Genotype-fitness maps of EGFR-mutant lung adenocarcinoma chart the evolutionary landscape of resistance for combination therapy optimization

SUMMARY

Cancer evolution poses a central obstacle to cure, as resistant clones expand under therapeutic selection pressures. Genome sequencing in relapsed disease can nominate genomic alterations conferring resistance, but sample collection lags behind, limiting therapeutic innovation. Genome-wide screens offer a complementary approach to chart the compendium of escape genotypes, anticipating clinical resistance. We report genome-wide ORF resistance screens for first and third generation EGFR inhibitors, and a MEK inhibitor. Using serial sampling, dose gradients, and mathematical modeling, we generate genotype-fitness maps across therapeutic contexts, and identify alterations that escape therapy. Our data expose varying dose-fitness relationship across genotypes, ranging from complete dose-invariance to paradoxical dose-dependency where fitness increases in higher doses. We predict fitness with combination therapy, and compare these estimates to genome-wide fitness maps of drug combinations, identifying genotypes where combination therapy results in unexpected inferior effectiveness. These data are applied to nominate combination optimization strategies to forestall resistant disease.

Graphical Abstract

eTOC Blurb

Bolan et al. generate genome-wide genotype-fitness maps for multiple targeted cancer therapies and use them to comprehensively identify drug-resistant genotypes, expose the effect of drug dosage on genotype resistance, show that unexpected buffering interactions make combination resistance unpredictable based on single-agent resistance, and algorithmically nominate treatment strategies to forestall resistant disease given these unexpected interactions.

INTRODUCTION

Targeted therapies have transformed cancer care with dramatic clinical responses (Rotow and Bivona, 2017). Nonetheless, disease evolution to resistance is the rule, limiting the effectiveness of these therapeutics. In response to the strong selection pressure exerted by targeted therapies, resistant sub-clones emerge resulting in recurrence (Alderton, 2014). Epidermal growth factor receptor (EGFR)-mutant lung adenocarcinoma (LUAD) provides an illustrative example. Among patients that initially respond to first generation EGFR tyrosine kinase inhibitors (TKIs), median progression free survival is limited to 10.2 months (Soria et al., 2018) as these patients acquire resistance through emergence of resistant sub-clones from vast intra-tumoral heterogeneity (Blakely et al., 2017; Alderton, 2014; Fisher, Pusztai and Swanton, 2013; Gerlinger et al., 2012).

While EGFR-mutant LUAD displays “oncogenic addiction” and is prone to undergo apoptosis when EGFR signaling is blocked, genomic alterations in the downstream signaling circuitry of EGFR can confer resistance through bypass signaling (Sharifnia et al., 2014). Consistently, many of the genomic alterations found in resistant clinical samples (EGFR “gatekeeper” T790M mutations, ERBB2 amplification, MET amplification, AXL upregulation, BRAF mutations, MAPK1 amplification, and PIK3CA mutations (Morgillo et al., 2016; Stewart et al., 2015)) likely confer resistance through this mechanism. Other genetic alterations, such as CTNNB1 mutations, CCNE1 amplification, and MYC amplifications, may confer resistance through other mechanisms, such as epithelial-to-mesenchymal transition (EMT), transformation to small cell lung cancer, and the cancer-stem-cell-like phenotype (Blakely et al., 2017; Rotow and Bivona, 2017). Resistance-conferring alterations remain unknown in 18-30% of patients (Sequist et al., 2011; Yu et al., 2013).

To overcome resistance, a third-generation EGFR TKI, osimertinib, has been developed to inhibit the most common EGFR TKI resistant alteration, EGFR^T790M (Soria et al., 2018). Nevertheless, intra-tumor clonal diversity prevails and LUAD evolves and recurs despite this therapy, with an emerging resistance profile including EGFR^C797S mutations, small cell lung cancer transformation, ERBB2 and MET amplifications, KRAS and BRAF mutations, and as of yet unknown mechanisms in many patients (Jia et al., 2016; Ho et al., 2017; Ortiz-Cuaran et al., 2016).

Combining drugs with low cross-resistance has been proposed as a strategy to overcome the central obstacle of tumor evolution (Bozic et al., 2013). However, the rapid progress in cancer drug discovery has resulted in a combinatorial challenge where the number of possible combinations, including varying doses and schedules, outpaces the ability to test these combinations through the traditional clinical trial paradigm. Further, the comprehensive charting of the resistance profiles of emerging agents via sequencing of relapsed samples lags behind the pace of therapeutic innovation (Morgillo et al., 2016; Sequist et al., 2011; Stewart et al., 2015; Sullivan and Planchard, 2016). Genetic perturbation screens can accelerate this process by charting the landscape of genotypes leading to therapeutic escape (Johannessen et al., 2010; Johannessen et al., 2013; Le et al., 2016; Wilson et al., 2015). However, genome-wide screens typically define resistant genotypes qualitatively as genotypes enriched through drug selection, relative to control. This relative measure is influenced by the spectrum of phenotypic strengths across the entire compendium of genotype-drug interactions. This precludes effective comparison of resistant genotypes across screen conditions and therapeutics, limiting the ability to discover drug combinations with minimal overlapping resistance.

We posited that an absolute measure of genotype growth – fitness (Shen et al., 2017) – will enable integration across treatment conditions. We therefore incorporated serial sampling into genome-wide open reading frame (ORF) over-expression screens, and resolved genotype-specific fitness (growth rate) to study the impact of EGFR inhibitors (first and third generation), a MEK inhibitor, and their combinations on tumor evolution. We defined the compendium of resistant genotypes for each of the therapeutic inhibitors, which notably include multiple transcription factors. We further show distinct dose dependencies across resistant genotypes. Finally, through combination therapy fitness mapping, we uncovered unexpected drug-drug interactions where genotypes show higher fitness in combination compared with monotherapy, laying the foundation for algorithmic combination therapy optimization.

RESULTS

Genotype-fitness mapping reveals candidate erlotinib-resistant ORFs, including transcription factors

To broadly define the compendium of genomic alterations capable of escaping EGFR inhibition, we generated genotype-fitness maps (fitness defined as ORF growth rate) using a genome-wide ORF over-expression screen (including 17,225 ORFs, representing 12,728 genes (wild-type and mutated)) in the EGFR-mutant LUAD PC-9 cellular model (Figure 1A, Data S1, see STAR Methods) (Ono et al., 2004). To calculate the fitness of each ORF, we cultured pooled transduced cells in 25 nM erlotinib to four different time points 0, 8, 15, and 22 days, in three replicates. We integrated cell number and ORF abundance at each time point, derived the corresponding cell number, and fit growth curves for each genotype to measure its fitness (Figure 1B). We found that exponential growth is an appropriate approximation in this experimental setting (Figure S1A).

Figure 1. Genome-wide open reading frame (ORF) over-expression screen enables genotype-fitness mapping of erlotinib-resistant alterations.

A. PC9 EGFR mutant cells were transduced with 17,255 ORFs, covering 12,728 genes and 35 mutant oncogenes. Cells for each of the four serial time-points were individually cultured with 25nM erlotinib or DMSO, in triplicates. B. Genotype (ORF)-specific fitness is calculated by (i) deriving the number of cells per ORF at each time point, and (ii) fitting exponential growth curves across the four time points. Examples of EGFR^{L858R, T790M}, ERBB2, and PIK3CA^E545G are shown. s – fitness in Δ(log2(cell number))/days. C. Fitness is shown across the replicates. ORF with fitness > 0.1 and fitness robust Z-score > 3 in all three replicates are shown (black), including genotypes previously associated with clinical relapse samples (red), or pre-clinical studies (blue). D. Resistance ORFs are annotated by affected gene pathways and processes, along with their fitness, and prior identification in clinical relapse samples, or pre-clinical studies.

See also Figures S1, Table S1, and Data S1.

We identified 37 ORFs that conferred resistance to erlotinib (Figures 1C, 1D, STAR Methods and Table S1; resistance defined as fitness > 0.1 Δ(log₂(cell number))/days and fitness robust Z-score > 3 in all replicates; P < 1×10⁻¹⁰ by SuperExactTest of the resistant ORF intersection between replicates (Wang, Zhao and Zhang, 2015)). The resistant ORFs included alterations previously discovered in resistant tumors, including wild-type and mutant EGFR (EGFR^T790M and EGFR^{T790M, L858R}), ERBB2, and PI3KCA^E545G (Jia et al., 2016; Morgillo et al., 2016) (Table S2). We also identified ORFs previously shown as resistant to erlotinib in a kinome-wide screen, including NTRK receptor tyrosine kinases (NTRK1 and NTRK2) and the adapter protein, CRKL (Sharifnia et al., 2014), as well as additional known preclinical resistance alterations (KRAS^G13D and ABCG2) (Berger et al., 2016; Elmeliegy et al., 2010; Neul et al., 2016). Fitness-based resistance correlated well with the traditional ORF fold-change enrichment metric (Figure S1B-D), but resulted in higher inter-replicate correlation (Figure S1E and S1F) and identified two additional clinically validated ORFs (EGFR^T790M,L858R and EGFR^WT), which failed to reach significance based on relative enrichment alone (Figure S1D).

