Resistance to targeted therapies as a multifactorial, gradual adaptation to inhibitor specific selective pressures

Despite high initial efficacy, targeted therapies eventually fail in advanced cancers, as tumors develop resistance and relapse. In contrast to the substantial body of research on the molecular mechanisms of resistance, understanding of how resistance evolves remains limited. Using an experimental model of ALK positive NSCLC, we explored the evolution of resistance to different clinical ALK inhibitors. We found that resistance can originate from heterogeneous, weakly resistant subpopulations with variable sensitivity to different ALK inhibitors. Instead of the commonly assumed stochastic single hit (epi) mutational transition, or drug-induced reprogramming, we found evidence for a hybrid scenario involving the gradual, multifactorial adaptation to the inhibitors through acquisition of multiple cooperating genetic and epigenetic adaptive changes. Additionally, we found that during this adaptation tumor cells might present unique, temporally restricted collateral sensitivities, absent in therapy naïve or fully resistant cells, suggesting the potential for new therapeutic interventions, directed against evolving resistance.

D espite inducing strong clinical responses, inhibitors that target abnormal activities of oncogenic tyrosine kinases (TKIs), including those directed against oncogenic ALK signaling in non-small cell lung cancers (NSCLCs), are rarely curative in advanced disease 1 . As ALK-TKIs typically fail to eradicate all of the tumor cells, residual tumors eventually acquire resistance and relapse.
In contrast to the large body of work on deciphering the molecular mechanisms of resistance, understanding of its evolutionary mechanisms is more limited. Based on quantitative analyses and the detection of resistance-associated mutations, resistance is often assumed to arise due to selective expansion of pre-existent rare, fully resistant subpopulations [2][3][4][5] . On the other hand, a growing body of experimental studies suggest that resistance can emerge de novo from drug-tolerant persister (DTP) cells, which can maintain residual disease and serve as a substrate for mutational or epigenetic conversions to fully resistant phenotypes 6 . Finally, as therapies can induce adaptive phenotypic changes on shorter time scales, acquired resistance has also been viewed from a differentiation/reprograming paradigm [7][8][9] .
Motivated by interest in evolutionary-informed therapy scheduling, we investigated the origin and dynamics that underlie the development of therapy resistance of an in vitro model of ALK + NSCLC, the patient-derived NCI-H3122 cell line. Upon exposure to different ALK-TKIs, these cells rapidly and predictably develop strong drug resistance. Resistance originates de novo, from weakly resistant heterogeneous subpopulations, which differ in fitness when exposed to different ALK-TKIs. Levels of resistance gradually increase under therapy, through acquisition of multiple cooperating genetic and epigenetic mechanisms, through TKIspecific phenotypic evolutionary trajectories. In contrast to therapy-naive or fully resistant cells, these evolving populations show strong collateral sensitivity to the dual epidermal growth factor receptor EGFR/HER2 inhibitor lapatinib, suggesting a temporally restricted opportunity to interfere with the development of resistance.

Results
Different ALK-TKIs select for distinct resistant phenotypes. To understand the evolution of resistance to different TKIs, we focused on the ALK + NSCLC cell line, NCI-H3122, which has been well-characterized in multiple mechanistic studies. Starting from resistant cell lines derived through a dose-escalation protocol in our previous study 10 , we continued dose escalation, eventually selecting for cells capable of growing in high, clinically relevant concentrations of the drugs (up to 1 µM crizotinib, 4 µM lorlatinib, 2 µM alectinib, and 200 nM ceritinib). Cells with evolved resistance to ALK-TKI (erALK-TKI) displayed strong collateral resistance toward other ALK-TKIs, with IC50's 5-100× higher than therapy naïve controls ( Fig. 1a and Supplementary Table 1). Despite the similar shift in IC50, erALK-TKI cells selected by different inhibitors displayed divergent responses to higher drug concentrations. Consistent with the clinical efficacy of alectinib and lorlatinib as a second-line therapy after failure of crizotinib 11,12 , high concentrations of alectinib and lorlatinib strongly inhibited cells with evolved resistance to crizotinib (erCriz) and ceritinib (erCer). The resistance was partially or completely maintained after drug holiday both in vitro (Supplementary Fig. 1a) and in vivo (Fig. 1b, c). Resistance was not an artifact of the dose-escalation protocols, as H3122 cells developed compatible resistance levels after 2-4 months of acute exposure to clinically relevant concentrations of the ALK-TKIs (Fig. 1d). Similar to the resistant cell lines derived by gradual exposure, resistant phenotypes in cell lines derived by acute exposure to ALK-TKIs were largely heritable ( Supplementary Fig. 1b). Next, we asked whether the differences in cross-sensitivities of erALK-TKI cells towards different ALK inhibitors are attributable to the choice of specific ALK inhibitor. To this end, we examined the sensitivities of three independently derived (acute exposure protocol) cell lines for each of the ALK-TKIs used. Consistent with previous findings (Fig. 1a), erLor and erAlec cells demonstrated stronger resistance to high concentrations of different ALK-TKIs, compared with erCriz and erCer cell lines (Fig. 1e).
Next, we asked whether the differences in ALK-TKI crosssensitivity of cells selected with different ALK-TKIs are associated with specific changes in EML4-ALK-dependent signaling pathways. We found that, in the absence of ALK-TKI, all of the erCriz cell lines displayed increased phosphorylation of EML4-ALK, which was in most cases associated with elevated total EML4-ALK protein levels, as well as increased STAT3 phosphorylation (Fig. 1f). In contrast, all of the erLor cell lines displayed reduced EML4-ALK and STAT3 phosphorylation. Instead, they expressed higher levels of EGFR, HER2, or both. To assess the phenotypes more globally, we examined mRNA expression levels of 230 cancer-related genes using the Nanostring nCounter GX human cancer reference panel. Principal component analyses (PCAs) and hierarchical clustering for mRNA expression were largely consistent with the immunoblot evaluation of phosphorylation ( Supplementary Fig. 2a, b). erCer, evolved through acute selection, as well as acutely and gradually evolved erCriz and erLor cells formed distinct clusters; in contrast, phenotypes of erAlec were more diverse.
