Abstract
Non-genetic fluctuations in the molecular state of a cell can lead to different cell fate choices in response to environmental perturbations. In the context of cancer, such fluctuations can lead to the formation of rare cells within the population that are able to survive treatment with targeted therapies. However, the biological processes governing the formation of these cells and their subsequent impact on therapy resistance remains largely unknown. We performed genetic screens using CRISPR/Cas9 to identify over 80 factors that modulate these rare cell fluctuations. These screens revealed several novel targets affecting cellular variability and, in turn, therapy resistance. Transcriptomic analysis revealed that many of these novel targets appeared to act via distinct mechanisms from those previously identified; for instance, knockout of either DOT1L or LATS2 appeared to increase resistance by increasing the degree of cellular differentiation, as opposed to knockout of MITF or SOX10, which achieve the same effect by decreasing cellular differentiation. We show that the timing of inhibition of these new variability-altering factors can affect the degree of resistance to targeted therapies. Our results suggest that cellular plasticity may be subject to regulatory processes that play critical roles in cell fate determination.
Introduction
The advent of therapies designed to inhibit particular oncogenic targets has raised the potential to cure certain cancers. However, while these therapies are able to kill most of the tumor cells, often a few cells remain that are able to ultimately repopulate the tumor (Garraway and Jänne 2012; Trunzer et al. 2013). There are numerous mechanisms by which these rare cells become resistant. While these mechanisms are often genetic in origin, for instance a mutation affecting the site of action of the drug, non-genetic variability between otherwise identical cells has emerged as another mechanism by which cells can be rendered resistant (Sharma et al. 2010; Gupta et al. 2011; Pisco and Huang 2015; Shaffer et al. 2017, 2018; Cohen et al. 2008; Weinreb et al. 2018; Rambow et al. 2018; Su et al. 2017). A commonality in the study of genetic versus non-genetic resistance mechanisms is the goal of determining the molecular “state” of a cell that confers distinct ultimate fates such as death or therapy resistance—a mutation in the case of genetic mechanisms versus high levels of expression of particular marker genes in the case of non-genetic mechanisms. However, a critical difference is that while genetic mutations largely arise through spontaneous, stochastic processes, non-genetic fluctuations can in principle occur due to the changes in activity of specific biological pathways, yielding the potential to enhance or inhibit the formation of these rare cells by targeting those pathways specifically. Yet, to date, little is known about pathways that affect non-genetic rare-cell variability, leaving their therapeutic potential unrealized.
As an example, in melanoma, several groups have identified transient rare-cell populations within otherwise homogeneous cells that are able to proliferate in the face of treatment with BRAFV600E inhibitors (Fallahi-Sichani et al. 2017; Tirosh et al. 2016; Rambow et al. 2018; Shaffer et al. 2017; Torre et al. 2018). These cells, which we refer to as pre-resistant cells, are often marked by transiently high expression of several resistance marker genes, such as EGFR, NGFR and AXL. When drug is added to enriched populations of these cells, they are far more likely to be drug resistant than other cells in the population (Shaffer et al. 2017) (Fig. 1A, top). Furthermore, once these cells are exposed to drug, they are reprogrammed such that the transient pre-resistant phenotype is converted to a stably drug-resistant phenotype characterized by massive changes in signaling and gene expression profiles.
These results serve as an example of linking the molecular state of the cell (before the addition of the drug) to the eventual fate of the cell (upon adding drug) (Symmons and Raj 2016; Raj and van Oudenaarden 2008; Raj et al. 2010; Wernet et al. 2006; Süel et al. 2007; Maamar, Raj, and Dubnau 2007; Mojtahedi et al. 2016; Rambow et al. 2018). This mapping from cellular state to phenotypic fate has two key components. One is the fluctuations that give rise to distinct cellular states in the original population. The other is conversion of that molecular difference into a distinct phenotypic fate; in this case, the biological processes that convert a pre-resistant cell into a stably resistant colony. We were interested in dissecting the molecular regulators of this mapping.