We identified 27 additional putative erlotinib resistant ORFs, including several ORFs likely capable of reconstituting the downstream EGFR signaling, such as tyrosine kinases (e.g., PDGFRB, NTRK3, and FGFR2^{A315del,Q778Pfs*11}), an adapter protein (GRB2), a GTPase (HRAS), and a MAP kinase (MAP2K8). Several other categories of genes were identified (Figure 1D) including G-coupled protein receptors (GPCRs), which have been shown to drive resistance to crizotinib in ALK-driven LUAD through activation of protein kinase C (Wilson et al., 2015). In addition to reactivation of mitogenic signaling, EMT and acquisition of stem-cell-like properties have been described as resistance mechanisms to erlotinib (Shien et al., 2013). Consistent with this path to resistance, our candidates include two transcription factors (FOXA1 and SOX15, Figure 1D) previously linked to these phenotypes (Iwafushi-Doi et al., 2016; Wang et al., 2013; Chang et al., 2015; Blakely et al., 2017, Liau et al., 2017; Mu et al., 2017).

Genotype-fitness-dosage mapping reveals differential dose dependency behaviors across resistant ORFs

Spatial, microenvironmental, and pharmacodynamic heterogeneity results in large variability in inter and intra-patient drug exposure (Fu, Nowak and Bonhoeffer, 2015; Lankheet, 2015). To explore potential dose-fitness dependencies across genotypes, we generated genotype-fitness maps across a dose gradient (DMSO control, 5, 15, 25 and 40 nM erlotinib) (Figures 2A-2C, S2A, and STAR Methods). We uncovered a total of 68 erlotinib-resistant ORFs across the dose spectrum, representing five clusters of dynamic dose-dependent behavior (Figures 2D, S2B, S2C, and STAR Methods).

Figure 2. Fitness-dose landscape reveals distinct genotype-specific dependencies of erlotinib resistance, including dose-invariance, dose-dependent sensitivity, and paradoxical higher fitness advantage at dose-increase.

A. Genotype-fitness maps are derived across the erlotinib dose range (DMSO control, 5nM, 15nM, 25nM, and 40nM). B. Genotype-fitness landscape for all ORFs across increasing erlotinib doses. ORFs are sorted in ascending order by composite fitness across all doses. C. Genotype-fitness-dose landscape as in (B) shown by robust Z-score of fitness. Clinically associated and representative ORFs are labeled. D Fitness across erlotinib doses (left) of 68 resistant ORFs detected across the genotype-fitness-dosage landscape (robust Z-score > 3, in all three replicates at any erlotinib dose). Resistant ORFs were grouped by k-means into clusters (C1-5), with control ORFs added in gray for comparison (left). Resistant ORFs are annotated according to their involvement in cellular pathways or processes (center). Fitness mean and 95% confidence interval for each cluster is displayed across doses with comparison to the ORF library population mean and 95% confidence interval (right). E. Examples of cluster 1 (C1: ERBB2, KRAS^G13D), cluster 2 (C2: NTRK1), cluster 3 (C3: BRAF^V600E), cluster 4 (C4: PIK3CA^E545G), and cluster 5 (C5: GPR52) are shown with mean fitness and standard error compared to the ORF library population mean. Standard error is calculated over 6 or 3 replicates for each ORF and dosage.

See also Figure S2, Tables S2, S3, and Data S1.

The first cluster (C1) represents the ORFs with the most robust resistance, showing no decrease in fitness as erlotinib dosage is increased (Figures 2D, 2E, and Table S3). This group includes the two most frequent alterations identified in clinical samples, EGFR^T790M and ERBB2 (Stewart et al., 2015). C1 also includes mutated KRAS, a frequent LUAD driver and a reported resistance candidate to EGFR-directed therapy in vitro (Berger et al., 2016). The dose invariance of these genotypes suggests that they are capable of reinstating the full signaling effect of mutated EGFR, resulting in complete bypass of erlotinib, validated through restoration of pERK and pAKT levels in the presence of erlotinib (Figures S2D and S2E).

While mutant KRAS was shown to overcome EGFR inhibition, KRAS mutations have not been commonly observed in erlotinib relapse samples. Our data suggest this discrepancy may arise from a fitness disadvantage relative to the median population fitness in the absence of EGFR inhibition (Figure 2E and Figure S2F), consistent with mutual exclusivity between EGFR and KRAS mutations observed in untreated LUAD (Unni et al., 2015). This phenomenon may result from overstimulation of the MAP kinase pathway by concomitant EGFR and KRAS mutations, which leads to impaired proliferation (Unni et al., 2015). This overstimulation is abrogated when EGFR is inhibited with erlotinib, allowing enhanced growth. The fitness disadvantage without therapy may translate into lower clonal abundance of KRAS mutants in pretreatment EGFR-mutated LUAD, favoring the emergence of alternative resistant alterations with no impact on growth prior to therapy, such as EGFR^T790M. Consistent with these data, the application of next generation EGFR inhibitors, designed to overcome T790M mutations, have exposed RAS mutations in increasing frequency in disease relapse (Ortiz-Cuaran et al., 2016).

The second cluster (C2) conferred moderate resistance, though showing erlotinib dose dependency, with declining fitness as erlotinib concentration is increased (Figures 2D, 2E, and Table S3). This cluster includes an infrequent erlotinib-resistant alteration – EGFR^WT amplification – as well as known drivers of LUAD, such as NTRK1 amplification, and FGFR2 mutations (Helsten et al., 2015; Ji et al., 2013; Rosell and Karachaliou, 2016). This group includes several other tyrosine kinases (e.g., PDGFRA-B, NTRK2-3, PDGFRB^S180F, KIT^L576P, and FGFR2^{A315del, Q778Pfs*11}), a GTPase (HRAS), an adapter protein (CRKL), and an efflux transporter (ABCG2). The dose dependency of C2 ORFs may be due to an inability to completely reinstate EGFR signaling across the entire dose range, due to insufficient signaling overlap with EGFR or saturation of the underlying mechanism, as in the case of the efflux transporter ABCG2.

The third cluster (C3) showed atypical behavior, with no fitness difference up to 25 nM erlotinib, and a marked fitness advantage exposed at the highest erlotinib dose (Figures 2D, 2E, and Table S3). C3 genotypes include two clinically described alterations, BRAF^V600E, and AXL, which is upregulated in 20% of resistant tumors (Zhang et al., 2012). As the C3 ORFs BRAF^V600E, RAF1, and MAP3K8 specifically activate the MAP kinase pathway (Figure S2G), we speculate that other ORFs represented in C3 (AXL, RET^M918T, SRC^Y530F, MST1R, and RASGRPs) may function through similar MAP kinase pathway activation, which may not provide a fitness advantage until EGFR signaling is sufficiently subdued. The paradoxical dependency of these genotypes on a high degree of EGFR inhibition is reminiscent of “addiction” of melanoma cells to sustained BRAF or MEK inhibition after resistance has emerged (Thakur et al., 2013; Sale et al., 2019).

The fourth cluster (C4) showed moderate fitness advantage relative to the population median across all dosages (Figures 2D, 2E, and Table S3). Most notably, this cluster included three different activating PIK3CA mutations (PIK3CA^H1047R, PIK3CA^E545G, and PIK3CA^C420R). In addition, C4 included adapter proteins (GRB2, CRK), transcription factors (SOX15, NR2E1, MSX2), and GPCRs (GPR161, HTR4, P2RY11). Given the enrichment in PIK3CA mutations, C4 ORF over-expression may more prominently engage signaling through the PI3 kinase/AKT pathway, which is one of the major downstream outputs of mutant EGFR. Unlike MAP kinase signaling activators, PI3KCA mutants show a fitness advantage compared to the population fitness in the absence of erlotinib. This is consistent with PIK3CA and EGFR mutation co-occurrence in pre-treatment LUAD sequencing data (Yu et al., 2018) and may explain how these resistant clones can emerge despite the more modest fitness advantage they confer to drug treated cells.

Lastly, cluster 5 (C5) showed a moderate fitness advantage at intermediate doses, which was then suppressed at higher erlotinib doses (Figures 2D, 2E, and Table S3). This cluster is highly enriched with GPCRs and thus may represent activation of the PKC or PKA pathways, as has been suggested for a subset of these genes in NSCLC and melanoma (Johannessen et al., 2013; Wilson et al., 2015). In addition, this cluster was enriched with transcription factors (YEATS4, FOXA1, FOXA3), suggesting that cellular-reprogramming-mediated resistance may have different evolutionary dynamics compared with more traditional bypass mechanisms. This is reminiscent of glioblastoma, where cellular plasticity was associated with cell persistence rather than complete drug resistance (Liau et al., 2017).