To test whether the observed differences reflected different proportions of shared phenotypic subpopulations or populationwide phenotypic changes, we performed single-cell RNA sequencing of erALK-TKI cell lines, focusing on two cell lines per specific ALK-TKI, with the highest divergence in PCA analysis of NanoString data ( Supplementary Fig. 2a). Uniform manifold approximation and projection (UMAP) 13 dimension reduction of single-cell expression data revealed relative phenotypic homogeneity within individual erALK-TKI lines, with a high degree of similarity among cell lines derived with the same inhibitor (Fig. 1g). Thus, despite a degree of stochasticity, reflective of evolutionary contingencies 14 , acquired resistance to specific ALK-TKIs is associated with phenotypes that are convergent within the same inhibitor, but divergent between different inhibitors.
Resistance originates from diverse tolerant subpopulations. Acquired resistance is often assumed to reflect a simple expansion of therapy-resistant subpopulations 15,16 . Whereas the observed predictable distinctions between resistant phenotypes, selected in different ALK-TKIs contradicts this notion, we decided to interrogate the pre-existence of fully resistant phenotypes more directly. To this end, we seeded treatment-naive NCI-H3122 cells at clonogenic densities in the presence of ALK-TKIs or vehicle control (dimethylsulfoxide (DMSO)). At 10 days post seeding, the majority of treatment-naive cells plated in DMSO and erALK-TKI cells plated in DMSO or ALK-TKIs formed macroscopic colonies. In contrast, in the presence of ALK-TKIs, treatmentnaive cells formed only microscopic colonies, consistent with tolerance, even upon plating as many as 10,000 cells (Fig. 2a). At the same time, limiting dilution experiments revealed that 1:338-1:660 of treatment-naive NCI-H3122 cells can give rise to robustly growing drug-resistant colonies within 10 weeks of drug exposure ( Fig. 2b-d). Whereas these observations cannot exclude pre-existence of rare (<0.01%) fully resistant subpopulations, together with the observed inhibitor-specific divergence of resistance phenotypes, they indicate that acquisition of resistance de novo via tolerant intermediates must be more common in our experimental system. This inference is consistent with time-lapse microscopy examination of the dynamics of drug resistance emergence, where robust growth in the presence of ALK-TKIs was observed after significant delay (Supplementary Video 1).
The existence of a distinct weakly resistant phenotypic state, which is a precursor to bona fide resistance, termed tolerance or persistence, has been widely studied in microbiology 17,18 . Based on the observation of a similar phenomenon in the context of response to TKIs, the term DTPs has been introduced to describe a weakly resistant subpopulation in EGFR + NSCLC 6 and later in other cancers 9,19,20 . Potentially, DTP cells can reflect a distinct pre-existing subpopulation or arise, either stochastically or deterministically, in response to drug-induced stress. To discriminate between these scenarios, we used tracing with selectively neutral DNA barcodes, an approach that has been previously used to demonstrate the pre-existence of EGFR-TKI resistance in EGFR + NSCLC 21 . Following the transduction of H3122 cells with a high-complexity lentiviral ClonTracer library at a low multiplicity of infection (MOI) (so that most of the transduced cells are labeled with a unique barcode) and elimination of non-transduced cells with puromycin selection, we achieved~100× expansion of the barcoded cells. After taking a baseline aliquot, cells were split into parallel quadruplicate cultures and then exposed to 0.5 µM alectinib, lorlatinib, crizotinib, or DMSO control. After 4 weeks of incubation, barcode frequencies were enumerated by sequencing and were compared with the baseline frequencies (Fig. 2e). Evidence of both negative and positive selection was observed in all treatment groups (including DMSO controls), as barcode diversity, captured with Shannon diversity index, decreased ( Supplementary Fig. 3a), whereas several subpopulations expanded ( Supplementary  Fig. 3b). Spearman's ranking of positively selected barcodes revealed a strikingly high degree of correlation between replicates, indicating pre-existence of stable weakly resistant subpopulations (Fig. 2f). However, correlation between samples treated with different ALK-TKIs was either absent or much less pronounced, indicating that distinct selective pressures exerted by different ALK-TKIs might amplify distinct pre-existing tolerant subpopulations. Unsupervised hierarchical clustering analysis revealed a partial overlap between positively selected subpopulations across multiple ALK-TKIs, indicating that some of the pre-existent phenotypes were fit under multiple ALK-TKIs (Fig. 2g). In contrast to the fitness cost of classical persistence under baseline growth conditions, considered to be a form of bet hedging 18 Fig. 3c). Thus, taken together, this reveals that resistance to an ALK-TKI in H3122 cells originates from pre-existent, heterogeneous subpopulations with ALK-TKI-specific pan-TKI tolerance.