With the advent of CRISPR/Cas9 technology, it is now possible to perform genetic screens to identify regulators of various molecular processes. For therapy resistance, virtually all screens have been designed to detect changes to cellular fate only—i.e., changes in the final number of resistant cells, typically measured as a proliferation phenotype (Shalem et al. 2014; Konermann et al. 2014; Strub et al. 2018; Joung et al. 2017). However, the non-genetic paradigm of drug resistance consists of multiple processes, specifically first the aforementioned transient fluctuations in single cells that results in the generation of many cellular states within an otherwise homogeneous population, and then the subsequent drug-induced selection and reprogramming (Shaffer et al. 2017). Each of these processes may in principle have distinct regulatory mechanisms, and so the opportunity remains to direct screening techniques at these specific aspects of non-genetic drug resistance; for instance to identify factors that affect the initial fluctuations in the cellular state (Fig. 1A, middle). These factors may then also affect the overall degree of drug resistance, but potentially through new, previously undiscovered mechanisms that allow for new therapeutic targets that affect drug resistance in ways not revealed by classical resistance screens.
We here describe the results of genetic screens designed to capture modulators of single cell state variability that subsequently affect cell fate decisions. Specifically, in the context of melanoma, we performed pooled CRISPR/Cas9 genetic screens to reveal modulators of the rare cell state we have previously identified as being responsible for drug resistance. We recover several factors known to play key roles in melanoma, and also identify several new factors. Using RNA sequencing to measure the transcriptome profiles induced by knocking out these factors, these factors organized into a few distinct functional classes, suggesting that these sets of factors may affect cell state variability and consequently cell fate selection through a small number of distinct modes. Drugs targeting these new orthogonal mechanisms display a variety of synergistic effects when coupled with therapy, sometimes depending on the relative timing of drug application.
Results
CRISPR/Cas9 genetic screens identify factors that modulate cellular states
We wanted to identify factors that affected the fluctuations in cellular state that lead to single cells being resistant to drug. We took advantage of a clonal melanoma cell line (WM989 A6-G3) that we have extensively characterized as exhibiting resistance behavior in cell culture that is broadly comparable to that displayed in patients. Phenomenologically, in cell culture, we observe that upon addition of a roughly cytostatic dose of the BRAFV600E inhibitor vemurafenib (1µM), the vast majority of cells die or stop growing, but around 1 in 2,000-3,000 cells continues to proliferate, ultimately forming a resistant colony after 2-3 weeks in culture in vemurafenib. We have previously demonstrated that before the application of drug, there is a rare subpopulation of cells (pre-resistant cells) that express high levels of a number of markers, and that these cells are far more likely to be resistant than other cells (Shaffer et al. 2017). In order to identify modulators of the fluctuations that lead to the formation of this subpopulation of pre-resistant cells, we designed a large scale loss-of-function pooled CRISPR genetic screen (which we dubbed the “state screen”) comprised of ~13,000 single guide RNAs (sgRNAs) targeting functionally relevant domains of ~2,000 proteins, with roughly six distinct single guide RNAs per domain (1402 transcription factors targets, 481 kinase targets, 176 epigenetic targets; each single guide RNA targets an important functional domain, see Supplemental tables 1-3) (Huang et al. 2018; Tarumoto et al. 2018; Brien et al. 2018). To conduct the screen, we stably integrated Streptococcus pyogenes Cas9 (spCas9) into the WM989-A6-G3 cell line, creating the line WM989-A6-G3-Cas9, and isolated a single cell clone, which we refer to as WM989-A6-G3-Cas9-5a3. We verified that this cell line was capable of editing the genome and that it still contained pre-resistant cells marked by the expression of drug-resistance markers (Supplemental fig. 1).