Genotype-fitness mapping of osimertinib reveals differential resistant alterations compared to erlotinib

Osimertinib is increasingly used in the treatment of EGFR-mutant LUAD as it is effective against EGFR^T790M and prolongs progression free survival (Soria et al., 2018). Given the limited understanding of its resistance profile (Jia et al., 2016), we defined the genotype-fitness landscape of osimertinib. We identified 44 resistant ORFs (fitness > 0.1 and robust Z-score > 3 in all replicates, P = 4 × 10⁻⁶ by hypergeometric test of resistant ORF intersection between replicates), including KRAS^G13D, BRAF^V600E, and RET, which have been reported in clinical resistance cases (Ho et al., 2017; Ortiz-Cuaran et al., 2016; Piotrowska et al., 2018) (Figures 3A, 3B, Data S1, and Table S1). Chief among the 41 additional resistant ORFs we identified were those capable of activating the canonical signaling pathways of EGFR. These include tyrosine kinase receptors (e.g., PDGFRA-B, NTRK1-3, RET^M918T, FGFR2^{A315del,Q778Pfs*11}, and KIT^L576P), a non-receptor tyrosine kinase (SRC^Y530F), the mitogenic ligand of FGFR (FGF6), adapter proteins (CRKL, RASGRP1, and CCDC25), and several genes involved in the MAP kinase pathway (e.g., HRAS, MAP3K8, and DUSP10).

Figure 3. Genotype-fitness maps and resistant ORFs of a third generation EGFR inhibitor (osimertinib) and a MEK inhibitor (binimetinib).

Genotype-fitness is shown across the replicates, with enhancement on fitness > 0 and > 0.1 for osimertinib 40 nM (A) and binimetinib 5uM (C), respectively. ORF with fitness > 0.1 and robust Z-score > 3 in all replicates are shown (black), including genotypes previously associated with clinical relapse samples (red), or pre-clinical studies (green). Resistance ORFs are annotated by affected gene pathways and processes for osimertinib (B) and binimetinib (D), along with their fitness, and prior identification in clinical relapse samples, or pre-clinical studies. E. ORF fitness robust Z-score in the presence of 40nM erlotinib or 40nM osimertinib. Cross-resistant ORFs were defined as having fitness >0.1 and robust Z-score > 3 in all replicates in the presence of either drug and are shown in red. F. As in (E), with 40nM erlotinib or 4uM binimetinib. G. As in (E), with 40nM osimertinib or 4uM binimetinib.

See also Figure S3, Table S1, S4, and Data S1.

Although erlotinib and osimertinib share 17 resistant ORFs, there were several clinically relevant differences that may underlie osimertinib’s increased efficacy. As anticipated, the EGFR^T790M genotypes were differentially sensitive to osimertinib. Also consistent with previous findings, we found that clones expressing EGFR^L8581R and ERBB2 are more sensitive to osimertinib than erlotinib (Liu et al., 2018; Masuzawa et al., 2017). Furthermore, we confirmed that ABCG2 is less effective at exporting osimertinib (Chen et al., 2016). In addition, osimertinib was effective against another common alteration – AXL upregulation. This added effectiveness in suppressing escape genotypes beyond EGFR^T790M may further contribute to the prolonged progression free survival of osimertinib as compared to erlotinib. Of note, while EGFR^C797S was not present in the ORFeome library, single ORF experiments confirmed higher osimertinib resistance compared with erlotinib (Figure S3A). Finally, the clusters of the genotype-fitness-dosage landscape for osimertinib (Figures S3B-S3D) showed a high level of agreement with those of erlotinib, further supporting the robustness of the inferred dose-genotype interactions.

EGFR mutant LUAD can resist binimetinib by MAP kinase pathway reactivation and transcriptional reprogramming

Several of the EGFR TKI resistant ORFs we detected are known druggable targets (Table S4). Moreover, the strong enrichment of MAP kinase pathway mediators among resistant ORFs in both erlotinib and osimertinib highlights the central role of MAP kinase pathway reactivation in driving resistance to EGFR inhibition, consistent with ongoing trials testing binimetinib, a MEK1/2 inhibitor, in combination with EGFR inhibitors (e.g., ).

We thus performed genotype-fitness mapping of single-agent MEK1/2 inhibition to inform combination strategies, as well as anticipate potential alterations resistant to the combination. We identified 57 binimetinib resistant ORFs (Figures 3C, 3D, Data S1, and Table S1; fitness > 0.1 and robust Z-score > 3 in all replicates, P < 1×10⁻¹⁰ by hypergeometric test of the resistant ORF intersection between replicates). While the clinical experience with MEK inhibition in lung cancer is limited, our genome-wide screen data is consistent with the melanoma experience, revealing enrichment in regulators of MAP kinase activity (FDR adjusted P = 1.6 × 10⁻³, Table S1). However, we did not observe any tyrosine kinase nor PI3 kinase ORFs as binimetinib resistant alterations potentially due to shunting of background EGFR signaling to the PI3-AKT pathway during MEK inhibition (Mendoza, Er and Blenis, 2011). Indeed, treatment of parental PC9 cells with binimetinib results in an increase in pAKT, in contrast to erlotinib or osimertinib treatment (Figure S2D). Further emphasizing the central role of MAP kinase reactivation in resistance to binimetinib, we identified a host of additional MAP kinase-related resistant ORFs (e.g., KRAS, BRAF^V600E, and MAPK1, Figure 3D), as well as MAP kinase pathway transcriptional targets (e.g., FOS, ETV1, and ATF3).

Cellular reprogramming may also contribute to binimetinib resistance. Members of the FOS family have also been implicated in transcriptional programs driving cell proliferation, differentiation, and EMT (Vallejo et al., 2017; Johannessen et al., 2013; Muhammad et al., 2016; Shao et al., 2014). Additionally, binimetinib resistant ORFs include transcription factors such as POU5F1/OCT4, SOX15, and HOXA6, which have been described to drive lineage reprogramming and stem-like properties, and are frequently overexpressed in cancer (Kumar et al., 2012). Finally, we observed multiple apoptosis regulators (BCL2L1, BCL2L2, and FASTK), which may mediate resistance through overcoming the pro-apoptotic stress of oncogene inhibition (Carné Trécesson et al., 2017).

Cross-drug comparisons of fitness maps to guide therapeutic combinations that minimize overlapping resistance

To rationally design combinations with low cross-resistance, we capitalized on the absolute fitness metric which allows direct comparison of genotype resistance across the three therapeutic contexts tested; erlotinib, osimertinib and binimetinib. In contrast to the high cross-resistance between the two EGFR inhibitors (Figure 3E), EGFR TKIs and binimetinib showed far fewer cross-resistant ORFs (Figures 3F and 3G). For example, receptor tyrosine kinases, which show strong resistance to EGFR TKIs (Figures 2D and 3B), do not result in resistance to binimetinib (Figure 3D). On the other hand, apoptosis regulators and some cell transporters, which lead to resistance to binimetinib (Figure 3D), are sensitive to EGFR inhibition (Figures 2D and 3B).

Thus, with the majority of resistant ORFs expected to be suppressed with combination therapy, only five ORFs were predicted to be cross-resistant to erlotinib and binimetinib and only two were predicted to be cross resistant to osimertinib and binimetinib (Figures 3F and 3G). KRAS^G13D and BRAF^V600E were predicted cross-resistant alterations in both combinations. These data suggest the potential effectiveness of combined EGFR and MEK inhibition in anticipating clonal evolution to resistance in EGFR-driven LUAD.

Unexpected drug-drug interaction is revealed by genotype-fitness mapping of resistance to combination therapy

Central to the application of genome-wide screens to combinatorial drug design is knowing whether intersecting resistance profiles from single-agent screens will be sufficient to predict cross-resistance. Thus, we empirically tested the drug combinations and observed the predicted cross-resistant ORFs based on the single-agent data (e.g., KRAS^G13D, BRAF^V600E, ABCG2, and SOX15 for erlotinib and binimetinib combination; Figures 4A, 4B, Data S1, S2, and Table S1). However, using the same fitness thresholds employed for single agents (fitness > 0.1 and robust Z-score > 3 in all replicate; P < 1×10⁻¹⁰ by hypergeometric test of the intersection of resistant ORF across replicates), 28 additional resistant ORFs were detected with erlotinib and binimetinib combination therapy that were not predicted by the monotherapy experimental data. Similarly, 26 additional resistant ORFs were detected with osimertinib and binimetinib combination therapy. Notably, only a minority of these unexpected combination resistant ORFs showed even borderline resistance in the single-agents screens (Figures S3E-S3G), demonstrating that combination therapy truly exposes drug-drug-genotype dependencies.