Gradual development of ALK-TKI resistance. Next, we asked how tumor cells progress from tolerance toward full resistance. The prevalent assumption in the modeling, experimental, and clinical communities is that resistance results from a single-hit transition via acquisition of a resistance-conferring mutation or an epigenetic switch 2,22 . Should this be the case, a binary distribution of growth rates, corresponding to tolerant and resistant cells, would be expected in clonogenic assays after ALK-TKI exposure, with longer exposure times leading to a higher proportion of large resistant colonies (Fig. 3a). As expected from the predicted elimination of sensitive subpopulations, clonogenic proportion in 0.5 µM crizotinib progressively increased from the initial 2.6% to 26% at week 3, whereas the clonogenic proportion in 0.5 µM lorlatinib increased from 1% to 17% (Fig. 3b). However, although longer ALK-TKI exposure lead to an increase in the average colony size, this increase was apparently homogeneous, suggesting a gradual development of resistance (Fig. 3c). Analysis of the colony size distributions at the intermediate (2 and 3 weeks) time points using Kolmogorov-Smirnov (KS) statistics revealed that the observed colony sizes cannot be explained by mixed sampling from distributions of tolerant and erALK-TKI cells ( Supplementary Fig. 4a, b). Consistent with the lack of tolerance-associated proliferation penalty in the barcoding experiment inferences, we did not observe substantial differences in the size of colonies formed by cells pre-treated by crizotinib or lorlatinib for varying durations of time in the absence of the drugs (Supplementary Fig. 4c, d). KS analyses consider transitions from tolerance to resistance within the 1-3 weeks of growth under ALK-TKI prior to the clonogenic assay. However, the transition could also occur during colony growth. Therefore, to test compatibility of the experimental data with a single-step transition more rigorously, we developed an agent-based mathematical model of the experimental assay, which simulated growth both prior and during the clonogenic assay across a range of (epi)mutational probabilities and numbers of (epi)mutational steps (Fig. 3d, e and Supplementary Methods). To account for the variability in observed colony sizes, the maximal proliferation probability was set for each simulated colony individually, based on random sampling of sizes of colonies produced by resistant cells. We calibrated the initial proliferation probability using just the median colony size, assuming minor variability in the proliferation rates of such cells. At each cell division, a cell can increase in division rate, reflecting adaptive (epi-)mutations. Cells can transition from the initial to maximal division probability by either a single (epi)mutational step (n = 1) or multiple (n > 1) steps, representing fractional increments of the single transition ( Fig. 3d    can differ based on two parameters: (epi)mutational probability and the number of mutational steps required to reach maximum proliferation rate. We explored outcomes with the full range of possible (epi)mutational probabilities (0-1) and the numbers of mutational steps ranging from 1 to 300. With each choice of parameters, we generated 100,000 in-silico stochastic simulations ( Fig. 3f and Supplementary Video 2). We found that a single mutational step provided the poorest fit to the data for all mutation probabilities. The best fit was achieved with a number of mutational steps in the range of 3-100 (Fig. 3g, h). Considerations of cell death, bi-directionality of (epi)mutational changes in cell fitness and variability in fitness effects of (epi) mutations did not change the poor fit of a single-step transition ( Supplementary Fig. 4e, f and Supplementary Methods). Therefore, our analyses are inconsistent with the prevalent assumption of resistance achieved through a single (epi)mutational change, which converts tolerance to resistance, and instead support a gradual improvement of cell fitness under therapy.
Next, we decided to characterize the mechanistic underpinning of the evolution of resistance. Extensive prior studies of mechanisms of resistance to ALK-TKIs, including studies using the H3122 model enabled us to ask whether the previously identified ALK-TKI resistance mechanisms are consistent with a single-hit acquisition of resistance. Clinical resistance to ALK inhibitors is frequently associated with point mutations in the kinase domain of ALK, which reduce drug binding 23 . However, targeted Sanger sequencing of PCR-amplified cDNA revealed lack of hotspot ALK mutations in erALK-TKI cell lines ( Supplementary Fig. 5). Given the frequent association of clinical ALK-TKI resistance with EML4-ALK amplification 24 and the observed increase in the expression of EML4-ALK in some of the erALK-TKI-resistant cell lines (Fig. 1f), we interrogated EML4-ALK amplification status in the treatment-naive and erALK-TKIresistant cells (lines "0" from Fig. 1f), using the mutational breakapart fluorescence in-situ hybridization assay. The majority of treatment-naive H3122 cells displayed four copies of the wildtype allele and one copy of the fusion allele, with a minor subpopulation where the fusion gene signal could not be detected. Some of the erALK-TKI cells displayed amplification of the mutant allele (Fig. 4a). Extrachromosomal amplification of oncogene-containing DNA has been recently implicated in the rapid evolution of TKI resistance 25 ; however, examination of metaphase spreads revealed that the amplified alleles were localized within the same chromosome. Notably, we observed substantial heterogeneity in the amplification status of EML4-ALK, both between and within erALK-TKI cell lines (Fig. 4b).
The majority of erCriz cells and a fraction of erAlec and erCer contained amplified EML4-ALK. In contrast, erLor cells not only lacked EML4-ALK amplification but also contained a significantly higher proportion of cells with undetectable mutant allele (p < 0.0001 in a χ 2 -test) compared with the treatment-naive cells, suggesting that genomic loss of EML4-ALK might be selectively advantageous under the more potent ALK-TKI.
To investigate the functional importance of the observed changes in EML4-ALK copy numbers, we transfected treatmentnaive erCriz and erLor cells with constructs co-expressing Cas9 and one of two different ALK-targeting guide RNAs, and selected for puromycin-resistant colonies. No colonies could be observed for erCriz cells, suggesting a critical dependency on EML4-ALK (Fig. 4c). Naive H3122 cells formed few small colonies, resembling tolerant colonies formed upon exposure to an ALK-TKI (Fig. 2a). Puromycin-resistant naive cells, transfected with guide RNA directed against ALK expressed EML4-ALK protein, displayed normal ALK expression (Fig. 4d), indicating a strong selective disadvantage of losing EML4-ALK expression and selection of variants that uncouple antibiotic resistance from guide RNA expression. In contrast, erLor cells formed multiple large colonies consistent with a lack of growth inhibition (Fig. 4c), despite complete ablation of the protein expression of the EML4-ALK gene (Fig. 4d). This observation is consistent with reduced baseline EML4-ALK expression in erLor cells (Fig. 1f) and suggests that erLor cells completely lose EML4-ALK addiction. Given that EML4-ALK amplification resulting in overexpression is considered to provide a bona fide resistance mechanism to ALK inhibition 26 , we asked whether the observed increase in EML4-ALK expression is sufficient to account for ALK-TKI resistance. To this end, we retrovirally overexpressed EML4-ALK protein in H3122 cells, resulting in levels of total and phosphorylated EML4-ALK, which closely resemble those observed in EML4-ALK-amplified erALK-TKI cells (Fig. 4e). After exposure to crizotinib, these cells retained residual levels of ALK phosphorylation similar to those observed in the erALK-TKI cells (Supplementary Fig. 6a). However, cells with EML4-ALK overexpression displayed only a marginal increase in crizotinib resistance (Fig. 4f), suggesting that although ALK amplification contributes to resistance, it is insufficient to fully account for it.