Our screening strategy consisted of first transducing the pooled library of single guide RNAs into a population of cells using lentivirus at a low multiplicity of infection. Given that the pre-resistant cells show up only rarely within the population, we needed to ensure that we had sufficient cell numbers in order to effectively sample differences in the frequency of pre-resistant cells. Thus we expanded the culture to maintain around 50,000-250,000 total cells per each individual single guide RNA, for a total of roughly a billion cells per screen. We then used a combination of magnetic sorting and flow cytometry to isolate cells that were positive for both EGFR and NGFR expression, both of which are prominent markers of the pre-resistant rare cell subpopulation. We then sequenced the single guide RNAs in this sorted subpopulation to determine which single guide RNAs were over- or under-represented as compared to the unsorted total population. Here, over-representation suggests that knockout of the gene leads to an increased frequency of NGFRHIGH/EGFRHIGH cells and vice versa (Fig. 1B). To select “hits” from the screen, we designed a series of criteria to identify and rank targets (see methods for a detailed description of the selection criteria). For ranking targets, we considered a target in the screen to be a Tier 1 “hit” if we detected a two-fold change in abundance for ≥ 75% of the single guide RNAs (Tier 2 ≥ 66%, Tier 3 ≥ 50%, Tier 4 < 50%).
By these criteria, we obtained a set of 61 targets identified as factors (Tier 1 or Tier 2) affecting the frequency of NGFRHIGH/EGFRHIGH cells in our screen (Fig. 2, Supplemental table 4). Of these, 25 increased the frequency of NGFRHIGH/EGFRHIGH cells, while 36 decreased the frequency. Reassuringly, many genes known to be important factors in melanoma biology emerged from our screen, including SOX10 and MITF, both of which are key melanocyte transcription factors whose downregulation is often associated with increased drug resistance (Sun et al. 2014; Almeida et al. 2019; Hartman and Czyz 2015; Wellbrock and Arozarena 2015). Other tumor suppressors include RUNX3, whose expression is often lost in melanoma (Kitago et al. 2009; Bae and Choi 2004) and LATS2, a regulator of the Hippo pathway (Harvey, Zhang, and Thomas 2013).
Additionally, several new factors not previously associated with resistance to BRAFV600E inhibition emerged from our screen. These include genes such as DOT1L, which encodes an H3K79 methyltransferase associated with melanoma oncogenesis (Zhu et al. 2018), and BRD2, which encodes a protein that is a member of the BET family, often overexpressed in human melanoma (Segura et al. 2013).
In order to assess the robustness and generality of our results, we performed secondary, targeted screens in WM989-A6-G3-Cas9 (non-clonal) cells as well as in another BRAFV600E human melanoma cell line, 451Lu-Cas9. This screen included sgRNAs targeting 34 of the Tier 1 and 2 factors we identified in the state screen, as well as another 52 factors from Tiers 3 and 4 (these lower confidence hits may also have been Tier 1 or 2 hits in the “fate” screen described below; Supplemental table 5). We found that 25 of the 34 high confidence hits showed at least a two fold change in the frequency of NGFRHIGH/EGFRHIGH cells concordant with the effects detected in the original screening clonal cell line (WM989-A6-G3-Cas9-5a3). In 451Lu-Cas9 cells, 20 of the 34 targets also showed a change in the frequency of NGFRHIGH/EGFRHIGH cells, with 11 of those exhibiting at least a two-fold change, indicating that the factors identified by our screen are likely general and that our results are not entirely dependent on the specific cellular context of the original WM989-A6-G3-Cas9-5a3 cell line (Supplemental Fig. 3, Supplemental table 4). Not all the factors were validated by this secondary screen, which may be due to technical challenges associated with performing screens for rare cell phenotypes (see discussion). Also, even Tier 4 hits displayed qualitative agreement in these secondary screens, although quantitatively they were not as strong as the effects of the Tier 1 and 2 hits, suggesting that our screen has isolated many strong effect hits but that there may be several additional factors with smaller effect sizes that we were unable to detect.
