Mutational fitness landscape and resistance

Abstract

Robust technology has been developed to systematically quantify fitness landscapes that provide critical opportunities to improve our understanding of drug resistance and define new avenues to develop drugs with reduced resistance susceptibility. We outline the critical importance of drug resistance studies and the potential for fitness landscape approaches to contribute to this effort. We describe the major technical advancements in mutational scanning, which is the primary approach used to quantify protein fitness landscapes. There are a number of complex steps to consider in planning and executing mutational scanning projects including developing a selection scheme, generating mutant libraries, tracking the frequency of variants using next-generation sequencing, and processing and interpreting the data. Key experimental parameters impacting each of these steps are discussed to provide to aid in planning fitness landscape studies. There is a strong need for improved understanding of drug resistance, and fitness landscapes provide a promising new approach.

Introduction

The promise of fitness landscapes to study drug resistance

The efficient and systematic quantification of drug resistant mutations provides a powerful tool that can contribute to the development of inhibitors with reduced potential for resistance (Figure 1). Once a selection scheme and library have been developed, fitness landscapes can be efficiently generated in parallel with hundreds to thousands of inhibitors. These comprehensive maps of mutations that are compatible with function in the presence of inhibitors could be mined to determine structure-resistance relationships and to guide the development of inhibitors with reduced susceptibility to resistance. In addition, resistant maps also provide the opportunity to identify inhibitors with non-overlapping resistance profiles that could be used in combination to decrease susceptibility to resistance. With the available advancements in technology and the increasing threats to human health from drug resistance, it is an apt time to apply fitness landscapes to understand and combat resistance.

Figure 1.

Fitness landscapes can assess drug resistance potential early in the drug development process. The current throughput of mutational scans is compatible with rapid and comprehensive identification of drug resistant mutations for thousands of compounds. Determination of resistance landscapes at the lead development stage of drug development would enable rationale strategies to limit the evolution of resistance.

A number of studies have demonstrated the biological relevance of resistance fitness landscapes. In bacteria, a number of groups have determined fitness landscapes of β-lactamase with inhibitors¹ that have recapitulated clinically relevant mutations as well as provided an understanding of the mechanism underlying resistance. In bacteria, mutational scanning has also been utilized to identify resistant mutations in deoxyxylulose phosphate reductoisomerase to the inhibitor fosmidomycin², and in RNA polymerase to the inhibitor rifamycin³. These studies demonstrate the feasibility of using fitness landscapes to comprehensively identify drug resistant mutations.

Systematic mutational scanning approaches have also identified clinically relevant drug resistant mutations in oncogenes and viruses. Prior to clinical observation, a fitness landscape of the BRAF^V600E oncogene that causes melanoma identified L505H as a resistant-causing mutation. While most clinical resistance to BRAF drugs have been caused by mutations in other genes (e.g. MEK), L505H was subsequently identified as causing resistance in some patients⁴. In BCR-ABL1-driven leukemia, clinically relevant resistance mutations have been well studied and are predominantly caused by mutations in BCR-ABL1. A fitness landscape in mammalian cells of a region of BCR-ABL1 near the active site revealed a strong correlation between the clinical frequency of resistant mutations and experimental measurements⁵. A nearly complete fitness map of ERK2 mutations⁶ identified 15-fold more resistant mutations than a prior study using random mutagenesis. Similarly, in multiple studies of viruses, systematic mutational scans have both recovered known resistance mutations and identified previously unappreciated resistant variants^{7, 8}. Together, these findings demonstrate the ability of fitness landscapes to predict clinical resistant outcomes and serve as comprehensive guides in developing inhibitors with reduced susceptibility to resistance.

Development of approaches to measure an experimental fitness landscape

The tremendous efficiency of next-generation sequencing powers most approaches to determine fitness landscapes (Figure 2). Next-generation sequencing yields hundreds of millions of individual reads in a lane, enabling tracking of the frequency of a hundred thousand mutational variants in a single experiment⁹. This approach can be used to follow randomly introduced mutations¹⁰ or when combined with systematic mutational approaches to illuminate comprehensive fitness landscapes¹¹. These sequencing-driven approaches have broad applications in the investigation of sequence-function-fitness relationships including drug resistance¹². The general approach, termed mutational scanning, is sufficiently robust and effective that it has been applied by a large number of laboratories to a wide array of genes and selection schemes. Here, we provide a guide to the use of systematic fitness landscapes for the analysis of drug resistance.