Figure 4. Genome-wide screen with drug combinations reveals genotypes with cross-resistance, as well as genotypes with higher than expected fitness with combination therapy.

Genotype-specific fitness with combination of 15nM erlotinib and 1uM binimetinib (A) and 5nM osimertinib and 1uM binimetinib (B), across replicates. Resistant ORFs were defined as fitness > 0.1 and robust Z-score > 3 across replicates. Resistant ORFs expected from cross-resistant ORFs in single-drug genotype-fitness mapping as in Figure 3F and G are shown in red. Unexpected resistance ORFs are shown in black. C. Residual plot of the ORF fitness measured from the empirical combination of erlotinib and binimetinib compared to the combination fitness predicted by a generalized linear model using ORF fitness from single-agent erlotinib and single-agent binimetinib experiments. Resistant ORFs expected from cross-resistant ORFs in single-drug genotype-fitness mapping as in Figure 3F are shown in red. Unexpected resistance ORFs are shown in black. D. As in (C), except with osimertinib and binimetinib combination therapy. Resistant ORFs expected from cross-resistant ORFs in single-drug genotype-fitness mapping as in Figure 3G are shown in red.

See also Figure S4, Table S1, Data S1, and S2.

To evaluate potential ORF-specific drug combination interaction, we modeled the combination fitness of each ORF based on the single agent fitness (Figures 4C, 4D, 5A-5C, S3B S3H, S4A, and S4B). The model for both drug combinations predicted an additive effect of the two drugs based on the response across the majority of clones (see STAR Methods). For example, the PIK3CA^H1047R ORF replicates had a mean fitness of −0.01 Δ(log2(cell number))/days in the erlotinib and binimetinib combination, similar to the model predicted fitness of 0.02 (Figure 5C). Likewise, the CRKL ORF replicates treated with the combination had a mean fitness of 0.09, close to the model-predicted fitness of 0.12. Similar behavior was observed in the osimertinib and binimetinib combination (Figure 5C).

Figure 5. Characterization and validation of drug-drug interactions across resistant genotypes.

A. Resistant ORF fitness across single-agent erlotinib (15nM), single-agent binimetinib (1uM), predicted combination fitness as in Figure 4, and measured combination fitness (15nM erlotinib and 1uM binimetinib). Resistant ORFs from the combination and high-dosage single-agent experiments were selected and those with consistent interaction behavior in both combinations are shown. (B) Z-score of the fitness residual of the measured combination fitness compared to the predicted combination fitness for erlotinib and binimetinib (left) and osimertinib and binimetinib (right). ORFs with Z-score > 2 are shown in red, and genes are annotated by gene categories highlighted in Figure 1-3. C. Representative genotype-combination interactions for buffering vs. additive ORFs. ORF fitness (2-6 replicates with standard error bars) are shown across single-agent therapy (no shading), model prediction for combination therapy (grey), and measured combination therapy (red). D. Drug printing allows high resolution validation of potential drug-drug interaction across a dose range for individual ORFs. E. Expected additive impact is shown for control ORF HcRed, dose invariance for EGFR^{T790M/C797S/L858R}, and paradoxical increased fitness for BRAF^V600E and ERBB2 with increased erlotinib doses. F. Fitness is measured via luminescence values in drug condition at day 5 as a proportion of luminescence in drug condition at day 0, and compared between binimetinib monotherapy (including 0.6, 0.8, and 1uM n = 12), vs. binimetinib in combination with erlotinib (including 0.6, 0.8, and 1uM binimetinib and 6nM and 8nM erlotinib, n = 24; left). P-values derived from t-test.

See also Figure S4, Table S1, Data S1, and S2.

In contrast, a small subset of genotypes revealed an unexpected buffering behavior in the presence of the drug combination, with higher fitness than predicted by the model (Figures 4C and 4D). For example, the KRAS^G13D ORF replicates had a measured mean fitness 0.61 in the erlotinib and binimetinib combination, significantly higher than the model-predicted fitness of 0.18 (Figure 5C). In another example, the mean ERBB2 ORF replicate combination fitness was measured at 0.42, well above the model-predicted fitness of 0.10 (Figure 5C). Similar findings were observed with binimetinib in combination with osimertinib. Overall, we identified 23 ORFs that showed buffering behavior in both combinations (i.e., binimetinib with erlotinib or osimertinib; Figures 5A and 5B, Table S1, and see Figures S4C-S4F for buffering ORFs in each combination individually). Notably, this group was enriched with ORFs that positively regulate the MAP kinase pathway (FDR adjusted P = 1.5 × 10⁻¹⁰), including several receptor tyrosine kinases (CSF1R, NTRK1-3, FGFR2^{A315del, Q778Pfs*11}, and PDGFRB^S180F), KRAS^G13D, and HRAS. Furthermore, several of these ORFs, including ERBB2 and KRAS^G13D, showed strong buffering such that their fitness in combined EGFR and MEK inhibition was paradoxically higher than that with single-agent MEK inhibition (Figures 5A-5C).

To validate these potential drug-drug interactions, we conducted drug treatment experiments for BRAF^V600E, ERBB2, EGFR^{T790M, C797S, L858R}, and control lentivirus across a range of EGFR TKI and binimetinib dosages (Figure 5D). The expected additive behavior was observed for the control ORF and erlotinib dose invariance for EGFR^{T790M, C797S, L858R}. Moreover, the putative buffering ORFs BRAF^V600E and ERBB2 show increased growth with the addition of erlotinib compared with binimetinib alone (Figures 5E and 5F). Similar results were found in validation experiments using osimertinib and binimetinib (Figure S4G), and with KRAS^G13D (Figure S4H).

Optimizing combination treatment strategies to minimize the evolution of resistance

Our data suggest that while combination therapy will suppress the majority of genetic variants in a given malignant population, some genotypes with buffering interactions will be better countered by binimetinib monotherapy (e.g., ERBB2 and KRAS^G13D) (Figures 5A-5C). Thus, alternating combination therapy with binimetinib monotherapy may provide greater overall growth suppression than combination therapy alone. To nominate an optimized treatment policy based on fitness measurements, we applied previously proposed optimization models (Chmielecki et al., 2011). Briefly, we modeled the overall growth of the malignant population across a range of therapeutic regimens that include varying proportions of the treatment cycle using erlotinib monotherapy, binimetinib monotherapy and the combination (Figure 6A and STAR Methods). As expected, when the modeled malignant population included only cells harboring ERBB2 or KRAS^G13D, two of the genotypes with strong buffering behavior, the optimal regimen was single-agent binimetinib throughout the cycle (Figures 6B and S5A, see Figure S5B and S5C). In contrast, when the modeled malignant population included only cells harboring PIK3CA^H1047R, a genotype with additive effect, the optimal regimen was identified as combination of erlotinib and binimetinib throughout the cycle (Figure 6C, see Figure S5D for combination with osimertinib).

Figure 6. Genotype-fitness map therapeutic optimization nominates cycling between binimetinib monotherapy and combination.

A. Schematic of treatment cycle optimization data; y-axis represents the proportion of time of the treatment cycle with drug A, and x-axis represents the proportion of time of the treatment cycle with drug B. Varying proportions across the optimization hypotheses space are illustrated for informative examples. In all subsequent analysis, a tumor growth model is initialized with 750,000 cells, and the population size is measured after 50 days, as in (Chmielecki et al., 2011). B. In a homogenous population of EGFR mutant cells that also harbor ERBB2, optimization landscape shows the lowest cell number is achieved with monotherapy binimetinib, as expected from the inferior inhibition of this genotype with combination therapy (as in Figure 5C). In contrast, for ORFs that show additive effect of the combination (as in Figure 5C), the optimal regimen includes combination therapy across the entire cycle (example given for PIK3CA^H1047R in (C)). To model a clonally diverse malignant population, the starting frequency of clones containing all genotypes across the genome-wide library was assigned proportionally to their fitness in DMSO controls (i.e., prior to therapy). The model shows optimization for the entire population involves partial binimetinib monotherapy alternating with binimetinib and erlotinib combination (D), with similar results with binimetinib and osimertinib combination (F). E. Growth over time is shown for the same starting clonally diverse population, comparing modeled growth with binimetinib monotherapy (red), combination erlotinib and binimetinib (blue), or the optimized treatment cycle of combination (40% of time) and single-agent binimetinib (60% of time, black). Error bars represent two standard deviations. Similar results shown with binimetinib and osimertinib combination (G). H. The optimized regimen is stable to varying frequencies of the ERBB2 clone until reaching a clonal frequency of 1:100 cells with combination of binimetinib with erlotinib.

See also Figure S5, Data S1, and S2.

