Anticancer drug resistance: an update and perspective

Abstract

Driver mutations promote initiation and progression of cancer. Pharmacological treatment can inhibit the action of the mutant protein; however, drug resistance almost invariably emerges. Multiple studies revealed that cancer drug resistance is based upon a plethora of distinct mechanisms. Drug resistance mutations can occur in the same protein or in different proteins; as well as in the same pathway or in parallel pathways, bypassing the intercepted signaling. The dilemma that the clinical oncologist is facing is that not all the genomic alterations as well as alterations in the tumor microenvironment that facilitate cancer cell proliferation are known, and neither are the alterations that are likely to promote metastasis. For example, the common KRas^G12C driver mutation emerges in different cancers. Most occur in NSCLC, but some occur, albeit to a lower extent, in colorectal cancer and pancreatic ductal carcinoma. The responses to KRas^G12C inhibitors are variable and fall into three categories, (i) new point mutations in KRas, or multiple copies of KRAS G12C which lead to higher expression level of the mutant protein; (ii) mutations in genes other than KRAS; (iii) original cancer transitioning to other cancer(s). Resistance to adagrasib, an experimental antitumor agent exerting its cytotoxic effect as a covalent inhibitor of the G12C KRas, indicated that half of the cases present multiple KRas mutations as well as allele amplification. Redundant or parallel pathways included MET amplification; emerging driver mutations in NRAS, BRAF, MAP2K1, and RET; gene fusion events in ALK, RET, BRAF, RAF1, and FGFR3; and loss-of-function mutations in NF1 and PTEN tumor suppressors.

In the current review we discuss the molecular mechanisms underlying drug resistance while focusing on those emerging to common targeted cancer drivers. We also address questions of why cancers with a common driver mutation are unlikely to evolve a common drug resistance mechanism, and whether one can predict the likely mechanisms that the tumor cell may develop. These vastly important and tantalizing questions in drug discovery, and broadly in precision medicine, are the focus of our present review. We end with our perspective, which calls for target combinations to be selected and prioritized with the help of the emerging massive compute power which enables artificial intelligence, and the increased gathering of data to overcome its insatiable needs.

1. Introduction

Cancer can evolve from somatic mutations (cancer.sanger.ac.uk) (Akdemir et al., 2020b; Martincorena and Campbell, 2015; Tate et al., 2019). These mutations frequently accumulate in oncogenes and tumor suppressor genes (Morjaria, 2021). Mutations accumulate over a relatively long period. Some mutations emerge randomly, some are inherited, whereas others develop due to mutagens. Estimates of the number of mutations that are required for a normal human cell to progress to advanced cancer, including by mathematical models based on the relation between age and incidence, vary (Tomasetti et al., 2015).

A driver mutation confers upon cancer cells a growth advantage for its neoplastic transformation (Pon and Marra, 2015). It differs from passenger mutations which do not contribute to the development and progression of the cancer. Only a small fraction of mutations identified in a cancer patient occur in driver genes. Drivers are commonly identified by their high frequencies of occurrence (Brown et al., 2019; Chen et al., 2020). However, they can also be rare (Nussinov et al., 2020b, 2019c). Drivers commonly lead to an observable conformational change. At the same time, ‘latent drivers’ may also contribute to a gain of function, even though the associated conformational change may be unobservable, and on their own, their contribution is minor (Nussinov et al., 2019a; Nussinov and Tsai, 2015). Rare latent drivers can also be preexisting drug resistance mutations. Driver mutations can propel cancer initiation and progression. They characterize molecular profiles of tumors and help in predicting clinical outcomes for the patients. They are sought after as druggable targets and are used in making therapeutic decisions (Nussinov et al., 2019b; Saleem et al., 2019; Zsakai et al., 2019). Their frequency points to their common occurrence. Not all driver mutations are drug resistance mutations and not all drug resistance mutations occur in driver genes (Beckman and Loeb, 2020; Loeb et al., 2019; Reiter et al., 2019). Driver mutations activate the protein even in the absence of an incoming activation signal, making it an escape regulatory control. On the other hand, drug resistance mutations can make a protein escape inhibition by certain drugs, commonly via changes in the active site shape or surface. In addition, alternative mechanisms of chemoresistance occur that are not associated with qualitative and/or quantitative changes in the driver genes; these include: impaired drug influx, enhanced drug efflux via multidrug resistance pumps of the ATP-Binding Cassette (ABC) superfamily, drug compartmentalization away from its target protein, metabolic drug inactivation, as well as induction of anti-apoptotic mechanisms (Assaraf et al., 2019; Gonen and Assaraf, 2012; Li et al., 2016; Shahar and Larisch, 2020; Wang et al., 2021; Wijdeven et al., 2016; Zhitomirsky and Assaraf, 2016). Resistance to imatinib (STI571), an orthosteric ATP-competitive drug that inhibits oncogenic mutants of Bcr-Abl fusion tyrosine kinase, is a classic example (Burchert, 2007; Gao et al., 2021) (see below). Mutations in Bcr-Abl’s ATP-binding site or the activation loop can result in steric clashes blocking imatinib binding, as observed upon T315I substitution (Nicolini et al., 2013; Zhang et al., 2010). The T315I mutation also blocks second-generation tyrosine kinase inhibitors (TKIs) such as nilotinib (AMN107) and dasatinib (BMS-354825) (O’Hare et al., 2005; Tamai et al., 2018). Another example is the case of epidermal growth factor receptor (EGFR) mutations T790M and C797S. In a second resistance scenario, gene amplification can spur an alternative signaling pathway (Genovese et al., 2017). Mesenchymal-epithelial transition factor (MET or c-MET, a.k.a. HGFR, hepatocyte growth factor receptor) encoded by the MET gene (Aissa et al., 2021), or other receptor tyrosine kinase (RTK) signaling pathways that can bypass the targeted RTK (Harbinski et al., 2012; Wilson et al., 2012), such as IGF-1R (insulin-like growth factor 1 receptor) (Crudden et al., 2015), FGFR1 (fibroblast growth factor receptor 1), HER2 (human epidermal growth factor receptor 2), and AXL (tyrosine-protein kinase receptor UFO), which share downstream pathways, PI3K/AKT/mTOR (PI3K, phosphatidylinositol 3-kinase; AKT, protein kinase B; mTOR, mechanistic target of rapamycin) and MAPK (mitogen-activated protein kinase), provide examples. Activation of EMT (epithelial-to-mesenchymal transition), NF-κB (nuclear factor kappa B), or AXL either by chemotherapy or a single small molecule inhibitor (Kim et al., 2019; Taniguchi et al., 2019; Zhu et al., 2019), dysregulation of growth factor cell signaling, a major driver of most human cancers (Gillis and McLeod, 2016), and loss of function of tumor suppressor genes (Gao et al., 2021) have all suggested that drug resistant cancer cells can exploit diverse and distinct mechanisms (Aissa et al., 2021).