Figure 2.

Mutational scanning approach to measure fitness landscapes. First, a DNA library is generated that includes all possible amino acid changes in a target gene. Next, the library is introduced into a biological form, usually cells. Each cell will contain a single mutant copy. In the next step, a selection is performed on the cells to enrich or deplete variants based on their function. For example, a drug may be applied that will hinder the growth of cells harboring a mutant with normal or defective function, but permit the growth of cells harboring a resistant mutant. Finally, next-generation sequencing is used to estimate the frequency of each mutant before and after selection. Enrichment or depletion of a mutant in the sequencing reactions is used to calculate the function of the variant under the selection conditions.

Approaches to determine fitness landscapes have undergone many key developments since the initial report from about 10 years ago¹¹. For many applications, systematic and extensive coverage of all possible mutations is critical. For example, in the analysis of drug resistance, any mutation that is unassessed causes a blind spot that hinders clinical predictions and development of inhibitors with reduced likelihood of resistance evolution. The development of approaches to systematically generate and analyze all possible amino acid changes was critical for the generation of comprehensive fitness landscapes^{11, 13}. Another key technical advancement has been the use of nucleic acid barcodes to track mutant variants^{5, 14, 15}. The pre-association of barcodes with mutations and subsequent tracking of barcodes reduces the impacts of sequencing errors on fitness estimates and enables the analyses of mutations spread across large regions of a gene that cannot be directly sequenced by the limited read length of next-generation sequencing platforms. Numerous additional approaches to generate and analyze systematic mutations have been developed and these are covered later in this work. These approaches provide a broad technical foundation for quantifying fitness landscapes and investigating the underpinnings of function and resistance.

A wide array of selections have been developed for mutational scanning readouts that range from in vitro affinity to growth impacts in microbes, viruses, and mammalian cells. Many selections for in vitro binding involve the use of either yeast^{16, 17} or phage display¹⁸ where binding is quantified based on the enrichment of variants in bound populations. In these binding studies, fluorescently labeled ligands are frequently used so that flow accelerated cell sorting (FACS) can be applied to isolate bound and unbound fractions. In a conceptually related approach, the separation of active and inactive enzyme variants using microfluidic sorting has provided a fitness landscape of a glycosidase with a fluorogenic substrate¹⁹. GFP fusions can be used to analyze the steady state expression level of mutant variants that can serve as a proxy for function for some genes²⁰. While in vitro assays can provide fitness landscapes, in our experience assays based on growth rate have generally proven faster to develop and more reproducible. Growth assays for mutational scans have been developed in bacteria^{1, 21}, yeast^{15, 22}, viruses²³, and mammalian cells^{5, 6, 24, 25}. Of note, the relationship between protein function and fitness are typically non-linear, making it challenging to relate the two types of measurements^{17, 26}. While both in vitro and growth assays can provide information on drug resistance mutations, growth provides direct experimental fitness measurements that should be of greater biological relevance.

Steps involved in using a fitness landscape to study drug resistance

Developing a selection to distinguish the function of mutant variants

More than any other step, the development of a robust selection step is critical to the success of fitness landscape studies. There are a large number of potential options that can have tremendous impacts on the reproducibility of results, the clarity of interpretations, and biological relevance. Ideally, planning the selection system should be done with a deep understanding of the gene and organism of focus so that the studies are both feasible and relevant. We suggest the following as a rough guideline in considering different selection options.

Because of the influence of the selection scheme on conceptual conclusions, we recommend pilot experiments testing multiple options. For example, there may be three potential cell lines that could be used to study drug resistance in an oncogene. While you may prefer one cell line because of its genetic relationship to the native cancer, it may prove challenging to work with experimentally (e.g., it may be challenging to efficiently introduce mutant variants into the genome). Pilot experiments with mutant variants including positive controls (e.g., those with known resistance) and negative controls (e.g., internal stop codons that destroy function) can provide information regarding feasibility including the efficiency of genetic manipulation as well as biological relevance. The efficiency of genetic manipulation influences the complexity of variants that can be analyzed. As a general guideline, we aim for 10-fold or greater coverage of the number of variants in a library at each genetic manipulation step. This limits bottlenecks that can skew the frequency of or eliminate variants from the analyses. We also suggest analyzing a small library focusing on a critical region of the gene as a valuable approach to assess potential selection schemes. These pilot studies can also be used to explore different strengths and lengths of selection to identify experimental parameters that provide the best signal to noise. Performing a biological repeat with the small library will reveal the accuracy of measurements in each setup. In all selection experiments, it is important to have an unselected control in order to quantify mutant frequency in the absence of selection. We often analyze variant frequencies in the plasmid library and just after introduction into the selection system (e.g., integration into the chromosome) as separate assessments of mutant frequencies without selection.