A cell survival “fate” screen reveals a distinct set of factors that modulate drug resistance
Our variability “state” screen is conceptually distinct from more conventional resistance screen designs for which the screened phenotype is overall survival and proliferation in the face of drug. Our hypothesis was that those more conventional screens would identify a different set of factors because the underlying biological processes may be different. To test this hypothesis, we also performed a parallel, conventional genetic screen for resistance to find modulators that specifically altered cellular “fate”; i.e., formation of resistant colonies when exposed to BRAFV600E inhibition. Here, we used the same cells and the same library of single guide RNAs, but instead of isolating NGFRHIGH/EGFRHIGH cells, we added vemurafenib and grew the cells until large resistant colonies formed, at which point we isolated DNA and sequenced it to look for over- and under-represented single guide RNAs as before (Fig. 3A). Here, if a single guide RNA is overrepresented in the resistant cell population, it means that knocking out the gene it targets resulted in an increase in the number of cells that survive the drug and acquire resistance, and vice versa for under-representation of single guide RNAs.
As before, we ranked the targets into tiers based on the percent of single guide RNAs that exhibited at least a two-fold change in representation throughout the screen (Tier 1, ≥ 75%; Tier 2, ≥ 66%; Tier 3, ≥ 50%; Tier 4, < 50%). We identified 24 factors (Tier 1 and 2) that emerged as hits from our fate screen (Fig. 3B, Supplemental table 4). Among those, there were a number of factors that affect or act downstream of signaling pathways such as MAPK (CSK) (Okada et al. 1991), Wnt/B-catenin (KDM2A)(Lu et al. 2015), and Hippo (LATS2) (Harvey, Zhang, and Thomas 2013). Of these, 20 of the 24 emerged as predicted enhancers of the number of resistant cells, while the remaining 4 were predicted to decrease the number of resistant cells. Of all 24 hits, five were also identified by our state screen (Fig. 3C).
The lack of correspondence between the two screens most likely reflects the fact that distinct biological processes may play a more dominant role in either single cell variability or other aspects of the overall acquisition of resistance. If it were possible to run the screen for overall acquired resistance to saturation—i.e., isolate all possible factors affecting resistance—then we would in principle be able to find all variability factors that affected the resistance phenotype as well. However, bottlenecks in the screening process owing to the rarity of the resistance phenotype at the single cell level meant that some factors (e.g. CSK) may come to dominate the fate screen, making it difficult to fully capture all factors through just the conventional resistance screen alone.
As with the state screen, we also performed a secondary, targeted fate screen in both WM989-A6-G3-Cas9 and 451Lu-Cas9 cells to test for the generality of the factors we had identified. This time we included nine of the top factors affecting vemurafenib resistance in our screen (as well as 77 targets that either affected vemurafenib resistance but did not pass the thresholds to be called a hit, or affected the frequency of NGFRHIGH/EGFRHIGH cells, see Supplemental table 5). In WM989-A6-G3-Cas9, we found that seven of the nine targets replicated the effect that we observed originally. For 451Lu-Cas9, the same seven factors showed similar effects, suggesting again that the factors isolated may not be specific to the cellular context (Supplemental fig. 4).
Factors affecting variability also affect overall drug resistance to different degrees
We designed our state screen to identify factors that increase or decrease the percentage of NGFRHIGH/EGFRHIGHcells. The implicit assumption was that increasing the percentage of NGFRHIGH/EGFRHIGH cells would be associated with an increase in the number of cells that went on to acquire drug resistance. However, EGFR and NGFR are, in this context, just markers of the pre-resistant state, and most of the factors identified in the state screen did not appear in our fate screen for factors affecting overall resistance. Thus, it was conceivable that factors identified from our screen might simply affect the transcriptional regulation of these genes but not affect the frequency of cells being in the pre-resistant state per se. Thus, we conducted extensive validation experiments to demonstrate that: 1. the factors identified actually did change the frequency of cells expressing NGFR as predicted by the screen, and then 2. that these changes in the frequency of NGFRHIGH cells translated into changes in the number of resistant colonies upon application of vemurafenib.