In the current review we ask: Are drug resistance mechanisms that emerge following targeting of common driver mutations also common? That is, would targeting a certain Ras driver, e.g., KRas^G12C, in different patients lead to the same drug resistance modality? This question portends critical drug treatment decisions (Nussinov et al., 2014). Below, we first outline our premise. We then briefly overview some drug resistance mechanisms. We follow with those exploited by human tumors harboring a common KRas mutation as examples. We further discuss why the mechanisms of resistance of common driver mutations are expected to differ. We next comment on genomic heterogeneity and the pre-existence of both driver and resistance mutations, and finally take up the key question of whether one can forecast the likely emerging resistance mechanisms. Drug resistance mechanisms are chiefly determined by genomics, pre-existing and evolving mutations (Aissa et al., 2021; Bar-Zeev et al., 2017; Friedman, 2016; Levin et al., 2021; Lim and Ma, 2019; Prieto-Vila et al., 2019; Taylor et al., 2015), and cell type and state which determine the gene expression status. Single cell transcriptomics can capture rare drug resistant cells in a tumor clone (Bhang et al., 2015; Hata et al., 2016; La Monica et al., 2019). Loss of H3K4me3 (histone 3 lysine 4 trimethylation) in EGFR mutant in erlotinib (Tarceva)-treated cells (Lantermann et al., 2015) along with changes in chromatin-modifying proteins HDAC9 (histone deacetylase 9), NCOR1 (nuclear receptor co-repressor 1), MLL1 (a histone methyltransferase), and EED (embryonic ectoderm development protein) demonstrate large transcriptomic alterations in drug resistant clones, emphasizing the cardinal role of epigenetics and chromatin organization (Nussinov et al., 2021a). Transcriptional changes associated with drug resistance, including those emerging from gene copy number mutations can seal cell fate, by allowing, or restricting, cancer re-growth, and as such can improve the predictive power of genomic sequencing for rare drug resistant mutations (Aissa et al., 2021; Sun et al., 2019). We end with a discussion of the almost limitless number of possible target combinations to address the many potential drug resistance mechanisms. In our view, the way to select and prioritize the combinations is by artificial intelligence (AI) strategies powered by the fast-increasing massive computational power and clinical and inhibitor databases (Kim, 2021; Nguyen et al., 2021). With the growing size of model parameters deep learning requires data-intensive computations. The rapid increase in the volume of data coupled with advanced hardware and software methods could break the performance wall of traditional approaches and provide more reliable target combination to be submitted for testing in the expensive clinical trials.