Given the insufficiency of EML4-ALK overexpression to confer full resistance, we interrogated the functional impact of the most common resistance-associated point mutation, L1196M, which is considered to be a clear case of a single hit resistance mechanism to crizotinib 26 . Surprisingly, at low expression levels, achieved with <1 retroviral MOI, L1196M expression only moderately decreased crizotinib sensitivity (Fig. 4f). Higher overexpression levels of the mutant protein blocked crizotinib from downregulating phosphorylation of EML4-ALK, as well as its main downstream effector ERK ( Fig. 4e and Supplementary Fig. 6a, b), and provided crizotinib resistance at a level that is similar or higher to that observed in erCriz cells (Fig. 4f). Whereas the results with L1196M overexpression are consistent with previously reported sufficiency to confer resistance in NIH-H3122 cells 27,28 , and ALK mutations were reported co-occur with EML4-ALK amplification in some patients 29,30 , these observations suggest that two mutational hits (mutation and overexpression) might be required to achieve full resistance. Thus, common resistance-associated mutational changes, assumed to provide single hit resistance, reduce drug sensitivity but do not provide for a full adaptation to the ALK-TKI induced selective pressures.
Examination of resistant cell lines with Oncomine Focus Assay 31 failed to detect additional common mutations. However, CytoScanHD single-nucleotide polymorphism (SNP) array revealed additional genetic changes, including recurrent chromosomal amplifications in chromosomes 2, 3, 12, and 17. Interestingly, two out of three examined erLor lines displayed Chr12 p12.1-p11.1 amplification containing KRAS, whereas a third one contained Chr1 p13.2-p12 amplification containing NRAS (Supplementary Fig. 7a). Notably, chromosomal amplification of genomic regions containing KRAS and NRAS were associated with elevated transcript levels of these proto-oncogenes ( Supplementary Fig. 7b). Given the growing evidence for the importance of non-genetic mechanisms in therapy resistance, we used RNA sequencing (RNA-seq) analysis to examine changes in gene expression, accompanying erALK-TKI phenotypes. After a 48 h drug holiday (to reduce the direct impact of ALK-TKI on gene expression), erALK-TKI cells displayed multiple gene expression changes, previously implicated in TKI and chemotherapy resistance, including increased expression of multiple receptor tyrosine kinases (EGFR, HER2, FGFR, AXL, EPHA2, etc.), cytokines, extracellular matrix (ECM) and ECM receptors, and other types of molecules, suggesting a complex, multifactorial nature of resistant phenotypes (Fig. 5a). Co-expression analysis of cell lines in the CCLE database revealed that the resistance-associated genes, upregulated in erALK-TKI cells, belonged to distinct gene co-expression clusters, suggesting that the resistant phenotypes cannot be explained by a single coordinated transcriptional switch ( Supplementary Fig. 8a, b), whereas gene set enrichment analysis revealed several shared enriched gene sets (Supplementary Fig. 8c). Interestingly, all of the examined resistant cell lines displayed an epithelial to mesenchymal transition (EMT) signature. Whereas EMT has been described as one ALK-TKI resistance mechanism 32,33 , our results suggest a need for a more nuanced interpretation, given the predictably distinct phenotypic characteristics of erALK-TKI cell lines, evolved under different ALK-TKIs, as these differences would be missed under the umbrella of the EMT/stemness explanation.   EGFR  ERBB3  ERBB2  MET  EPHA2  EPHA4  EPHB4  AXL  FGFR1  FGFR4  FYN  LYN  IL6R  OSMR  LIFR  IL6ST  STAT3  IL13RA1  IL15RA  IL4R  IL6R*  WNT5A  NXN  WNT7A  CTNNB1  IL18R1  IL1R1  IL1RAP  NT5E (CD73)  IL1RL1  NFKB2  KITLG  AREG  EGF  HBEGF  NRG1  NRG3  IL8  IL7  IL11  IL15  PDGFB  PDGFC  SPARC  SNAI2  VIM  CD44  WWTR1  S100A4  FN1  COL1A1  COL1A2  COL4A1  COL4A2  COL4A3  COL6A1  COL6A2  COL6A3  ITGA1  ITGA2  ITGA3  ITGA5  CD44*  MCL1  BCL2L1  PPP2R3A  PPP2R2C  PPP4R1  NRP1  NRP2  S100A9  MUC1  ADRB1  NRAS  KDM5B Fig. 5 ALK-TKI resistance integrates multiple mechanisms. a Normalized mRNA expression for the indicated genes, previously associated with chemotherapy or targeted therapy resistance, in erALK-TKI cell lines across the indicated functional categories. b UMAP analysis of single-cell mRNA expression data from cells exposed to 0.5 μM alectinib for the indicated time duration. c Violin plot of expression levels of ALK and HER2 following indicated duration of exposure to the indicated ALK-TKIs. *p < 0.05, **p < 0.01, ***p < 0.001, and ****p < 0.0001 of a Mann-Whitney U-test. d ALK-TKI sensitivity of engineered cell lines, lentivirally overexpressing indicated genes to the indicated ALK-TKIs (0.5 μM), as determined by Cell Titer Glo assay. Mean ± SD of experimental replicates (n = 6, representing separate wells) are shown. e Impact of combination of individual resistance mechanisms toward sensitivity to indicated ALK-TKIs (0.5 μM), determined by Cell Titer Glo assay. *p < 0.05, **p < 0.01, ***p < 0.001, and ****p < 0.0001 of ANOVA analysis with Dunnett's (d) or Tukey's (e) multiple comparison correction. Mean ± SD of experimental replicates (n = 6, representing separate wells) are shown. To gain further insights into the dynamics of phenotypic changes during the evolution of resistance, we analyzed single-cell transcriptomes after different exposure times to ALK-TKIs. UMAP analysis revealed gradual phenotypic progression from naive to resistant phenotypes ( Fig. 5b and Supplementary Fig. 9). Even a brief (4 h) alectinib exposure substantially impacted cell phenotypes, suggesting that acquisition of resistance might reflect not only the action of drug-imposed selection but also direct drug-induced cell adaptation. To gain further insight into the temporal dynamics of the acquisition of resistance-conferring expression changes, we analyzed the expression of ALK and HER2 at single-cell levels, at different time points, after exposure to an ALK-TKI. Consistent with high levels of ALK expression and genetic amplification of EML4-ALK in erCriz cell lines, the expression of ALK increased upon exposure to crizotinib, but not alectinib. In contrast, expression of HER2 became elevated within 48 h of exposure to both drugs, suggesting a direct adaptive response, then decreased upon prolonged incubation with crizotinib, while staying upregulated or increasing further under alectinib (Fig. 5c).