To measure the frequency of NGFRHIGH cells, we generated a population of ≥ 500 cells in which the target of interest had been edited and then performed immunofluorescence using antibodies that bound to NGFR (86 different targets, each targeted at least three times using different sgRNA designs for a total of ~288 different samples including controls). Most of the targets selected from the state screen showed an increase or decrease in the number of NGFRHIGHcells concordant with the effect on the primary screen, with 21 of 34 tier one and tier two targets showing at least a 50% increase or decrease over control (Fig. 4A-B). NGFR immunofluorescence performed on cells in which targets from tiers three and four were knocked out also seemed to confirm the effect measured in the primary screen, but to a lesser extent, with only 21 out of 49 targets increasing or decreasing the frequency of NGFRHIGH cells by 50% or more (Fig. 4B, tiers 3 and 4; Supplemental fig. 5). This confirmed that our screen was able to identify factors that modulated the expression of the pre-resistance marker NGFR, and further suggests that there are likely additional factors with moderate-to-small effect size that our screen was unable to detect reliably. (Note that we also organized the targets tier in the fate screen and show their effect on the frequency of NGFRHIGH cells in the bottom panel of Fig. 4B.)
We next took a subset of these knockout populations (33 different targets, 21 of which are tier 1 or tier 2 hits from the primary state screen), added vemurafenib, and counted the number of resistant colonies after three weeks in drug. Despite not being classified as a tier 1 or 2 hit in the primary fate screen, we found that 15 of the 21 factors tested showed an increase or decrease in the number of resistant colonies, consistent with the prediction that an increased or decreased percentage of NGFRHIGH cells would lead to a concomitant change in resistance (Fig. 4A, C, Supplemental fig. 6). We also tested seven tier 1 and tier 2 targets from the primary fate screen, six of which led to a change in the number of colonies resistant to vemurafenib (Fig. 4C, right panel; Supplemental fig. 6-7). Additionally, for DOT1L, where small molecule inhibitors exist, we confirmed that at non-cytotoxic doses of the drug, inhibition of this target led to dramatic changes in the number of colonies able to survive and proliferate upon BRAFV600E inhibition (Supplemental fig. 8). Furthermore, DOT1L inhibition led to an increase in the number of resistant colonies not only upon targeting BRAFV600E, but also upon MEK inhibition and simultaneous MEK and BRAFV600E inhibition (Supplemental fig. 9).
If the level of NGFR expression perfectly reflected the probability of a cell becoming resistant upon addition of drug, then changes in the number of resistant colonies should be directly proportional to changes in the number of NGFRHIGH cells. However, while the general trend indicated such a pattern, knockouts of individual genes varied widely in the degree to which this relationship held (Fig. 4D). For instance, knockout of EP300 resulted in a ~two-fold increase in the number of NGFRHIGH cells but only a small increase in the number of resistant colonies, while knockout of CSK resulted in only a small increase in the number of NGFRHIGH but had at least a six-fold increase in the number of resistant colonies. (Importantly, the number of colonies for the CSK knockout is an underestimate due to difficulties in accurately counting colonies in highly-confluent plates, see Supplemental fig. 10 for raw image of colonies.) This is probably why CSK was a dominant hit in our screen for full resistance for vemurafenib. Conceptually, a change in the number of resistant colonies without a proportional change in the number of NGFRHIGH cells could result from a change in the mapping between state and fate induced by the drug (perhaps best thought of as a change in “threshold”), or from a shift in the distribution of the internal state of the cells that is not reflected in a change in NGFR expression. While our results cannot conclusive resolve this difference, some of our results argue in favor of changes to the mapping between state and fate itself. For instance, the CSK knockout cell line showed an increase in the number of resistant colonies but also an increase in the number of resistant cells that do not form colonies (Supplemental fig. 10). This suggested that, in addition to the usual pre-resistant cells that form colonies, an additional set of cells in the CSK knockout line were now enabled to survive drug. This suggests that the “threshold” for cells to survive drug may have changed; i.e., the state-fate mapping has been altered by the removal of CSK.