2. Our premise: cells with common driver mutations are unlikely to develop common drug resistance mechanisms

The reasons for this premise include: (i) the mutant protein may be in different cell (tissue) types and (ii) the cell state may differ. The cell state can be described by, e.g., its gene expression profile, abundance of other proteins in the corresponding pathway, its post-translational modifications status, and morphology. These are impacted by environmental cues and the cell developmental status. Even for a specific tissue, or cell type where the mutation emerged, the time window is a key consideration (Mitchell et al., 2018). (iii) Often overlooked, yet, with a profound influence on the mechanism of the emerging drug resistance are the identity and distinct patterns of background mutations. Background mutations (Beckman and Loeb, 2020; Nussinov et al., 2021b; Saito et al., 2020; Vasan et al., 2019b; Zhang et al., 2021), either pre-existing the emergence of cancer or emerging early in its evolution (Haupt et al., 2021; Johnson et al., 2014) are likely to vary among patients. Considering that tumors may initially respond to treatment but, as not all the neoplastic cells are decimated, the minor population, or rare remaining drug resistant cells housing these mutations can seed cancer regrowth (Figure 1). The mutations are likely to include driver mutations, or combinations of strong drivers, latent drivers, or weak drivers (Zhang et al., 2021). Rare cells are also likely to accommodate a preexisting drug resistance mutation, which may, or may not be a driver. In one example, it reactivates cellular proliferation through hypermutated Rb (retinoblastoma), which plays a key role in negatively regulating cell cycle progression from the G1 to S-phase through two positive regulators of cell cycle entry, E2F transcription factors and cyclin dependent kinases, and the AKT/mTOR, which is also required for cell cycle progression from G1 to S phase (Kumar et al., 2016). Since the expression levels of the corresponding genes need to be high, coupled mutations in chromatin remodelers is commonly observed (Akdemir et al., 2020a). Additional contributors not discussed here include parameters such as cell metabolism (Gremke et al., 2020; Knoechel and Aster, 2015; Zaal and Berkers, 2018). Resistance may not necessarily involve genetic changes. Some cancer cells may survive chemotherapy by exploiting nutrients in their tumor microenvironment (TME) (van Gastel et al., 2020). Thus, parameters related to the TME also play cardinal role [e.g., Refs. (Getzenberg and Coffey, 2011; Pontiggia et al., 2012; Wojtkowiak et al., 2011)].

Figure 1.

The role of heterogeneity in drug resistance. An example of cells in the tissue containing a vast number of common driver mutant (e.g., KRas4B G12C) and a small number of rare drug resistance mutant cells (lower left). When a drug is taken, all cells are decimated except those with rare resistance mutation for the particular drug, resulting that those cells with the rare mutation proliferate (middle). A subsequent drug will remove the rare drug resistance mutant cells in later stage of chemotherapy (lower right). However, simultaneously taking both drugs have a better chance of success.

As to background mutations, at the time of diagnosis, a tumor may already accumulate more than 1 billion cancer cells (Offord, 2020). The effective mutation rate was calculated to be 6 × 10⁻⁷ per gene per generation leading to an accumulation of mutational diversity beyond a critical tumor size (Beckman and Loeb, 2020). This is in line with the expectation that all drug resistant mutations pre-exist in the tumor in at least one cell at diagnosis, even though not all driver mutations are resistance mutations, and not all resistance mutations occur in drivers (Beckman and Loeb, 2020; Loeb et al., 2019; Reiter et al., 2019). Even rare cells harboring such mutations may proliferate, thereby seeding a tumor subclone resistant to a single therapy. However, the probability that they harbor co-existing mutations that simultaneously resist polypharmacology, that is, a combination of drugs, is low. These are expected to emerge during treatment. This argues that genetic alterations – pre-existing and emerging during tumor evolution – are a critical factor in drug resistance (Beckman and Loeb, 2020; Loeb et al., 2019; Reiter et al., 2019), including mutations in chromatin remodelers which play critical roles in protein expression. Rare mutations are vastly important contributors to cancer initiation and progression (Nussinov et al., 2019a, b; Nussinov et al., 2019c), and to resistance (Beckman and Loeb, 2020), making their identification a vital aim (Nussinov et al., 2020b).

3. Drug resistance mechanisms emerging to common targeted drivers reflect cell-type specific cancer evolution

Several excellent reviews on drug resistance mechanisms were published in the last few years that describe the multifactorial nature of cancer drug resistance [e.g., Refs. (Aleksakhina et al., 2019; Andrei et al., 2020; Assaraf et al., 2019; Das et al., 2021; Du et al., 2020; Dunnett-Kane et al., 2021; Friedman, 2016; Garraway and Janne, 2012; Gottesman, 2002; Haider et al., 2020; Housman et al., 2014; Hussein et al., 2021; Jiang et al., 2020; Lee et al., 2020; Li et al., 2020; Lim and Ma, 2019; Long et al., 2020; Narayanan et al., 2020; Rolfo et al., 2014; Ruan et al., 2020; Sabnis and Bivona, 2019; Sarmento-Ribeiro et al., 2019; Sciarrillo et al., 2020; Shahar and Larisch, 2020; Sun et al., 2019; Wang et al., 2019; Yan et al., 2020; Zhong and Virshup, 2020)]. Thus, we herein only touch on several key determinants, and proceed to our focus on drug resistance mechanisms that were observed to emerge following the drugging of common driver mutations in common target proteins in patients with the same or different tumor types, e.g., lung, colorectal, and breast cancer. Key determinants of drug resistance include (Vasan et al., 2019a) (i) Tumor burden, or tumor load, defined as the total amount of tumor (cells/mass) distributed in the patients’ body, or the sum of the longest diameters of all measurable lesions (Bousquet et al., 2013; Czarnecka et al., 2019), (ii) Tumor growth kinetics, defined as the ratio of the slope of tumor growth before treatment and the slope of tumor growth on treatment between the nadir and disease progression, calculated for each patient (Le Tourneau et al., 2012), (iii) Tumor heterogeneity (Lawrence et al., 2013), which captures the distinct phenotypic profiles of tumor cells, e.g., morphology, gene expression, metabolism, motility, proliferation, and (iv) Metastatic potential, the immune system, the TME, druggability, and more (Vasan et al., 2019a). Cancer drug resistance is a frequently a multifactorial phenomenon and hence can occur via multiple mechanisms, including multidrug resistance (Buck et al., 2021; Gottesman, 2002; Gottesman et al., 2002; Li et al., 2016; Robey et al., 2018; Su et al., 2021; Wang et al., 2021) and detoxification (Lee et al., 2012), cell death inhibiting (apoptosis suppression) (Eberle et al., 2007; Jiang et al., 2021; Shahar and Larisch, 2020), altered drug metabolism, epigenetic (Jiang et al., 2021), enhancing DNA repair (Chiappa et al., 2021; Pecoraro et al., 2021), gene amplification (Assaraf et al., 1989; Bram et al., 2007; Mansoori et al., 2017) as well as drug sequestration in lysosomes or vesicles away from the cellular drug target (Goler-Baron et al., 2012; Ifergan et al., 2005; Zhitomirsky and Assaraf, 2015, 2016).