To explore possible epigenetic mechanisms underlying the observed stable changes in gene expression in erALK-TKIs cell lines, we analyzed the global repatterning of Histone 3.3 lysine 27 acetylation (H3K27ac), a posttranslational histone modification associated with strong enhancer elements using chromatin immunoprecipitation sequencing (ChIP-Seq) analysis. We found distinct global patterns of H3K27 acetylation at gene regulatory elements between naive and ALK-TKI-resistant lines (Supplementary Fig. 10a). Notably, statistically significant changes in H3K27 acetylation were observed in regulatory elements of genes with resistance-associated gene expression changes (Supplementary Fig. 10b), including genes previously implicated in therapy resistance. For example, consistent with observed differences in protein expression (Fig. 1f), we found new H3K27ac peaks in the vicinity of the ERBB2/HER2 gene in erAlec and erLor lines, but not in the erCriz lines ( Supplementary Fig. 10c). New peaks were also observed near genes with increased genomic copy numbers (EML4, K-RAS, and N-RAS) ( Supplementary Fig. 10d), suggesting that that stable upregulation of resistance-associated genes might result from a combined input of genetic and epigenetic changes.
To evaluate the functional significance of these transcriptional changes, we tested the impact of overexpression of selected genes previously implicated in TKI resistance. Lentiviral overexpression of HER2, FGFR, AXL, and SLUG significantly increased resistance to multiple ALK-TKIs (Fig. 5d). However, resistance levels observed in these engineered cell lines fell short of the levels observed in the erALK-TKI lines, suggesting insufficiency of individual mechanisms to fully account for resistance. On the other hand, combined overexpression of EML4-ALK and HER2 led to significantly higher resistance levels, compared with cells overexpressing either EML4-ALK or HER2 separately, although still failing to recapitulate resistance levels observed in erALK-TKI cells (Fig. 5e). These results support the notion that ALK-TKI resistance reflects a combined output of multiple genetic and epigenetic changes, which contribute to increased fitness in an additive or synergistic manner.
Temporal collateral sensitivities of resistance-evolving cells. It is frequently assumed that, according to the principle of an evolutionary fitness tradeoff 34 , resistance-conferring phenotypes are associated with strong fitness penalties outside of the drug [35][36][37][38] . This fitness penalty could enable evolutionary-informed adaptive therapy, by creating a tug of war between therapy sensitive and therapy-resistant populations with strategic treatment breaks 37 . Thus, we examined growth rates of erALK-TKI cell lines in the presence and absence of the drugs. Whereas in some cases cells taken off the drugs indeed proliferated slower than drug-naive control, some of the resistant cell lines proliferated at similar or even higher rates (Supplementary Fig. 11a). Interestingly, all of the examined erCriz and one out of three of the examined erAlec cell lines displayed higher rates of proliferation in the presence of the inhibitors, consistent with previously reported observations in melanoma 39,40 . The remaining cell lines were either modestly inhibited, or unaffected by the ALK-TKI used for their selection (Supplementary Fig. 11a).
Fitness of tumor cells is context dependent 41,42 and in vitro two-dimensional cultures do not capture death/proliferation dynamics within tumors in vivo. Further, non-cell autonomous interactions between phenotypically distinct subpopulations could significantly alter fitness in competitive settings 43,44 . Therefore, we compared baseline fitness of differentially labeled (green fluorescent protein (GFP) and mCherry) naive and erALK-TKIs cells in a competitive in-vivo context by implanting their mixtures either subcutaneously or into the lungs (via tail vein injections). Only erCriz cells in the lungs displayed signs of reduced fitness compared with therapy-naive cells (Supplementary Fig. 11b). In contrast, erLor cells had a significant selective advantage over naive cells both in subcutaneous and lung tumors. Therefore, resistance is not necessarily associated with a fitness penalty and might be linked to higher fitness outside of the drugs.
On the other hand, growth out of the primary therapy is just one of many potential contexts where an evolutionary tradeoff might be manifested. Application of a different drug, to which ALK-TKI-resistant phenotypes are collaterally sensitive, could provide an alternative strategy to create a fitness tradeoff 45,46 . Given the frequent upregulation of EGFR and HER2 in cells, resistant to the front-line ALK-TKI alectinib, we tested whether the dual EGFR/HER2 inhibitor lapatinib could create a collateral fitness tradeoff. Whereas H3122 cells failed to develop resistance to alectinib in the presence of lapatinib ( Supplementary Fig. 12a), erAlec cells were not sensitized to lapatinib as a single agent and only some of the independently derived erAlec cell lines were inhibited by the combination of alectinib and lapatinib ( Supplementary Fig. 12b). At the same time, evolving cells were remarkably sensitive to lapatinib, even as a single agent, although this sensitivity gradually diminished as cells became more resistant to alectinib (Fig. 6a, b). This collateral sensitivity cannot be explained by the pre-existence of lapatinib-sensitive, alectinibtolerant subpopulations, as pretreatment with lapatinib for up to 3 weeks did not substantially impact sensitivity of H3122 to alectinib (Supplementary Fig. 12c). Consistently, administration of lapatinib as a single agent, after 3 weeks in alectinib, or as a combination therapy, led to complete elimination of H3122 cells in vitro (Fig. 6c).
Encouraged by this observation, we asked whether alectinib/ lapatinib cycling or a combination treatment with the two drugs could outperform alectinib monotherapy in vivo. We found that lapatinib was completely ineffective in adaptive cycling with alectinib, as tumor regrowth during lapatinib cycles was indistinguishable from the regrowth observed with a drug holiday (Fig. 6d). Most likely, this lack of efficiency in vivo reflects its inability to reach the high concentrations required to achieve collateral sensitivity in vitro. Still, lapatinib significantly increased tumor sensitivity to alectinib in a combination therapy setting (Fig. 6d). Whereas, at this point, the clinical utility of our observations remains unclear, they provide a proof of principle that temporally restricted collateral vulnerabilities of evolutionary intermediates might be exploited therapeutically to improve responses, and potentially, to forestall the emergence of resistance in targetable lung cancers.