Next, to identify the optimal policy for a genetically diverse population, such as seen in human cancer, we constructed a heterogeneous model of a malignant cell population, where each genotype is represented by a cell number proportional to its fitness prior to therapy (based on genotype-fitness maps in DMSO control). In this context, where cells are an admixture of multiple genotypes, including clones with additive effect and buffering effect, the optimal policy was a regimen with an alternating schedule of binimetinib monotherapy and erlotinib and binimetinib combination (Figures 6D and 6E). Similar results were observed with the combination of osimertinib and binimetinib (Figures 6F and 6G) and with a greedy optimization algorithm (Figures S5E and S5F).

Nevertheless, stochastic variation in clonal frequencies may not be adequately captured in the fitness measured in the DMSO controls. To examine the impact of varying clonal frequency on the suggested optimal policy, we modeled the malignant population with increasing frequencies of genotypes associated with strong buffering behavior (e.g., ERBB2 and KRAS^G13D) and observed relative stability of the optimal policy up to clonal frequencies of buffering ORFs ERBB2 at 10⁻² and KRAS^G13D at 10⁻⁴ (Figures 6H and S5G, see Figures S5H and S5I for combination with osimertinib), suggesting that higher clonal frequencies would require adjustment of the combination therapy schedules. Notably, these clonal frequencies approximate the sensitivity of leading-edge liquid biopsy methods suggesting the ability to integrate clonal frequency measurement with genotype-fitness mapping in designing optimal combination strategies.

DISCUSSION

Combining drugs with low cross-resistance has long been proposed as a strategy to suppress resistance and address tumor evolution (Bozic et al., 2013). However, prospective identification of such combinations has lagged, as using clinical samples to characterize drug resistance profiles is limited by scale and limited to therapies in clinical use. This problem is compounded by an expanding number of potential drug combinations, and an even higher number of potential dose and schedule combinations. Here, we use serial sequencing to transform functional genomic screening into genotype-fitness maps, which can be readily compared across screen conditions. Thus, our approach can be used to guide drug combination development by comparing resistant alterations across drug dosages, between different drugs, and between single-agent drugs and their combinations.

Capitalizing on the ability to compare fitness across screen conditions, we observed varying dose-dependency across resistant genotypes, which may be of clinical relevance given the inter- and intra-patient variability in drug delivery (Fu, Nowak and Bonhoeffer, 2015). These distinct dose-dependencies affected two of the key signaling networks activated in EGFR-mutant LUAD – the MAP kinase and PI3 kinase pathways. Whereas the former shows a fitness advantage only in the presence of strong EGFR inhibition, the latter shows a more modest fitness advantage across the entire dose range. Mechanistically, the paradoxical increase in fitness with higher drug dose, seen with ORFs activating the MAP kinase pathway, may result from a potential toxic effect of MAP kinase pathway overstimulation (Unni et al., 2015; Leung et al., 2019) by MAP kinase pathway ORFs expressed in the context of mutated EGFR. These data also demonstrate the advantage of a cellular strategy capable of stimulating both pathways, such as via EGFR^T790M,L858R, ERBB2, and KRAS^G13D ORFs, which may synergize to result in the robust resistance across the dose range.

Genotype-fitness mapping also addresses the challenge of combination discovery and optimization by enabling direct comparison of resistance profiles across different drugs. For example, our data suggest that combination of osimertinib and binimetinib is likely an effective combination strategy given the low number of genotypes that confer resistance to both drugs. Nonetheless, when the combination was tested directly in a dedicated genome-wide screen, unexpected genotype-specific drug-drug interactions were observed that resulted in reduced combination therapy efficacy. This highlights the importance of using combination screens to expose the signaling network architecture of resistance and to rationally design combination therapy (Fitzgerald et al., 2006). The protective effect of EGFR inhibition in this context is reminiscent of a similar phenomenon observed in BRAF-mutant melanoma, where compensatory upregulation of MAP kinase signaling results in higher growth in the presence of BRAF inhibition than when it is withdrawn (Moriceau et al., 2015). As MAP kinase activation plays a central role across cancer, these drug-drug interactions observed in EGFR-driven LUAD may also be relevant for combination therapy in other settings. This conceptual framework can be readily integrated with loss of function screens in future studies to complement ORF screens and map fitness across the entire spectrum of genomic alterations potentially capable of conferring resistance.

In summary, we envision that wide application of genotype-fitness mapping across drug combinations may provide a tractable avenue to explore the vast combinatorial treatment hypotheses space, accelerating the development of effective combination strategies. To explore how these data may be applied for algorithmic treatment optimization, we modeled a clonally diverse cancer population, and identified a combination policy that optimizes growth suppression across a clonally diverse malignant population. Thus, genotype-fitness maps may be integrated into personalized treatment frameworks, considering the dominant clonal composition of each patient’s tumor. Through emerging liquid biopsy technologies, such clonal monitoring may further enable the application of genotype-fitness map insight in continuous optimization, to address resistant clones as they emerge (Adalsteinsson et al., 2017).

STAR METHODS

LEAD CONTACT AND MATERIALS AVAILABILITY

This study did not generate new unique reagents. Further information and requests for resources and materials should be directed to and will be fulfilled by the Lead Contact, Dan A. Landau, MD, PhD (dlandau@nygenome.org).

EXPERIMENTAL MODEL DETAILS

PC-9 (del E746_A750) lung adenocarcinoma cells have been previously described (Ono et al., 2004) and were obtained from The Broad/Novartis Cancer Cell Line Encyclopedia internal repository. The cells were maintained in Cellgro RPMI-1640 (Thermo Fischer Scientific, Waltham, MA, USA) and supplemented with 10% fetal bovine serum (Sigma Aldrich, St. Louis, MO, USA) and a combination of Giboco penicillin (100 units/mL) / streptomycin (100 μg/mL) (Thermo Fischer Scientific, Waltham, MA, USA). They were also tested to confirm the absence of mycoplasma contamination using the MycoAlert Detection Kit (Lonza, Basel, Switzerland) according to the manufacturer’s instructions.

Additional further steps may include examination of how differences in background germline and somatic genetics influence resistant alterations to further nuance genotype-fitness mapping. However, prior experience has shown that single cell lines tend to validate at a rate of 70-85% when tested in other models (Johannessen et al., 2010; Johannessen et al., 2013; Wilson et al., 2015).

METHOD DETAILS

PC-9 Dosage Optimization Experiments

In order to allow a comprehensive read-out of the compendium of erlotinib resistant alterations, erlotinib dosages were optimized to achieve maximal rescue with the positive control EGFR^{T790M, L858R} ORF at a minimal drug concentration (Johannessen et al., 2013; Wilson et al., 2015). This methodology is standard in the field to enrich for biology that reflects drug-resistant tumors as sublethal drug dosages have been shown to reveal physiological relationships in vitro with higher sensitivity (Yeh, Tschumi and Kishony, 2006). Initially, cumulative population doublings were measured for PC-9 parental cells and PC-9 cells infected EGFR^{T790M ,L858R}in the presence of DMSO, 30, 50, 100, or 200 nM erlotinib. Cells were passaged in 6-well dishes every 3-4 days for a total of 15 days. They were maintained at ~15% confluency if their population doublings were positive, or their seeding confluency was increased if their population doublings were negative (Figure S6A). These experiments were followed by measuring the cumulative population doublings of PC-9 parental cells at higher resolution (Figure S6B) within a lower range of erlotinib dosage: DMSO, 5, 15, 25, 40, 60, and 100 nM. The cells were passaged every 3-4 days for a total of 17-21 days in T75 flasks at their optimal growth density: DMSO cells were maintained at ~15% confluency at each passage, and cells treated with erlotinib were either passaged at ~30% confluency or their drug-media refreshed if the confluency visually appeared less than 50%. The sublethal erlotinib dosage of 25nM was chosen for Figure 1 based on these growth optimization experiments.

Functional screens are typically limited to a single concentration, often tuned to optimize technical success. However, while intra-tumoral drug concentrations are not often measured effectively, limited studies have suggested large variability (Lankheet, 2015). To study the response of resistance alterations to a range of dosages, we used the aforementioned optimization experiments to select the additional erlotinib dosages of 5, 15, and 40 nM and generated the genotype-fitness-dosage map in Figure 2.

We confirmed the clinical relevance of this dose range by comparing the resistant ORFs at each dosage to the clinical literature. Notably, eight of the ten clinically validated resistant alterations within our ORF library were detected across the dosage range (Table S2). Furthermore, different alterations were identified at different dosages, highlighting the value of testing multiple dosages. The two clinically validated resistant alterations that were not detected include MET and YES1, likely due to technical failures of these ORFs within the screening library as seen previously (Sharifnia et al., 2014; Fan et al., 2018).