At the same time, recent observations on the consequences of drugging KRas^G12C (Figure 2) indicate that resistance mechanisms emerging in different cells vary, questioning the effectiveness and wisdom of adopting a common drug regime for a common driver mutation (Awad et al., 2021; Dana-Farber Cancer Institute, 2021). Considering the high frequency of common drivers, this dilemma of effective treatment for cancers harboring common drivers has far reaching consequences. The quandary facing the physician stems from the realization that the alterations in the genomic and cellular environment that permitted the cancer to proliferate are unknown, and neither are those that are likely to evolve. The common KRas^G12C mutation emerged in different cancer types: even though most occurred in NSCLC (non-small-cell lung cancer), some were detected in colorectal cancer and pancreatic ductal adenocarcinoma (Seton-Rogers, 2020). The results from an analysis of patients with the KRas^G12C mutation are revealing (Awad et al., 2021). They indicated that the response of patients to KRas^G12C inhibitors led to cancer evolution that fell into the following categories (Figure 2), (i) Cancers with new mutations in KRas, at KRas^G13, KRas^R68, KRas^H95, and KRas^Y96, or multiple copies of KRas^G12C which lead to a higher expression level of the mutant protein; (ii) Mutations in genes other than KRAS, e.g., BRAF, MET, ALK, RET, MAP2K1; (iii) Transitioning to other cancers, e.g., transitioning from lung adenocarcinomas to squamous cell carcinomas. Overall, resistance to adagrasib (MRTX849), which earned an FDA ‘Breakthrough Therapy’ designation for KRas^G12C NSCLC emerged in 17 out of the 38 patients, with half of these presenting multiple coincident mechanisms (Awad et al., 2021). Mutations included G12D/R/V/W, G13D, Q61H, R68S, H95D/Q/R, Y96C, and allele amplification. Redundant or parallel pathways included MET amplification; emerging driver mutations in NRAS, BRAF, MAP2K1, and RET; gene fusions events in ALK, RET, BRAF, RAF1, and FGFR3; and loss-of-function mutations in NF1 and PTEN tumor suppressors.

Figure 2.

The mechanism of KRas4B^G12C drug resistance. Examples of covalent drugs, adagrasib (MRTX849) and sotorasib (AMG510), bound to Cys12 of GDP-bound KRas4B (upper panel). The outcomes of drug resistance for a common driver mutation G12C in KRas4B (lower panel).

Cell type and background mutational load with the subsequent emergence and selection of new drivers critically influence the resistance mechanism. The pre-existing mutational burden already has more than a single driver, even though only the KRas^G12C mutation was common in all patients. These mutations could be in the same gene (i.e., in cis), which is the case most of the time, or in different ones (in trans). In a pan-cancer analysis of 60,954 cancer samples Saito et al. (Saito et al., 2020) observed that in cis double/multiple driver mutations are more common than those in trans, in line with earlier observations (Kohsaka et al., 2017; Madsen et al., 2019). Especially, the background mutations need not be strong drivers. Functionally weak, or rare mutations, can confer enhanced oncogenicity in combination. Why the preference for the same gene versus different gene driver combinations? We believe that a second (or multiple) mutation emerging in the same gene has higher chances to elicit more potent outcome as compared to different genes, since many possible routes and key pathway nodes can engage in the same or parallel (or redundant) pathways (Nussinov et al., 2021a; Nussinov et al., 2020a). These relatively common mutational scenarios can clarify the clonal selection of suboptimal mutations that when coupled, can promote cancer emergence, tumor cell proliferation, progression and drug resistance.