Discussion
Despite substantial advances in deciphering the molecular mechanisms of resistance to TKI-based targeted therapies and the development of more effective drugs, advanced targetable lung cancers remain incurable, as tumors eventually acquire resistance and relapse. Developing strategies to interfere with evolving resistance has the potential to substantially improve long-term survival outcomes. However, this is contingent on a correct understanding of the underlying evolutionary dynamics. In contrast to the large body of experimentally derived knowledge on individual molecular mechanisms of resistance, our understanding of its evolutionary causes and dynamics remains less developed. The subject of evolutionary-informed therapies has received significant attention from mathematical modelers 5,16,47,48 . However, given the paucity of experimental data acquired to derive solid assumptions for building models, these modeling studies often have to rely on conjectures from reductionist, mechanism-centered studies, thus limiting their potential.
Acquired resistance is commonly viewed through two major conceptual frameworks. According to the first, currently dominant gene-centric framework, resistance arises from a selective expansion of genetically or epigenetically distinct (meta)stable subpopulations. The subpopulation(s) could either pre-exist therapy or arise de novo from mutational conversion of sensitive or tolerant cells. In both cases, clinical and experimental studies operating within this paradigm typically reduce acquired resistance to a single cause, such as a point mutation (e.g., the L1196M gatekeeper mutation) 30 , gene amplification (e.g., EML4-ALK 30 , cMET 49 , etc.), or a non-mutational stable change in the expression of a resistance-conferring gene (e.g., IGFR) 50 . Consequently, the assumption of a single hit resistance mechanism is highly prevalent in the mathematical modeling community 5,16,22 .
The second framework views resistance as the result of druginduced reprogramming, where phenotypic plasticity enables tumor cells to directionally rewire their signaling, metabolic, and gene-expression networks to cope with inhibitor-induced perturbations 9,51,52 . In this system-biology-based framework, resistance can be gradual and multifactorial; however, the causal role of selection is sometimes rejected due to its link with mutation-centric paradigms 9 .
The two frameworks are not strictly mutually exclusive, as they can be bridged in a two-step process involving reprogrammingmediated formation of tolerant subpopulations followed by an  Fig. 6 Collateral sensitivities of evolutionary intermediates. a Impact of pre-exposure to 0.5 µM alectinib for the indicated periods on sensitivity to alectinib and lapatinib, measured by Cell Titer Glo assay. One-way ANOVA was used for both drugs (p < 0.0001). Adjacent time points were compared using Sidak's multiple comparison tests (***p < 0.001 and ****p < 0.0001). Mean ± SD of experimental replicates are shown. n = 3, representing separate wells, for naive control and erAlec H3122 cells. n = 6 wells (3 wells each of 2 biological replicates) for cells pre-exposed to alectinib. Data for one of the biological replicates is missing in week 6. b Illustration of association of evolving resistance to alectinib with collateral sensitivity to lapatinib. c Lapatinib can prevent the development of resistance to alectinib in vitro, both as a combination treatment, or in drug cycling. Green shading indicates switching from alectinib to lapatinib monotherapy. Residual signal in the combination therapy and cycling groups reflects autofluorescence, as visual examination revealed lack of surviving tumor cells. Mean ± SD of experimental replicates (n = 3 representing separate wells for DMSO and lapatinib monotherapy, n = 6 for the remaining groups) are shown. d Change in volume of H3122 xenograft tumors treated with indicated therapies; treatment was initiated 3 weeks post tumor implantation. Green shading indicates switching from alectinib to lapatinib monotherapy or vehicle control. Mean values ± SE are shown; n = 6 for lapatinib monotherapy, n = 10 for alectinib monotherapy, switching between alectinib and lapatinib, and alectinib/lapatinib combination, n = 9 for vehicle control, replicates represent separate tumors. e Evolving resistance interpreted through a fitness landscape metaphor. Naive cells occupy a local fitness peak. Drug exposure reshapes the landscape, turning this fitness peak into a fitness trough. Different ALK inhibitors act on partially distinct outliers, directing their evolution toward distinct fitness peaks.
NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-020-16212-w ARTICLE (epi)mutational switch to full resistance. Our results, however, suggest an alternative scenario of gradual, multifactorial resistance, which integrates features of the two paradigms described above. We speculate that selective pressures imposed by therapies act on phenotypic heterogeneity, stemming from both stochastic and drug-induced changes, leading to a gradual increase in population fitness through acquisition of additional genetic and epigenetic changes until a local fitness peak is reached (Fig. 6e). Different evolutionary trajectories under distinct ALK-TKIs reflect inhibitor-specific differences in both selective pressures and direct drug-induced adaptive responses, which can become hardwired if selectively advantageous. This model is closer to Darwin's original thesis "As natural selection acts solely by accumulating slight, successive, favorable variations, it can produce no great or sudden modifications" 53 rather than the currently prevalent mutation-centric re-interpretation of Darwinian selection within the clinical and basic science cancer research community 54 . Gradual, multifactorial acquisition of resistance has been recently observed in an elegant experimental study of acquired BRAFi resistance in melanoma 9 . The authors rejected the Darwinian explanation based on the lack of pre-existent resistance and thus interpreted their observations strictly within a reprogramming paradigm. We argue that their data are highly consistent with our model, which does not reduce Darwinian selection to genetically distinct pre-existing variants. Likewise, a recent single-cell sequencing study in triple-negative breast cancer demonstrated that although resistance to chemotherapy can be traced to pre-existent genetically distinct subpopulations, resistant cells acquire new phenotypic changes, indicating epigenetic modlifications 55 . Similarly, a recent study in hormone positive breast cancers demonstrated a multi-step underpinning of resistance, involving both genetic and non-genetic changes 51 . Therefore, our inferences with a model of ALK mutant NSCLC might be generalizable to a broader range of scenarios of acquired therapy resistance.