Prior studies in different systems can also provide a guideline for planning selection schemes. In studies of resistance in oncogenes, we suggest performing selection for growth in an immortalized cell line with variants integrated into a consistent location in the genome and tracked using a barcoding strategy because this provides results that are highly reproducible and predictive of clinical prevalence⁵. Similarly, in studying bacterial resistance, we recommend performing selections that mimic the clinical setting including genetic location (plasmid or genome depending on the gene) and similarity to the pathogenic strain or species if the actual pathogen is not feasible. Viruses provide some unique challenges because their compact genomes may not allow for neutral barcodes, though this has not been extensively examined. The tracking of mutant frequencies in viruses has typically involved directly sequencing of mutations, which can require specialized strategies to reduce the noise from misreads⁸. If growth assays are not feasible for viruses (e.g. due to biosafety concerns), in vitro selection schemes can provide reasonable alternatives. When using in vitro selections, we suggest examining the relationship between protein function and growth rate in order to predict growth effects from in vitro findings.

Generating systematic libraries of mutant variants

A wide array of approaches have been used to generate systematic libraries of mutants. These include approaches based on Kunkel mutagenesis²⁷, cassette ligation^{11, 13}, various PCR strategies²⁸, and nicking mutagenesis²⁹. While we reference these here for completeness, in our opinion commercial options for gene synthesis of site saturation libraries provide a reliable option and are the approach that we use for most current projects. Generating systematic libraries involves extensive optimization and dedication using all approaches that we have explored. Many companies use oligonucleotide directed gene synthesis approaches to generate linear DNA encoding all 19 possible amino acid changes at a position for a cost of about $50 per position. Of note, prior fitness landscape studies have demonstrated that synonymous mutations tend to have little to no observable impact. For most fitness landscape applications, it is sufficient to measure one codon for each amino acid. When generating libraries, we recommend leaving at least 20 base pairs at either end invariant for use as handles (e.g., primer binding sites) for downstream manipulations. As we describe below, these libraries can then be easily transferred to different plasmids and barcoded.

Transferring libraries into plasmids and genomes, and barcoding

Libraries can be efficiently transferred into plasmids and barcoded for most applications (Figure 3). We transfer libraries into plasmids using either Gibson Assembly or NEBuilder HiFi DNA Assembly enzyme blends (both from New England Biolabs). Following the vendor recommended conditions, we routinely obtain greater than 100,000 independent transformants from a single reaction. The efficiency and flexibility of this process makes it simple and fast to transfer complex libraries while maintaining variant diversity. For libraries with larger than 10,000 variants, the reactions can be scaled up to achieve 10x coverage.

Figure 3.

Outline of approaches to transfer libraries into plasmids and add barcodes. (A) Regions of 20–30 identical bases enable efficient transfer of mutant versions of a gene into a plasmid using Gibson Assembly (NEB) or related enzyme blends. (B) Duplex DNA barcodes (e.g., N18) with constant regions at both ends and overhangs that match with plasmid ends can be efficiently ligated.