The residual resistant population seeds regrowth of tumor cells that no longer respond to the drugs (Cree and Charlton, 2017). Even for a single cell, the resistance mechanism can be expected to be complex, encompassing both gene regulation at the chromatin level (Akdemir et al., 2020b) and resistance mutations. Drug resistance may be due to specific mutations, which in some cases it is, but in many others rapid resistance originates from multiple mechanisms (Di Nicolantonio et al., 2005; Glaysher et al., 2010; Glaysher et al., 2009), and especially as already detailed recently (Su et al., 2021), including also pre-existing resistance mutations. A tumor can compensate for EGFR (HER1 in humans) blockade through the activation of alternative signaling pathways such as amplification of MET as well as through changes in the TME (van der Wekken et al., 2016), or resistance to EGFR inhibition (van der Wekken et al., 2016), or neuroendocrine differentiation (Oser et al., 2015). Below we discuss regulation, the role of chromatin status and alternative signaling pathways which should be considered in combinatorial drug regimens as well as the heterogeneity of tumor clones and pre-existence of rare resistance mutations which might be captured by single-cell transcriptional changes associated with drug tolerance (Aissa et al., 2021). We also consider how can personalized strategies be predicted from pre-existing genomic and proteomic profiles.

4. The frequencies of common driver mutations vary with cell type and age

As examples, we focus on Ras drivers. Ras mutations have emerged in 27% of all human cancers (Hobbs et al., 2016). They occur in all Ras isoforms albeit at different frequencies (Altmuller et al., 2017; Bera et al., 2019; Blons et al., 2014; Bournet et al., 2016; Burd et al., 2014; Der, 2014; Lampson et al., 2013; Lu et al., 2016; Munoz-Maldonado et al., 2019; Newlaczyl, 2016; Nussinov et al., 2018; Nussinov et al., 2021c; Pellicer, 2011; Poulin et al., 2019; Prior et al., 2020; Prior et al., 2012; Pylayeva-Gupta et al., 2011; Rajasekharan and Raman, 2013; Raso, 2020; Rezaei Adariani et al., 2021; Russo et al., 2014; Simanshu et al., 2017; Stark et al., 2012; Stephens et al., 2017; Tsai et al., 2015; Xu et al., 2013; Yang and Kim, 2018). HRas is the least (4%), KRas is the most highly mutated (85%) whereas NRas is at 11% (Table 1). 98% of the mutations are at G12, G13, and Q61 positions in KRas. However, the mutation frequencies at these locations vary among the isoforms. G12 mutations predominate in KRas, whereas Q61 mutations are rare. On the other hand, Q61 is the dominant alteration in NRas. The relative frequencies of the other mutations and their ranked order vary as well among the isoforms. Even for the same isoform, the frequencies differ across cancer types. For example, NRas Q61 and G12 mutations are frequent in melanoma and acute myeloid leukemia (AML), respectively. On the other hand, among the KRas G12 mutations, G12D is the most common in pancreatic ductal adenocarcinoma (PDAC) followed by G12V, but not G13 and Q61. G12D is the predominant G12 mutation in KRas (41%) and NRas (52%), and G12V in HRas (57%) (Table 1). As to G13, G13D is the most frequent in KRas (89%) and NRas (50%) but it is rare in HRas (3%), where G13R (85%) dominates (Hobbs et al., 2016). KRas A146T mutation occurs in certain cancers but not in PDAC, and A146T is associated with better overall survival than G12 mutations of KRas. In general, mechanisms driving allelic selection in cancer reflect the activation mechanism of the mutation implicating the aggressiveness of the specific cancer (Poulin et al., 2019). G12V and G12D interfere with GTPase-activating protein (GAP)-assisted hydrolysis. Q61L mutation also reduces KRas intrinsic hydrolysis and GAP’s action and increases intrinsic nucleotide exchange (Hobbs et al., 2016; Smith et al., 2013). Likewise, G13D also interferes with GAP-mediated hydrolysis (Smith et al., 2013). Recent comprehensive characterization of Ras mutations in colon and rectal cancers in old and young patients (Serebriiskii et al., 2019) showed remarkable differences as well. Epidemiological stratification indicated that cancers expressing different mutant forms of KRas exhibit distinct clinical behaviors (Gibbs et al., 1984; Haigis, 2017; Hobbs et al., 2016; Ogura et al., 2013; Poulin et al., 2019; Witkiewicz et al., 2015).

Table 1.

Ras mutations in human cancer. The parentheses indicate the predominant mutations.

Ras isoform	Frequency of mutations	G12 mutation	G13 mutation	Q61 mutation	Others
KRas	85%	83% (G12D 41%)	14% (G13D 89%)	2% (Q61H 58%)	2%
NRas	11%	23% (G12D 52%)	12% (G13D 50%)	63% (Q61R 47%)	3%
HRas	4%	33% (G12V 57%)	27% (G13R 85%)	37% (Q61R 43%)	3%

Tumors harbor tens to thousands of mutations (Mateo et al., 2020). The vast majority are passenger mutations, not observed to encode functional advantage. Analysis of over 9000 tumors from 33 tissues identified 258 genes with driver mutations (Bailey et al., 2018), 142 of which were from a single tumor type, and 87 from several types. The number of driver mutations in a single tumor type varies between 2 (kidney) to 55 (uterine). Especially relevant to our discussion, every patient has a unique combination of mutations and copy number variants (Rubio-Perez et al., 2015). These emphasize that tumor type and state, e.g., age as observed in Ras driver mutations (Serebriiskii et al., 2019). The differential background mutation load in the tumor cell are pivotal in tumor evolution (Gerstung et al., 2020), and we likewise expect in the emerging drug resistance mechanism.

