recombinant proteins [45]. When the last 13 acidic residues on the C-terminus were eliminated, there was a drastic reduction in the catalytic competence of the enzyme by lowering both the substrate affinity and the specific productivity. Furthermore, the capability of the protein to form a tetramer was significantly compromised. From the experiment above, it was concluded that the region of glutamic and aspartic acid residues in the C-terminus of the enzyme is necessary for the complete function of NT5C2. Another study, on bovine NT5C2 found that the C-terminus is perhaps involved with the modulation of enzyme function [46]. In our detailed SNP analysis, the D549 is well conserved among orthologs of NT5C2 (Fig. 7b) but is not conserved throughout the NT5C enzyme family (data not shown). D549E was predicted as deleterious by both PolyPhen-2 and SIFT. However, D and E are both negatively charged amino acids and the following acidic C-terminus shows a mixed pattern of the two so it is not mechanistically clear how this mutation could alter the enzyme activity of NT5C2. In agreement with this, we do not find significant correlations with survival or toxicity for NT5C2 D549E.

HELB T980I and CTDP1 T221M, unlike the other four SNPs that we predicted to directly affect the gemcitabine pharmacologic pathway, were selected because they could potentially affect the pathway indirectly in order to establish if more remotely related SNPs are still useful candidates for further studies. According to our criteria, HELB T980I passed three criteria i.e. significant results in SIFT, PolyPhen-2 and high MAF in Singaporean Chinese (Additional file 1: Table S2) while CTDP1 T221M passed two criteria (borderline) i.e. the SNP is in a domain region and has significant FoldX stability change (Additional file 1: Table S2). HELB T980I lies close to the phosphorylation sites in the phosphorylation regulated subcellular localization (PSLD) domain at the C-terminus of the protein. The PSLD domain has been suggested to play a significant role in regulating the subcellular localization of HELB [47]. The same study also suggested a possible role of HELB in DNA repair. Another study gives further evidence that HELB is recruited by Replication Protein A to mitigate replication stress [48]. So, we hypothesized that the SNP could impair HELB’s DNA replication stress mitigating effect of gemcitabine-induced DNA damage which could lead to reduced recovery from gemcitabine-induced replication stress, and hence to a better anti-cancer activity of gemcitabine. CTDP1 dephosphorylates a phosphorylated C-terminal domain of POLR2A to facilitate Pol II recycling for transcription [49]. There is evidence suggesting that CTDP1 may play a role in DNA damage response as well [50]. However, from our clinical results, we could not see any correlation between these two SNPs and the patient outcome. Therefore, it can be suggested that when we choose SNPs in the final step, we should expect a high chance for correlation with drug response only for those that are affecting the pharmacologic pathway directly.

Overall, our bioinformatics approach can be used to select a small number of potential causative SNPs that can be rationalized with molecular mechanisms of their effects. Our workflow can also be applied to any pathway of interest that could affect other phenotypes. Nowadays, using GWAS to find SNPs associated with common diseases or adverse drug reactions are the norm but even though thousands of case samples have been used, one review found that about 30% of the results lead to a null finding [51]. Furthermore, significant GWAS hits are often not easily explainable to have functional effects e.g. when they are intronic or synonymous SNPs. Although our study may be limited in sample size, we managed to find that three out of six final candidate SNPs are associated with patient outcome. However, more association studies and possibly molecular and cellular studies are needed to further establish the value of these three SNPs as biomarkers of patient outcome. While we have to acknowledge that this approach of selective filters does not guarantee to find all SNPs involved in a studied phenotype, the main benefit is that one can reduce the space of possible candidates to a small experimentally tractable number with higher chance of being relevant. We do believe that other candidate SNPs that passed three out of five criteria would also be potential candidates to be tested further. We hope our approach and findings will pave the way to more meaningful biomarkers and personalized treatment options in the future.