Finally, our data suggests that evolutionary intermediates of acquired resistance may present temporally restricted opportunities for therapeutic interventions, which is consistent with the previously reported existence of collateral sensitivities in mutational intermediates in the evolution of resistance in Ph+ ALL 56 . These findings support the notion that explicit interrogation and nuanced understanding of the evolutionary mechanisms of acquired drug resistance could open the way to designing evolutionary-informed therapy approaches, focused on staying a step ahead of resistance, rather than just reacting to it when it (inevitably) arises.
Short-term viability assays. Short-term cell viability was measured using Cell Titer Glo reagent (Promega). Typically, 2000 cells per well were plated in a 96-well plate (Costar). Drugs were added 24 h later and the assays were performed using manufacturer-recommended protocol 3-4 days after drug addition. Viability was calculated first by subtracting the luminescence of empty wells then by division by average DMSO luminescence. Dose-response curves were generated by fitting the following equation to the data: 50 is the half maximal inhibitory concentration, k is the hill slope, and b is the luminescence as the drug concentration approaches ∞.
Clonogenic assays. Cells were plated in the presence of ALK-TKIs or DMSO vehicle control at varying densities into 6 cm dishes or multi-well plates in duplicates or triplicates, and were grown for 10 days, at which point they were fixed and stained with crystal violet, following protocols described in ref. 57 . To measure the evolution of gradually increasing resistance, nuclear mCherry expressing H3122 cells were plated at~400,000 cells per 6 cm dish, allowed to attach overnight, and exposed to DMSO vehicle control, 0.5 μM crizotinib, or 0.5 μM lorlatinib the following day. After culturing for 1-3 weeks, cells were collected and seeded in 96well plates (Costar) at 50 cells/well for DMSO control, or between 50 and 500 cells for crizotinib and lorlatinib. Number of colonies and colony sizes were measured 1 week later for colonies larger than 1000 pixels (~5 cells), based on fluorescent area. To minimize the impact of variability in seeding numbers, clonogenic survival in the presence of ALK inhibitors was normalized to clonogenic data in the DMSO controls.
Determining frequency of resistance-initiating cells. Cells were plated with initial seeding densities of 400, 400, 800, 1400, and 400 cells/well in DMSO (0.1%), crizotinib (0.5 µM), ceritinib (0.1 µM), alectinib (0.5 µM), and lorlatinib (0.5 µM) treated plates, respectively, in a 96-well plate (Falcon). Five additional 2× dilutions were generated from these wells, each with 10 separate wells. ALK inhibitors were added after 24 h. After 7 weeks of treatment, wells containing no colonies larger than~50 cells were counted. The natural log of the proportion of wells without colonies was fitted linearly against the initial cell number in each well. The number of cells per resistance-initiating cell was calculated as 1 1Àe slope and the error of this value as Error slope e slope 1Àe slope ð Þ

.
Clone-tracing assay. H3122 cells were transduced at 10% efficiency (as defined by fluorescence-activated cell sorting analysis of dsRed expression) with ClonTracer neutral DNA barcode library 21 , kindly provided by Frank Stegmeier (Addgene #67267). Barcode containing cells were selected with puromycin and expanded for 23 days. Cells were expanded from 10 5 cells (after puromycin selection) to 3.4 × 10 7 cells. Quadruplicate cultures of 1.5 × 10 6 cells each were plated into 10 cm dishes in the presence of 0.1% DMSO, 0.5 µM crizotinib, 0.5 µM alectinib, and 0.5 µM lorlatinib. Two aliquots of 1.5 × 10 6 cells were frozen to serve as an initial mixture. Cells were collected following 4 weeks of cell culture. Genomic DNA was extracted using proteinase K digest, followed by phenol-chloroform purification. Barcodes were amplified, sequenced, and analyzed following protocols described in ref. 21 and an updated procedure provided on the Addgene website https://www.addgene.org/ pooled-library/clontracer. A first round of amplifications was performed for 35 cycles using the primers 5′-TCG ATT AGT GAA CGG ATC TCG ACG-3′ and 5′-AAG TGG ATC TCT GCT GTC CCT G-3′ for 35 cycles. A second round was performed for 15 cycles using second-generation clonTracer primers with the following sequence: 5′-CAA GCA GAA GAC GGC ATA CGA GAT-Variable Sequence-GTG ACT GGA GTT CAG ACG TGT GCT CTT CCG ATC TCT AGC ACT AGC ATA GAG TGC GTA GCT-3′. Barcode enrichment analyses focused on barcodes that were enriched above the most frequent barcode within the average of the two baseline samples (0.0077). Heatmap.2 in R (using default parameters) was used for heatmap clustering. Corrplot 58 in R (using spearman correlation coefficients) was used to visualize correlations.
Fluorescent in-situ hybridization analyses. Cells were cultured in 10 cm dishes in 10 ml of 1640 medium supplemented with 10% FBS. After 24 h, 100 µl of colcemid (10 µg/ml, Life Technologies, Carlsbad, CA) was added to each culture and culture replaced to incubator for another 1 h before collecting. Cell suspension was transferred to 15 ml cortical tubes. Culture medium was removed by centrifuge. Cell pellets were subjected to hypotonic treatment with 0.075 M KCl and then fixed with Methanol and acidic acid at 3:1 v/v. Cell suspensions were dropped to slides and treated with 0.005% (wt/vol) pepsin solution for 10 min, followed by dehydration with 70%, 85%, and 100% (vol/vol) ethanol for 2 min each. Hybridization was performed by adding 10 µl of ALK dual-color break-apart probe on each slide (Cytocell, Cambridge UK), a coverslip was placed, and the slides were sealed with rubber cement. The specimens were subjected to denaturation at 75°C for 3 min and hybridized at 37°C for 16 h. The slides were washed in 0.4× saline-sodium citrate at pH 7.2 and then counterstained with 4′,6-diamidino-2-phenylindole (DAPI). Results were analyzed on a Leica DM 5500B fluorescent microscope. Cell images were captured in both interphase and metaphase cells.