5. Cell type, background mutation load and the complexity of forecasting emerging resistance mechanism

Vigorous divisions of a cell harboring a mutation that confers a selective advantage can make the cell an initiator of a fast-growing mutant clone. By the time the cancer is diagnosed, the tumor may already consist of ≥10⁹ cells the replicating genomes of which accumulated a large number of mutations (Alberts et al., 2002; Beckman and Loeb, 2020; Offord, 2020). The emerging heterogeneous clonal populations can be viewed as substrates for cancer cell evolution, including those related to drug resistance. Thus, the rapid emergence of resistance to drugs, driven by the background mutational load, is not surprising. High drug doses may block the pathway. However, mutations that circumvent drug blockage and permit signaling – through the same pathway or via a by-pass pathway – can rapidly produce drug resistance, posing a massive therapeutic challenge. This underscores the critical question of forecasting drug resistance mechanism. Common driver mutations in a common protein, do not necessarily imply common resistance mechanisms (Figure 1).

To tackle drug resistance (National Cancer Institute, 2020; Volm and Efferth, 2015), efforts have been focusing on identifying mutations that impair drug binding in the active site cavity of the target protein, including the often so-called ‘gatekeeper’ residues, and on designing second generation drugs. Chemotherapy regimens mostly considered a combination of an orthosteric or allosteric drug targeting the mutant protein combined with drugs aiming at a downstream protein in the same pathway, or in an alternative signaling pathway that can take over in drug resistance (Nussinov et al., 2021b; Zhang et al., 2021, 2020). These additional proteins and pathways are largely selected based on experience, and drug availability (Cheng et al., 2019). EGFR mutants in NSCLC (~11%, (Neel and Bivona, 2017)) can provide an example (Lim and Ma, 2019). Secondary resistance mutations on the EGFR kinase domain can reduce the binding affinity of first- and second-generation drugs. A bypass pathway can take over. Sequentially prescribed drugs can temper resistance, only to cede to a small but fast growing subclonal population of resistant cells that becomes the dominant species. Additional examples include mutations in B-Raf (~7%), and gene rearrangements involving ALK and ROS1 (1–2%). These are protein kinases. The mutations hyperactivate downstream signaling promoting cell growth, proliferation, and survival (Neel and Bivona, 2017). Individualized genetic network approaches have been shown to reveal new therapeutic vulnerabilities in 6,700 cancer genomes promising new venue in such endeavors (Liu et al., 2020). Further, mutations that perturb protein-protein interaction networks were shown to be highly correlated with patient survival and drug resistance (and sensitivity) (Cheng et al., 2021).

The emerging mechanism of drug resistance depends on the cell type, cell state, and the set of background and evolving mutational load. A key factor that varies across cell types and dynamically fluctuates with cell state is epigenetics and chromatin accessibility of the regulatory regions of the respective genes (Akdemir et al., 2020b). Relatively lower density of chromatin can flag genes that can become available for transcription. In drug resistance, the expression level, and the potency of the mutation(s), which depends on the activation mechanism, are critical. Extremely high expression level – even on its own – can already drive cancer. However, it is not only the expression level of the mutant protein which is vital. The expression levels of all other protein nodes in the relevant signaling pathway should be at sufficiently high levels to allow the signal to propagate down to the cell cycle, to activate the corresponding transcription factors in gene expression.

6. Common driver mutation, genetic and epigenetic clone heterogeneity

Genomic instability and intratumoral heterogeneity are the driving force in drug resistance (Lim and Ma, 2019). Mutational burden, including in chromatin remodeler genes, and gene copy number alterations facilitate tumor evolution (Gerstung et al., 2020) and drug resistance. Even for a common driver mutation, such as the KRas^G12C discussed above, the different tumor cell types (tissues) and states (in different patients) imply heterogeneity in epigenetics (Baylin and Jones, 2011; Bennett and Licht, 2018; Dawson and Kouzarides, 2012; Easwaran et al., 2014; Fardi et al., 2018; Feinberg et al., 2016; Flavahan et al., 2017; Lee et al., 2020; Nebbioso et al., 2018; Suva et al., 2013; You and Jones, 2012), chromatin organization and gene accessibility (Makova and Hardison, 2015). Heterogeneity implies that resistance mutations are likely to exist prior to treatment, including allosteric ones like the T315I (Abl1 isoform IA; T334I for Abl1 isoform IB) substitution that alters the ATP-binding site in the Bcr-Abl fusion tyrosine kinase discussed above which blocks the binding of imatinib, as well as second-generation TKIs such as nilotinib and dasatinib (Nicolini et al., 2013; O’Hare et al., 2005; Tamai et al., 2018; Zhang et al., 2010) (Figure 3). Screening drug libraries discovered the allosteric drug GNF-2 (Zhang et al., 2010) and second-generation GNF-5, as well as asciminib (ABL001) that bind to the Bcr-Abl C-terminal myristate pocket and re-sensitize it to imatinib and nilotinib. The allosteric agent GNF-5, together with imatinib, an ATP-competitive drug, overcame the drug resistance induced by T315I (Adrian et al., 2006; Wylie et al., 2017; Zhang et al., 2010). Thus, even though mutations at the myristate site confer resistance to asciminib, they retain sensitivity to nilotinib (or imatinib, or dasatinib), making their combined treatment successful (Schoepfer et al., 2018). Allosteric asciminib together with orthosteric ponatinib (AP24534) were found to strongly counter drug resistance mutations Y253H (Y272H for IB) and E255V (E274V for IB) (Eide et al., 2019). Thus, rare cells with rare pre-existing drug resistant mutations from the repertoire of background mutations can take over, and the mechanism of drug resistance that is again based on the T315I substitution, may or may not display common features.