We introduce DNA barcodes into plasmid libraries using a ligation strategy (Figure 3). We recommend using restriction sites with 8 base recognition sites (e.g., NotI and PacI) for this step because it will reduce the likelihood that the enzyme cuts a mutant variant. For yeast applications, we typically use a random 18 base barcode (N₁₈) either upstream of the promoter or downstream of the 5’ UTR. This barcode includes 4¹⁸ or roughly 10¹¹ possible combinations. Even for large variant libraries (e.g., with 10⁶ variants), only a minute fraction of possible barcodes will be utilized. Therefore, almost all variants are barcoded with sequences that differ from all other variant barcodes by more than one base. Thus, when barcodes are sequenced, most errors in sequencing result in an unused barcode that can be detected and either corrected or ignored. Of note, the diversity of the barcode also makes it unlikely that a perfect pair will be present when samples of N₁₈ oligos are annealed. For this reason, we typically add at least 12 bases of invariant sequence at either end of each N₁₈ oligo to facilitate stable duplex ends for ligation. In addition, we have found that purifying the vector on a silica column without gel purification provides the best results. For efficient ligation reactions, we dephosphorylate the vector prior to ligation and phosphorylate each oligo individually prior to annealing them. It is also critical to dilute the annealed oligo so that the ligation reaction contains a ratio of about 1 vector molecule for every two duplex barcode molecules. Using these optimizations, we typically obtain about 50,000 independent transformants from a standard barcode ligation reaction, and this can readily be scaled up when necessary. Individual colonies from the N₁₈ ligation transformation typically have a mixture of two barcodes such that the barcode diversity is twice the number of transformants. This does not have a large impact on downstream steps, but can be confusing when first encountered. For barcoding mammalian genes, we have used an intron barcoding strategy ⁵. The main difference of the intron barcode from N₁₈ is that it avoids recognition motifs involved in mRNA splicing. As with the N₁₈ barcode, the intron barcode can also be introduced into plasmid libraries using ligation.

While many mutational libraries will contain copies of the wildtype (WT) gene because of the manner in which they are built, we have found it useful to add a small fraction of additional WT molecules. We do this by ligating barcodes into a plasmid containing the WT gene and isolating plasmid DNA from about 100 colonies scraped from a plate of bacterial transformants. We then identify the barcodes associated with the WT sample by sequencing the barcodes as a small fraction of an Illumina sequencing lane.

Multiple strategies have been used to efficiently introduce variant libraries into the genomes of mammalian cells. Of note, genome integration is critical for reproducible results in mammalian cells as introducing libraries into random locations using transduction strategies²⁴ can lead to variable translation efficiencies with dramatic impacts on phenotypes including growth and drug resistance. Genomic integration efficiencies of 8% have been achieved using either an optimized CRISPR/Cas9 strategy⁵ or a Bxb1 recombinase strategy³⁰. For studies of resistance in cancer, these strategies provide tools that can readily interrogate comprehensive fitness landscapes for most oncogenes.

Next-generation sequencing and data processing

The use of barcodes to track variant frequencies provides multiple benefits, but require an additional sequencing step to associate barcodes with mutant variants. In addition to reducing the impact of misreads on estimates of variant frequency that was discussed previously, barcoding can reduce sequencing costs when analyzing variant frequency. Because barcodes are typically 18–20 bases, they can be read using short reads that are cost effective, especially when fitness landscapes will be quantified under multiple conditions (e.g., different drug environments). Of note, with Illumina sequencing, the length of the DNA analyzed by sequencing influences the diameter of the bridge amplicons on the chip. Larger DNA generates larger diameter amplicon clusters and fewer reads per lane. It is necessary to have a sufficiently large DNA that it can form bridges and we typically use DNA that is about 270 base pairs in length for optimal Illumina sequencing. As discussed further below, reading barcodes of 18–20 bases enables highly efficient Illumina sequencing reactions for the quantification of variant frequencies.

For the association of barcodes with mutant variants (Figure 4), we recommend using circular consensus sequencing (CCS) available on PacBio instruments. By performing multiple reads of single molecules of circular DNA, this approach provides both long read lengths (>10 kb) and outstanding accuracy (>99% for each base call). Of note, we have found that PCR of large (>1 kb) fragments of DNA containing point mutant variants can lead to up to 25% of molecules recombining. For this reason, we typically generate samples for PacBio sequencing by cutting with restriction enzymes close to the ends of the barcode and location of mutations, blunting the ends, and sending the sample for purification on a BluePippin instrument (Sage Biosciences), followed by circularization, and PacBio sequencing. It is important not to expose the DNA to UV light or ethidium bromide during the purification steps as these will damage the DNA and impair PacBio sequencing. The resulting CCS files of libraries with up to 100,000 barcode variants typically contain multiple independent reads of each barcode with identical mutations associated for almost all (>99%) barcode containing reads.

Figure 4.