In order to allow a comprehensive read-out of the compendium of resistant alterations for osimertinib and binimetinib, dosages of these drugs were optimized to achieve equipotent growth suppression to the moderate (15 nM) and high (40 nM) dosages of erlotinib (Figure S6C). PC-9 population doubling time was measured for 0.5, 1, 5, 10, 20, and 40 nM osimertinib and 0.5, 1, and 4 uM binimetinib. 5 and 40 nM osimertinib, 1 and 4 uM binimetinib, and combinations of the moderate dosages (15 nM erlotinib + 1 uM binimetinib and 5 nM osimertinib + 1uM binimetinib) were selected for the screening experiments. The exposure of biologically and clinically expected resistant ORFs in these experiments suggest that the selected dosages were appropriate drug concentrations.

Screening of the CCSB-Broad Lentiviral Expression Library

The Center for Cancer Systems Biology (CCSB)-Broad Lentiviral Expression library has been previously described (Yang et al., 2011). We note that some genes are represented by multiple ORFs in the library due to random redundancies in the library generating procedure. The library consists of 17,255 ORFs representing 12,728 unique genes and 35 ORFs representing genes with known cancer drivers introduced. For each single-agent or combination drug genotype-fitness map, the ORF with the highest fitness was selected to represent its respective gene or mutant gene.

For the analyses involving the fitness across different drug conditions, the genotype-fitness maps were intersected after drop-outs were removed in each condition, the fitness of each ORF was summed across all drug conditions, and the ORF with the highest average fitness across all conditions was selected to represent its respective gene or mutant gene. These analyses included the genotype-fitness-dosage landscape, the single-agent cross-resistance analysis, the generalized linear models, the buffering analysis, and the evolutionary modeling analyses. The 35 oncogenic variants included in the library constitute clinically validated drivers at the time the library was developed. EGFR^{T790M, L858R} was added as a positive control of erlotinib resistance. KRAS^G13D was included instead of the more powerful KRAS^G12 mutations, as these tend to overtake pooled cellular screening experiments.

In the first set of experiments, PC-9 cells were infected with the library in three replicates (5.4 × 10⁷ cells/replicate, 36% infection efficiency) followed by selection with puromycin and achieving allelic representation of 1153 cells per ORF. Cells were then separated into 48 replicates. For three replicates, the cells were immediately lysed and treated for gDNA extraction. The remaining 45 replicates were treated with either 0.01% DMSO (drug vehicle), 5 nM erlotinib, 15 nM erlotinib, 25 nM erlotinib, or 40 nM erlotinib (9 replicates each). For each drug condition, cells in the replicates were lysed and gDNA was isolated and sequenced as described below after 7 days, 14 days, or 21 days of exposure (3 replicates each). For drugs, single-use 1000X working stocks were made of each drug in DMSO. At time of drug addition, working stocks were thawed and diluted to 10X in media. Cell suspension and media were combined in a flask, then the 10X drug solution was added to bring the final drug concentration to 1X.

In the second set of experiments, PC-9 cells were infected with the library in two replicates (1.08 × 10⁸ cells/replicate, 45% infection efficiency), achieving an allelic representation of 1981 cells per ORF. Cells were then separated into 30 replicates. For two replicates, the cells were immediately lysed and the ORF barcodes were sequenced and counted. The remaining 28 replicates were treated with either 0.01% DMSO (drug vehicle), 5 nM osimertinib, 40 nM osimertinib, 1 uM binimetinib, 4 uM binimetinib, 15 nM erlotinib + 1 uM binimetinib, or 5 nM osimertinib + 1 uM binimetinib (4 replicates each). For each drug condition, cells in the replicates were lysed, amplified, and sequenced as described below after 7 days or 14 days of exposure (two replicates each). Drug additions were performed as described above.

Drug-Sensitivity Measurements

PC-9 cells were infected with either control lentivirus HcRed, EGFR^{T790M, C797S, L858R}, BRAF^V600E, or ERBB2. Expression was confirmed by western blot. Cells expressing each ORF were seeded into 384-well, white-walled, clear bottom plates at a density of 800 cells/well. 24 hours after seeding, combination drugs were administered using an HP D300 Digital Dispenser in matrix format. The erlotinib and binimetinib matrix included concentrations of 8 nM, 6 nM, 4 nM, 2 nM, and 0 nM erlotinib by 1 uM, .8 uM, .6 uM, .4 uM, .2 uM, and 0 uM binimetinib (4 replicates each). The osimertinib and binimetinib matrix included 5 nM, 4 nM, 3 nM, 2 nM, 1 nM, 0 nM osimertinib by 1 uM, .8 uM, .6 uM, .4 uM, .2 uM, and 0 uM binimetinib (4 replicates each). Drug was administered such that the final volume of DMSO did not exceed 0.5%. The cells were then incubated for 96 hours and cell luminescence was measured using the Cell Titer-Glo assay (Promega, Madison, WI, USA) according to the manufacturer’s instructions.

Additionally, using the same methodology as above, PC-9 cells were infected with control lentivirus LacZ or KRAS^G13D and exposed to DMSO and single-agent erlotinib dosages of 0.4nM, 0.7nM, 1.3nM, 2nM, 3.3nM, 5.3nM, 8nM, 12.7nM, 20nM, 31.3nM, and 50nM. Similarly, PC-9 cells were infected with control lentivirus LacZ or EGFR^C797S and exposed to DMSO and either single-agent erlotinib or single-agent osimertinib at dosages of 50nM, 80nM, 120 nM, 200nM, 320nM, 506nM, 786nM, and 1252nM.

Immunoblotting

PC-9 cells infected with either control lentivirus HcRed, control lentivirus eGFP, KRAS, KRAS^G13D, EGFR, EGFR^{T790M, C797S, L858R}, BRAF or BRAF^V600E were exposed to either 0.01% DMSO (drug vehicle), 15nM erlotinib, 40nM erlotinib, 5nM osimertinib, 40nM osimertinib, 1uM binimetinib, 4uM binimetinib, 15nM erlotinib + 1uM binimetinib, 40nM erlotinib + 4uM binimetinib, 5nM osimertinib + 1uM binimetinib, or 40nM osimertinib + 4uM binimetinib. After 24 hours, the ORF-expressing cells were rinsed once with ice-cold PBS and lysed in 1% NP-40, 2mM EDTA [pH 8], 50mM Tris [pH 7.5], 25mM NaF, and 150mM NaCl with protease inhibitors (Roche, Basel, Switzerland) and Calbiochem phosphatase inhibitors (Millipore Sigma, Burlington, MA, USA). The lysate was reduced and denatured (95°C), then fractionated by SDS-polyacrylamide gel electrophoresis on 4%-20% Tris/Glycine gels (Invitrogen, Carlsbad, CA, USA). The resolved protein was transferred to nitrocellulose membranes, treated with LiCOR blocking reagents, stained with the antibodies listed in the Key Resources Table, and fluorescence was detected with an Odyssey CLx Infrared Imaging System.

QUANTIFICATION AND STATISTICAL ANALYSIS

DNA sequencing and ORF-specific read count

Genomic DNA (gDNA) was isolated using a Maxi (3 × 10⁷−1 × 10⁸ cells) kit according to the manufacturer's protocol (Qiagen, Hilden, Germany). PCR and sequencing were performed as described previously (Doench et al., 2016). Samples were sequenced and read counts normalized as previously described (Bockorny et al., 2018). Candidate mediators of resistance were defined as ORFs that had a LFC greater than or equal to 1.5 following treatment compared with the ETP sample at all drug concentrations employed. Standard deviation, standard error, and confidence intervals were calculated based on the replicates available in the experiments and where possible based on multiple ORFs covering the same gene.

RPM-based resistance analysis

Reads-per-million (RPM) was calculated for every experimental sample, any drug condition and any time-point, by dividing the ORF-specific read counts by the total number of million reads in this experimental sample. RPM was calculated on the cumulative read counts from the aggregate of all replicates and for each replicate separately in each experimental sample. The RPM fold-change for each ORF (indexed by i) was calculated for each drug condition at day 15 as follows,

$$RPM\_F{C}_{i,drug}={\log }_2\left(\frac{RP{M}_{i,drug}}{RP{M}_{i,DMSO}}\right).$$

ORFs were considered drop-outs and excluded from fold-change resistance detection if the reads-per-million were 0 in two or more replicates in either the drug or DMSO condition used in the calculation. Robust Z-score was determined for each ORF in each replicate of each condition based on RPM fold-change, using the following:

$${\widehat{R}}_{i,drug}=\frac{RPM\_F{C}_{i,drug}-median\left(RP{M}_{F{C}_{drug}}\right)}{MAD\left(RPM\_F{C}_{drug}\right)}$$

Where RPM fold-change_i corresponds with ORF specific RPM fold-change and median and median absolute deviation (MAD) are calculated over the RPM fold-change all ORFs. For each drug treatment condition, ORFs with a robust Z-score ≥3 (at least 3 median absolute deviations above the median) and RPM_FC >1 in all replicates were identified as resistant ORFs.