Figure 3.

Inhibitors for Bcr-Abl fusion tyrosine kinase. Molecular structures of the drugs bound to the kinase domain of Abl (upper panel). Orthosteric ATP-competitive drugs, such as imatinib (STI571), nilotinib (AMN107), dasatinib (BMS-354825), and ponatinib (AP24534), can inhibit oncogenic Bcr-Abl kinase activity. However, drug resistance mutations in ATP-binding site, such as gatekeeper mutation T315I (Abl1 isoform IA; T334I for Abl1 isoform IB), or the activation loop can change conformation of the active site, preventing the orthosteric drugs from the oncogenic inhibition. Allosteric drugs bound to the myristate pocket, such as GNF-2, GNF-5 and asciminib (ABL001), can re-sensitize the orthosteric drugs, overcoming the drug resistant. Drug structures were obtained from PubChem (https://pubchem.ncbi.nlm.nih.gov), a public chemical database at the National Library of Medicine (NLM) (Kim et al., 2021). Examples shown for the crystal structures of Abl1 autoinhibition, drug-bound kinase domains, and drug-induced Abl1 autoinhibition (lower panels).

7. Conclusions and perspectives

Resistance mechanisms include drug inactivation, qualitative and quantitative drug target alterations, drug efflux, enhanced DNA damage repair, cell death inhibition (Housman et al., 2014) as well as drug sequestration in lysosomes or vesicles away from the drug target. Recently, crosstalk with the TME has also been suggested to guide resistance evolution (Bhattacharya et al., 2021). However, the complexity of TME is daunting, and is only now beginning to be understood (Jin and Jin, 2020) and its pharmacology explored (Kaemmerer et al., 2021).

Resistance to kinase inhibitors is especially potent in the deadliest common cancers, such as pancreatic, lung, breast, colorectal, and prostate cancers (Geyer et al., 2006; Jonker et al., 2007; Knight et al., 2010; Miller et al., 2007; Moore et al., 2007; Sandler et al., 2006; Shepherd et al., 2005). Treatments which were given to those patients deemed with higher chances of favorable therapeutic response based on certain markers, fared better; however, treatments with only a single kinase inhibitor failed over time (Knight et al., 2010). This was not surprising since already early on, imatinib therapeutic regimen for Bcr-Abl fusion protein in chronic myeloid leukemia (CML) (Gorre et al., 2001), was observed to eventually lead to resistance mutations, such as the allosteric T315I mutation that interferes with imatinib binding as discussed above. ERBB2 (a.k.a. HER2) in breast cancer (Hudis, 2007), EGFR in lung cancer (Ciardiello and Tortora, 2008), and KRas in lung and colorectal cancer (Karapetis et al., 2008; Pao et al., 2005b) provide additional examples, as well as others [e.g., Refs. (Chen et al., 2004; Cools et al., 2003; Pao et al., 2005a; Tamborini et al., 2004)]. Emerging drug resistance can also involve combinations of mutations in the targeted kinase as well as with mutations in other proteins. These can involve constitutive mutations in a substitute kinase that can replace the drugged kinase, or alterations that lead to its overexpression. One such example involves the MET receptor tyrosine kinase in lung cancer cells becoming resistant to inhibitors following treatment of oncogenic mutant EGFR (Engelman et al., 2007). Alternatively, a mutation in the targeted kinase can couple with a mutation in a phosphatase whose inactivation can substitute the blocked kinase, as in the case of EGFR inhibitors in breast cancer cells (Sergina et al., 2007). Here, drug resistant cells respond by down regulating the phosphatase, thereby decreasing the potency of the kinase inhibitor. In another example, drug resistance to PI3K inhibitors can involve inactivation of phosphatase and tensin homolog (PTEN). The outcome of PTEN inactivation parallels that of PI3K activation (Carrera and Anderson, 2019; Jang et al., 2021; Lien et al., 2017; Papa and Pandolfi, 2019). Both result in an increase of the population of PIP₃ (phosphatidylinositol 3,4,5-trisphosphate) signaling lipid molecules. Tumor cells also often overexpress redundant oncogenic signaling via multiple kinases, as in the case of multiple RTKs (Stommel et al., 2007). Activation of a downstream kinase can also decrease the action of an inhibitor, enabling the bypassing of the targeted kinase (Karapetis et al., 2008; Mellinghoff et al., 2005; Pao et al., 2005b). Thus, resistance can take place via multiple mechanisms including mutations in the respective targeted driver oncogene (Neel and Bivona, 2017), mutations that activate a protein downstream or a parallel pathway (Nussinov and Tsai, 2014; Nussinov et al., 2017), and pro-survival mutations via a different pathway. An independent class of drug resistance encompasses histological transformation from one tumor cell lineage to another, as discussed above for lung cancer.