Outline of sequencing to associate barcodes with mutations and assess the frequency of barcodes. (A) Barcodes can be associated with mutations using PacBio circular consensus sequencing that provides long reads and high accuracy. Linear DNA containing the open reading frame and barcode need to be prepared. The addition of proprietary hairpin adapters and sequencing can be performed by commercial sources. (B) The frequency of each barcode in multiple samples can be assessed in the same lane of Illumina sequencing by using different index sequences for each sample. Index sequences are encoded on primers used to amplify the barcodes. All index sequences should differ from one another by two or more bases to reduce the likelihood of misreads leading to mistakes in assigning a read to the correct sample. The primer used to read the index in Illumina sequencing is called i7.

We process CCS data using custom analysis scripts that are available upon request. Our analysis pipeline includes organizing the data by barcode, and then for each barcode detailing the mutations associated with each read. For barcodes with two or more reads, we typically find complete agreement between the mutations in each read. In these cases, we accept the barcode and add it to a file listing the barcode and the associated mutations. This file is then used as a lookup table when we quantify barcode frequencies in samples from selection experiments.

In our experience, Illumina sequencing currently provides the most effective tool for analyzing barcode frequencies in samples from selection experiments (Figure 4). For accurate estimates of barcode frequency, we aim for an average of more than 100 reads of the average barcode. For an experiment with 100,000 barcodes, we would aim for 10⁷ reads for each condition (e.g., drug condition). The throughput of Illumina sequencing (~4×10⁸ reads per lane) enables us to analyze multiple conditions (up to 40 in this example) in a single lane. To distinguish samples in the sequencing sample, we add index sequences using a PCR strategy (Figure 4). We read the index sequence using the illumine index sequencing primer. Of note, because Illumina sequencing reads one base at a time from clonal clusters, it can be flooded if the same base is present at the same position in all clusters (especially when reading the initial bases used in defining cluster positions). To avoid sequencing failures, we blend samples from different genes in order to generate sequence diversity at all positions and/or blend in DNA with a random sequence (e.g. sheered phi-X DNA) at 25%.

We analyze Illumina barcode reads using custom scripts that are available upon request. Our analysis pipeline includes steps to organize the data by barcode and index sequence, tabulate reads of each barcode with each identifier, lookup the associated mutations with each barcode, and output the sequencing observations of each mutant variant under each condition. Fitness effects are estimated by comparing the frequency of each variant in an unselected control compared to its frequency under selection. Fitness effects are often normalized to WT and stops in order to facilitate comparison between different experiments and conditions.

Comparison of selection on the change in mutant variant frequency to WT can be used to identify mutations with functional impacts. The individual barcodes of WT provide an internal estimate of measurement variation. Multiple statistical approaches can be applied to assess variants that are significantly distinct from the WT distribution. One of the simplest approaches is to use twice the standard deviation to define a 95% confidence interval for WT fitness effects. Corrections can be made for multiple tests (e.g., Bonferroni corrections) and where experimental replicates are available variations in both WT and each mutant variant can be included in the statistical analyses.

Assessing evolutionary potential

Evolutionary potential is determined by both selection and mutation. While fitness landscapes described in this work are determined by selection only, evolutionary potential can be estimated by including a model of mutation. For example, in BCR-ABL1, we observed that drug resistant amino acid changes requiring more than one nucleotide change were rarely if ever observed in patients⁵. This is consistent with the known prevalence of single-base mutations in mammalian cancer cells. However, incorporating a simple mutational model provided predictions of evolutionary potential that closely matched clinical observations⁵. The availability of genome sequencing of parents and offspring for many organisms provides accurate mutational models that can be readily utilized to assess evolutionary potential from fitness landscapes.

Interpreting functional effects in the light of protein structure provides opportunities to understand mechanism and the biophysical principles that underly evolutionary potential. For example, structural modeling can indicate if drug resistant mutations act by directly interfering with inhibitor binding, or through longer-range allosteric or conformational changes. Also, identifying the set of mutations that are compatible with function in the absence of inhibitors provides the potential to structurally define target sites for drugs that can only be disrupted by mutations that also disturb function. Structural analyses of fitness landscapes can identify additional pockets beyond the active site that are critical for function (e.g., allosteric sites) that may be targeted by inhibitors.

Summary and conclusions

Drug resistance continues to increase in prevalence and new approaches to understand and combat it are critical for human health. The technology for measuring fitness landscapes can be widely applied to drug resistance and has the potential to aid in the development of next generation therapies with reduced potential for resistance. Given the state of resistance and fitness landscape technology, this is a critical time for collaborative efforts between scientists in academics, biotech, and pharmaceuticals.