RNA-seq analysis. Following 48 h of culturing in the absence of inhibitors (to minimize the impact of directly induced gene expression changes), RNA was isolated for erALK-TKI and parental H3122 cells using an RNAEasy Minikit (Qiagen). Reads were generated using a MiSeq instrument. Alignment was achieved using HiSat2 and the human hg19 reference genome. Normalized reads were obtained from DeSeq2. Analysis was performed with gene with more than 25 reads in at least one sample and more than two-fold change from ALK-TKI-naive cells in at least one resistant line.
Nanostring assay. Following 48 h of culturing in the absence of inhibitors (to minimize the impact of directly induced gene expression changes), RNA was isolated for erALK-TKI and parental H3122 cells using an RNAEasy Minikit (Qiagen ChIP-Seq analysis. H3122 cells were fixed with 1% formaldehyde for 15 min and quenched with 0.125 M glycine. Chromatin was isolated by the addition of lysis buffer, followed by disruption with a Dounce homogenizer. Lysates were sonicated and the DNA sheared to an average length of 300-500 bp. Genomic DNA (Input) was prepared by treating aliquots of chromatin with RNase, proteinase K, and heat for de-crosslinking, followed by ethanol precipitation. Pellets were resuspended and the resulting DNA was quantified on a NanoDrop spectrophotometer. Extrapolation to the original chromatin volume allowed quantification of the total chromatin yield. An aliquot of chromatin (30 µg) was precleared with protein A agarose beads (Invitrogen). Genomic DNA regions of interest were isolated using 4 µg of antibody against H3K27Ac (Active Motif, catalog number 39133, Lot number 01518010). Complexes were washed, eluted from the beads with SDS buffer, and subjected to RNase and proteinase K treatment. Crosslinks were reversed by incubation overnight at 65°C and ChIP DNA was purified by phenol-chloroform extraction and ethanol precipitation. Quantitative PCR (qPCR) reactions were carried out in triplicate on specific genomic regions using SYBR Green Supermix (BioRad). The resulting signals were normalized for primer efficiency by carrying out qPCR for each primer pair using Input DNA. Illumina sequencing libraries were prepared from the ChIP and Input DNAs by the standard consecutive enzymatic steps of end polishing, dA addition, and adaptor ligation. After a final PCR amplification step, the resulting DNA libraries were quantified and sequenced on Illumina's NextSeq 500 (75 nt reads, single end). Reads were aligned to the human genome (hg38) using the BWA algorithm (default settings). Duplicate reads were removed and only uniquely mapped reads (mapping quality ≥ 25) were used for further analysis. Alignments were extended in silico at their 3′-ends to a length of 200 bp, which is the average genomic fragment length in the size-selected library and assigned to 32-nt bins along the genome.
To define relevance toward transcriptional changes, RNA-seq reads were reanalyzed using HISAT2 and hg38 for proper comparison with ChIP-seq data. RseQC was used to check stranded information. HTSeq-count was used for reads counting. Raw ChIP-seq files were marked duplicate using picard and MACS2 was used to call peaks (q < 0.05). Peaks from four samples were merged using bedtools. Black list was removed. Featurecounts was used for read counting. All genes within ±20 kb of a peak are used for comparisons. Peaks with <15 reads in all samples with ChIP-seq and RNA-seq data (naive, erCriz0, erAlec0, and erLor0) were excluded. Genes with <25 reads in all samples with ChIP-seq and RNA-seq data were excluded. Gene and peak levels were normalized using the DESeq2 rlog function. Genes with a difference in rlog values between erTKI and naive cells of ≥1 are considered upregulated. Genes with a difference in rlog values between erTKI and naive cells of ≤ −1 or less are considered downregulated. All other genes are considered neutral. Peak values that are associated with multiple genes are used multiple times.
Tracks: ACS2 was used to call peaks. Partek was used to visualize chromosomal tracks.
Xenografts studies. erALK-TKI or parental H3122 were suspended in 1 : 1 RPMI/ Matrigel (ThermoFisher) mix and subcutaneously implanted into 4-6-week-old NSG mice, with two contralateral injections per animal, containing 10 6 tumor cells each. After 3 weeks, the animals were treated with 25 mg/kg alectinib (purchased from Astatech), 6 mg/kg lorlatinib (obtained from Pfizer), or vehicle control via daily oral gavage. Tumor diameters were measured weekly using electronic calipers and tumor volumes were calculated assuming spherical shaped tumors. Tumors in the collateral sensitivity experiments were treated with 11 mg/kg alectinib for 1 week before switching to 50 mg/kg alectinib. Xenograft studies were performed in accordance with the guidelines of the Institutional Animal Care and Use Committee of the H. Lee Moffitt Cancer Center. Animals were maintained under AAALAC-accredited specific pathogen-free housing vivarium and care and veterinary supervision following standard guidelines for temperature and humidity, with 12/12 light cycle.
Statistics and reproducibility. Figures and statistical analyses of experimental data were performed using GraphPad Prism 8.3 software, R version 3.6.1, and Python arrays 64 . Statistical tests are stated in the figure legends. P-values < 0.05 were considered as statistically significant. Exact p-values are provided in Supplementary Table 2. Continuous variables were expressed as the mean ± SEM or mean ± SD as indicated in the figure legends. Unless otherwise stated, the experiments were performed at least three times with similar results. Single-cell transcription profiling, Nanostring, bulk sample RNA-seq, CRISPR knockout, and clone-tracing experiment were performed for a single time with biological replicates indicated in the figure legend and Methods section. Immunoblot analyses were performed two times with similar results. The original unprocessed and uncropped gels/blots with molecular weight marker information are shown in Supplementary Figs. 13-15 and provided as a Source Data file.
Reporting summary. Further information on research design is available in the Nature Research Reporting Summary linked to this article.

Data availability
ChIP-seq data that support the findings of this study have been deposited in GEO with the accession code GSE144282. All other datasets during and/or analyzed during the current study are available from the corresponding author on reasonable request. ORFeome and MsigDB are publicly available.

Code availability
Code is available on GitHub at nryoon12/Gradual_Development_of_ALK-TKI_Resistance.