Malignant cells within a tumor are not genetically identical, and this heterogeneity poses major problems in cancer therapy (Gatenby and Brown, 2018). The mutations may not emerge following drug treatment. Instead, at least some, are likely to preexist it. However, being rare in the population, they are undetected. Decimation of the sensitive population by first generation drugs leads to their proliferation, and drugs that could otherwise have prevented the cancers from recurring are largely ineffectual. As tumors evolve, they become more heterogeneous, with cancer cells acquiring new mutations that make them resistant to certain treatments. A comprehensive single-molecule sequencing study revealed patterns of preexisting drug resistance mutations that suggested treatment strategies in Philadelphia-positive (Ph+) leukemias (Schmitt et al., 2018). The study pointed to Abl1 resistance mutations unlikely present at the time of diagnosis, explaining the success of targeted therapy; at the same time, patients with Ph+ acute lymphoblastic leukemia (ALL) were observed to often harbor multiple preexisting resistant cells with single mutants which proliferated, following treatment with Abl1 inhibitor of cells harboring sensitive mutations.

Mutations in specific signaling nodes, such as kinases or receptors, often couple with alterations in chromatin structure and the transcription machinery. Mutations in the pioneer transcription factor, forkhead box protein A1 (FOXA1, a.k.a. HNF-3α, hepatocyte nuclear factor 3α), provide an example (Arruabarrena-Aristorena et al., 2020). The study uncovered mutations in two different regions that are localized at the C-terminal forkhead domain of FOXA1 in estrogen receptor-positive (ER+) breast cancers. The mutations in the Wing2 subdomain (residues 247–269) and SY242CS mutation in the third β strand, away from the Wing2 region, can lead to resistance to aromatase inhibitors in distinct ways. Wing2 or SY242CS mutations do not require estrogen receptor activation. Modeling suggested that the SY242CS mutation changes FOXA1 protein shape, apparently enabling chromatin exposure at altered sites, leading to altered gene expression. Wing2 mutations promote the ability of the cells to respond to the estrogen when available. To date, mechanistic details are unavailable. Importantly, FOXA1 mutations co-occur with other mutations in breast cancer-related genes (e.g., PIK3CA, AKT1, or ARID1A). Also notable is the involvement of alternative transcription factor motifs observed for other FOXA1 mutations (i.e., R219S and R219C) in prostate cancer as well as for other transcription factors (Arruabarrena-Aristorena et al., 2020). These lead to distinct patterns in chromatin accessibility.

Going forward (Knight et al., 2010), our perspective to alleviate resistance calls for targeting combinations of key nodes in tumor cells, adopting a multi-drug strategy that concomitantly targets multiple proteins (Pao et al., 2005b). The combination should consider targeting one kinase with multiple drugs, i.e., competitive and allosteric, targeting nodes in the signaling network in addition to the mutant protein, selecting drug targets from the same or from bypass (so-called redundant) pathways, and chromatin modifiers and RNA polymerases. The challenge is to predict the best combinations of targets and ranking them with some score. With the vast number of possible combinations, this is a formidable task. However, the evolution of artificial intelligence techniques (Nussinov et al., 2021a), and the increased gathering of data which their deep learning requires for reliable prediction, have been made possible by massive advances in compute intensive capabilities. That data is clinical, structural (of proteins and chromatin) and includes tested inhibitors in other settings, which can be used for drug repurposing. Altogether, such a challenging future perspective can become a reality. We believe that progress will come through massive computational involvement.

To conclude, treatment decisions in precision medicine are not made based on the location of the tumor, but on genomic sequencing and identification of the mutations. Here we reasoned that even though knowledge of the mutation is critical, the cell, or tumor type, is important to consider as well. Genomic sequences are essential to obtain all relevant data, chiefly of the driver mutations in question as well as the identity of the background mutational load. Data indicate that co-existing double/multiple mutations are more common in the same protein as compared to different proteins and pathways, and they demonstrate higher sensitivity to drugs. However, more complete information – if, or when, available – including epigenetic and chromatin accessibility, and clinical manifestation, can be powerful in forecasting the mechanisms underlying the expected emergence of drug resistance. The technology is not yet available for broad use, but future cell-specific developments portend more accurate predictions. Tumor evolution is a vastly complex process that is still not entirely understood. Here we highlighted the role of mutational load in tumor initiation, progression, and chemotherapy resistance, and argued for enlisting and levying the increasingly available massive computational power.