Fitness (cell growth rate) based resistance analysis

The cumulative number of cells per time point and drug condition were estimated using a 500 uL aliquot of the total cell suspension, quantified on a Beckman Coulter Z2 Coulter Particle and Size Analyzer (Brea, CA, USA). ORF-specific cumulative number of cells per time point was calculated using the proportion of read supporting the specified ORF out of the total reads from the sample (RPM). ORF-specific growth rate (fitness) was calculated using linear regression over log₂ transformed cumulative number of cells per ORF at all time points collected - 0, 8 , 15, 22 days for the 25nM erlotinib experiments. Fitness was calculated using, 0, 8, and 15 days for the other erlotinib dosages. Fitness was calculated on the ORF-specific cumulative cell number from the aggregate of all replicates and for each replicate separately in each drug condition. Furthermore, ORFs were considered drop-outs and excluded from fitness resistance detection if the reads-per-million were 0 in two or more replicates in any time point used in the fitness calculation. Using fitness, we also observed a somewhat higher correlation across replicates, with R² of 0.92 (Pearson’s correlation), compared with the traditional fold-change enrichment approach, which had an R² of 0.87 (Figures S2E and S2F).

Robust Z-score was determined for each ORF in each replicate of each condition based on the fitness, using the following formula-

$${\widehat{F}}_i=\frac{f_i-median\left(f\right)}{MAD\left(f\right)}$$

Where f₁ corresponds with ORF specific fitness and median and MAD are calculated over all ORFs. For each drug treatment condition, ORFs with a robust Z-score > 3 (greater than 3 median absolute deviations above the median) and fitness > 0.1 Δ(log₂(cell number))/days in all replicates were identified as resistant ORFs. The Z-score threshold of > 3 was empirically optimized to balance recall and precision. To indirectly measure recall we define a set of clinically validated resistance variants based on a comprehensive literature survey. This survey includes the mutant genes, EGFR^{T790M, L858R} , EGFR^T790M, PIK3CA^E545G, PIK3CA^H1047R, BRAF^V600E, and the over-expression of ERBB2, AXL, MET (Morgillo et al., 2016), EGFR (Jia et al., 2016) and YES1 (Fan et al., 2018). We then measured positive enrichment as a function of Z-score threshold as seen in Figure S6D. To indirectly measure precision, we evaluated the enrichment of putative resistant ORFs nominated in two replicates that exhibit a positive growth rate (fitness > 0) in the third replicate. Double replicate nominated resistant ORFs exhibiting negative growth rates in the third replicate were considered false positives and those exhibiting positive growth rates were considered as true positives. Through this analysis, we found that a threshold of three Z-scores reflected a reasonable balance between the two (Figure S6E). It is worth noting that because we optimized for recall and precision, our model is optimized to make high quality drug resistance, but not drug sensitivity, detections. The fitness threshold (clonal doubling time of 10 days or shorter) was an added conservative metric to ensure that the resistance ORFs represent biologically meaningful alterations. Using this threshold, fitness identified two additional clinically validated ORFs (EGFR^T790M,L858R and EGFR^WT), which failed to reach significance based on relative enrichment alone (the former due to a technical reason – drop-out in DMSO control) (Figure S1D). Additionally, we tested thresholds with a Z-score of 2, 2.5, 3.5, and 4 and found that fitness identified a larger number of clinically annotated resistant alterations than fold-change enrichment across every threshold used (Figure S6F).

For our experiments with two replicates (osimertinib, binimetinib, and the drug combinations), we used a hypergeometric test to quantify the statistical significance of the observed intersection of independent replicates. However, this approach has not been applied to data with greater than two sets. As three replicates were used in the erlotinib studies, we have applied the SuperExactTest, which is an approach developed to efficiently calculate the exact probability of multi-set intersections (Wang, Zhao and Zhang, 2015). SuperExactTest was used with the MSET() R command with lower.tail=FALSE. We used this approach to measure the statistical significance of the overlap between three sets of independently sampled data. Furthermore, we have compared the three-set to the two-set approach for erlotinib resistance detection at 25 nM and found the results to be similar. The SuperExacTest produced a P value of 2 × 10⁻⁸³. For the hypergeometric test, P = 3 × 10⁻⁷² for the intersect of replicates A and B, a P value = 4 ×10⁻⁵⁰ for the intersect of replicates A and C, and a P value = 2 ×10⁻⁴⁵ for the intersect of replicates B and C.

The ordering of ORFs in the figures representing genotype-fitness-dosage landscapes was determined by calculating the difference of each ORF’s fitness and the population median, setting all negative values to 0, weighting each condition by drug dosage (DMSO=1, 5nM=2, 15nM=3 , 25nM=4, and 40nM=5), summing the weighted scores across all drug conditions, and ranking the ORFs in descending order.

Dose-genotype dependency clustering

ORF fitness was calculated and resistance ORFs were detected in DMSO and 5 nM, 15 nM, 25 nM, and 40 nM erlotinib using the methodology previously described. All resistant ORFs (n = 68) detected at any individual dosage were selected for k-means clustering. Fitness across DMSO and each dosage of erlotinib for each selected resistant ORF was clustered by k-means, using kmeans() R command with centers set to 5, using within groups sum-of-squares to optimize the number of clusters (Figure S2B). Hierarchical clustering with an L1 metric and ward linkage was also applied to assess dose response gene clusters and the results showed strong similarity to the k-means clusters (Figure S2C).

Using the same methodology, we clustered ORFs resistant to 5 and 40 nM osimertinib by fitness (Figure S3C). Consistent with a somewhat lower resolution afforded by two dosages (compared to four in erlotinib), we identified four clusters, compared with the five clusters that define the erlotinib dose-genotype interaction. The dose-genotype clusters show strong consistency between erlotinib and osimertinib. Similar to erlotinib C1, osimertinib cluster C1 shows no decrease in fitness as osimertinib dosage increases, and shares both of the expected overlapping ORFs, KRAS^G13D and CSF1R. Osimertinib cluster C2 is similar to erlotinib C2, as it shows moderate resistance, and decreasing fitness with increased osimertinib dosage. This cluster shared CRKL, FGFR^A315del, Q778Pfs*11, KIT^L576P, and NTRK2 with the corresponding erlotinib cluster. Osimertinib cluster C3 showed the same atypical behavior as C3 in erlotinib, with increased fitness advantage at higher dosage, with substantial ORF overlap (e.g., BRAF^V600E, SRC^Y530F, RET^M918T, MAP3K8, and HRAS). The high level of agreement across these two different drug models further supports the robustness of the inferred dose-genotype interactions.

Gene ontology enrichment

Gene Ontology enrichment was carried out using the GOrilla on-line tool with the two unranked lists setting, checking the detected resistance ORF set for each drug condition or the buffering ORF set against the background list of all non-drop out ORFs included in each respective analysis (Eden et al., 2009). FDR corrected p-values are reported for the annotation enrichment.

Generalized linear model and residual plots

Fitness of the ORFs was calculated cumulatively across replicates as previously described for single-agent 15 nM erlotinib, 5 nM osimertinib, and 1uM binimetinib, and 15 nM erlotinib + 1 uM binimetinib and 5 nM osimertinib + 1 uM binimetinib combinations. Drop-outs were excluded as previously described and additionally ORFs with cumulative fitness R-squared values < 0.8 in each sample were excluded. Fitness was calculated across day 0, 8, and 14 for osimertinib and binimetinib, and day 0 and 14 for erlotinib and binimetinib. Remaining ORFs (n = 12,453) that intersected across single-agent 15nM erlotinib, single-agent 1uM binimetinib, and combination 15 nM erlotinib + 1 uM binimetinib were included in a generalized linear model, using lm() R command with default settings, which predicted the combination fitness based on the single agent fitness across all ORFs. Residual fitness was calculated for each individual ORF comparing the actual combination fitness to that predicted by the generalized linear model. ORFs with a residual value > 2 standard deviations above the mean residual value of the population were considered “buffering” ORFs. The same methodology was repeated for osimertinib and binimetinib, using a total of 12,317 ORFs after removal of drop-outs and intersection among the different drug conditions. We note that the linear modeling does not address potential drug synergy. To determine the validity of this assumption, we measured the response additivity combination index (Foucquier and Guedj, 2015). Erlotinib and binimetinib combination yielded a near additive response (mean+/− SD of 1.08 +/−0.07) (Figure S6G), and osimertinib and binimetinib combination yielded and modest antagonistic response (1.56+/−0.07) (Figure S6H).

Buffering genotype validation

Cell viability was calculated using the drug treated luminescence as a percentage of ETP (ORF-infected cells treated at Day 0). For each of the 4 replicates, cell viability in each well was smoothed by taking the median of the adjacent cell in each direction. Contour plots were generated by taking the median of all 4 replicates. Box plots were generated by using the smoothed median values of each replicate.

Evolutionary modeling for treatment optimization

Genotype-fitness maps can be used to in-silico predict and optimize treatment regimens by using a simulated evolutionary model. Specifically, the fitness of each ORF (n = 12,728) in DMSO can be derived from the total initial population and the ORF specific fitness in DMSO with the following equation (assuming equilibrium distribution)-

$$N_i(t=0)=N(t=0)\frac{f_i}{{\sum }_i{f}_i}$$

Where f_i corresponds with ORF specific fitness, N_i correspond with ORF-specific population size and N corresponds with the total population size.

For each time point t and F_A, the fitness coefficient for condition A (single agent A), the ORF-specific population size is dependent on the ORF-specific fitness at that condition,

$$N_i\left(t\right)=N_i\left(0\right)2^{F_{A}t}$$

When considering a combination treatment with complex duty cycle between single agents and combined therapy, the overall growth rate depends on the superposition of fitness value in each condition and the fraction of time spent in each condition (single agent vs. combined therapy). This ORF-specific growth was simulated using the equation below-

$$N_i\left(t\right)=N_i\left(0\right)2^{(F_A{R}_A+F_B{R}_B+F_{AB}[1-(R_A+R_B)]t}$$

Where {F_A, F_B, F_AB} are the fitness coefficient for condition A (single agent A), condition B (single agent B) and combined therapy, respectively. Duty cycle of condition A alone (single agent), condition B alone and combination is defined by {R_A, R_B, [(R_A + R_B) − 1]}, where 0 ≤ R_A + R_B ≤ 1.

Different duty cycles of single agent and combination treatment exposure were simulated by applying ORF-specific fitness in each condition. Cycle times of single agent exposure were tested from 0 to 100% of the time by 10% increments. This approach was carried out using 15nM erlotinib, 1uM binimetinib, and the combination 15nM erlotinib + 1uM binimetinib for 50 days. It was repeated using 5nM osimertinib, 1uM binimetinib, and the combination 5nM osimertinib + 1uM binimetinib. Analytical simulation estimated the total number of cells at day 50 as a function of the drug regimen (duty cycle of the single and combined agents, Figure 6). Note that the cycle time noted in these figures correspond to the total fraction of time used with each drug (single agent time and combination time together).

For the above conditions, total tumor population size was inferred for different time points including (30, 35, 40, 45, 50 days) and plotted with confidence intervals. Standard deviation bars were generated from the gene with the largest number of ORFs in the library, IGHM. Standard deviation was calculated among the fitness from all 36 replicates of all 12 ORFs in DMSO and then applied to the total tumor population size at each time points.

This in silico procedure was also carried out for homogenous cell populations of 750,000 KRAS^G13D cells, ERBB2 cells, and PIK3CA^H1047R cells, and on varying clonal frequency with increasing frequencies of genotypes associated with buffering behavior (e.g., KRAS^G13D) between 1/700,000 to 1/1 cells.

To understand the precise temporal dynamics of drug cycling (i.e. which therapy to start with, how often to cycle), we used a greedy (local optimization) algorithm to nominate treatment policies that minimize tumor growth on a daily basis (Figures S5E and S5F). The population was composed of 170,255 cells – 10 cells per ORF. At each time point (once per day), tumor growth was simulated under four conditions, combination therapy (erlotinib 15 nM + binimetinib 1 uM), 15 nM erlotinib monotherapy, 1 uM binimetinib monotherapy, or DMSO, and a greedy algorithm was applied such that the treatment that minimizes net tumor size in the following time point is selected. We modified the greedy algorithm such that it only switched treatment conditions if tumor size shrank by at least 2%, to avoid frequent fluctuations. In this algorithm, cycling occurs in response to changing clonal frequencies, as the therapy that minimizes overall tumor growth depends on the remaining clonal sizes and growth rates. Genetically heterogeneous tumors grew the slowest under a treatment policy starting with combination (EGFRi and MEKi) therapy followed by a switch to MEKi monotherapy, and eventually cycling between both (Figures S5E and S5F).

DATA AND CODE AVAILABILITY

For reproducibility all input data are available on Mendeley (http://dx.doi.Org/10.17632/3mjb2z2b7g.1) and code available on GitHub (https://github.com/pobolan/LUAD-Tumor-Evolution).

Supplementary Material

1

Data S1. Single-agent ORF reads, reads-per-million, and fitness data, Related to Figure 1-6

2

Data S2. Combination ORF reads, reads-per-million, and fitness data, Related to Figure 4-6

4

Table S1. Gene set enrichments for resistant alterations, Related to Figures 1 and 3-6

KEY RESOURCES TABLE

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Antibodies
EGFR	Cell Signaling	2239
Phospho-EGFR	Cell Signaling	3777
AKT	Cell Signaling	2920
Phospho-AKT	Cell Signaling	4056
ERK	Cell Signaling	9107
Phospho-ERK	Cell Signaling	4094
cofilin	Cell Signaling	5175
Chemicals, Peptides, and Recombinant Proteins
Erlotinib	Selleck Chemicals	S7786
Osimertinib	Selleck Chemicals	S7297
Binimetinib	Selleck Chemicals	S7007
Critical Commercial Assays
Cell Titer-Glo	Promega	G7570
Beckman Coulter Z2 Coulter Particle and Size Analyzer	Brea	6605700
Deposited Data
Gene set enrichments for resistant alterations	This work	Table S1 http://dx.doi.org/10.17632/ddchgv6kwm.1
Single-agent ORF reads, reads-per-million, and fitness data	This work	Data S1 http://dx.doi.org/10.17632/fwk37v7xr5.1
Combination ORF reads, reads-per-million, and fitness data	This work	Data S2 http://dx.doi.org/10.17632/6c5vpv7s92.1
Raw data (ORF counts)	This work	http://dx.doi.org/10.17632/3mjb2z2b7g.1
Experimental Models: Cell Lines
PC-9	Broad/Novartis Cancer Cell Line Encyclopedia internal repository	PC9_LUNG
Recombinant DNA
Center for Cancer Systems Biology (CCSB)-Broad Lentiviral Expression library	GPP	CP0012
EGFR	GPP	BRDN0000560664
EGFR (T790M, L858R)	GPP	BRDN560707
KRAS	GPP	BRDN0000550205
KRAS (G13D)	GPP	TRCN0000489180
BRAF	GPP	BRDN0000561861
BRAF (V600E)	Generated in-house	N/A
ABCB1	GPP	TRCN0000487121
SOX15	GPP	TRCN0000473478
MET	GPP	TRCN0000471392
CTNNB1	GPP	BRDN0000550122
CCNE1	GPP	BRDN0000397568
MYC	GPP	BRDN0000553495
MEK2	GPP	TRCN0000486073
PDGFRA	GPP	TRCN0000491475
NTRK1	GPP	TRCN0000469111
KIT	GPP	TRCN0000491939
FOXA1	GPP	TRCN0000481564
AXL	GPP	TRCN0000480014
MAPK1	GPP	BRDN0000998163
ERBB2	GPP	TRCN0000488513
eGFP	GPP	BRDN0000559466
HcRed	GPP	BRDN0000464765
Software and Algorithms
R (version: 1.0.153)	R Development Core Team	https://www.R-project.org
GOrilla	European Union FP6 funded Multi Knowledge Project	http://cbl-gorilla.cs.technion.ac.il/
SuperExactTest	Wang, Zhao and Zhang, (2015)	https://cran.r-project.org/web/packages/SuperExactTest/index.html
DMSO fitness analysis code	This work	https://github.com/pobolan/LUAD-Tumor-Evolution
5nM erlotinib fitness analysis code	This work	https://github.com/pobolan/LUAD-Tumor-Evolution
15nM erlotinib fitness analysis code	This work	https://github.com/pobolan/LUAD-Tumor-Evolution
25nM erlotinib fitness analysis code	This work	https://github.com/pobolan/LUAD-Tumor-Evolution
40nM erlotinib fitness analysis code	This work	https://github.com/pobolan/LUAD-Tumor-Evolution
Binimetinib fitness analysis code	This work	https://github.com/pobolan/LUAD-Tumor-Evolution
Osimertinib fitness analysis code	This work	https://github.com/pobolan/LUAD-Tumor-Evolution
Erlotinib and binimetinib combination fitness analysis code	This work	https://github.com/pobolan/LUAD-Tumor-Evolution
Osimertinib and binimetinib combination fitness analysis code	This work	https://github.com/pobolan/LUAD-Tumor-Evolution

^*GPP – Broad Institute Genomic Perturbations Platform.

Highlights.

• Generate genome-wide genotype-fitness maps for EGFR inhibitors and a MEK inhibitor

• Expose differential dose-fitness relationships of resistant genotypes

• Identify genotypes where drug combination is less effective than the single-agent

• Nominate an alternating treatment schedule to forestall resistant disease