Transcriptomics reveals multiple mechanisms for cellular state and fate modulation
Our screens revealed a large number of factors affecting pre-resistance and full resistance (state and fate) that act across a range of biological processes, including a variety of signaling pathways and transcriptional regulatory mechanisms. Interestingly, a priori, no particular pathways appeared to dominate the set of identified factors; however it was possible that seemingly unrelated genes nevertheless affect pre-resistance and resistance through common biological processes.
To look for such pathways, we used RNA sequencing to measure genome-wide transcript abundance levels for 266 knockout cell lines targeting 85 different proteins (each targeted with 2-3 separate single guide RNAs; see supplementary table 4). These 85 proteins included 34 targets selected from the state screen, nine selected from the fate screen (five of these nine targets overlapped with 34 hits from the state screen), with the remaining targets being tier 3 and tier 4 targets in the primary screens. (Also, of the 85 protein targets, 45 validated by testing for changes in the frequency of NGFRHIGH cells by immunofluorescence and/or changes in the number of resistant colonies upon adding vemurafenib; Fig. 4, Supplemental figs. 5 and 7.)
Initially, we clustered the transcriptome profiles from the different cell lines, including only genes differentially expressed in at least one sample (Supplemental fig. 11). We found that while the transcriptomes induced by some gene knockouts were clearly distinct (such as MITF, SOX10 and KDM1A), many others appeared to coalesce and show relatively small differences, despite the fact that our validation results showed that these knockouts exhibited qualitatively different effects on resistance. We thus reasoned that while the sets of genes whose expression changes in our knockouts may be non-overlapping, these genes could still belong to similar categories of biological processes; i.e., different knockouts may all affect different genes all within a common pathway, for instance differentiation. Thus, using the transcriptome of each knockout, we performed a gene set enrichment analysis (GSEA, see methods) and obtained an enrichment score for a number of biological processes from the Gene Ontology terms database (Fig. 5)(Subramanian et al. 2005). Using these enrichment scores, the knockout lines clustered in a more obvious pattern. Notable clusters include MITF and SOX10, as before, which formed a distinct cluster from e.g. DOT1L, LATS2, RUNX3 and GATA4, which form a cluster with a rather opposing transcriptome profile. This latter cluster primarily consists of genes whose knockout results in an increase in vemurafenib resistant colonies. Interestingly, knockout or downregulation of MITF and SOX10 are also well known in the literature to increase drug resistance (Sun et al. 2014; Rambow et al. 2018), although by an apparently distinct mechanism in this context. (Note that the role of MITF in therapy resistance in general is complex (Bai, Fisher, and Flaherty 2019).
The clusters we identified also seem to correspond to distinct phenotypic profiles, meaning the resultant changes in the frequency of NGFRHIGH cells and number of resistant colonies. For instance, the transcriptomes of the knockouts in cluster 1 seem to mimic many aspects of the transcriptomes of NGFRHIGH, EGFRHIGH, NGFRHIGH/EGFRHIGH, and even vemurafenib resistant melanoma cells (Fig. 5A). Knockout of these targets showed a strong correspondence between the frequency of NGFRHIGH cells and the number of colonies that developed under BRAF inhibition, suggesting that the increase/decrease in the frequency of pre-resistant cells was the cause of increased/decreased resistance. Often, this relationship was relatively proportional, as was the case for the knockout of LATS2, JUNB, FOSL1, and CBFB; Fig. 5B). For MITF and SOX10, however, the relationship between the frequency of NGFRHIGH cells and the number of resistant colonies was much weaker, with very large changes in the latter but not the former. Accordingly, our transcriptomic analysis suggests that these knockouts lead to changes in gene expression that are distinct from those of NGFRHIGH/EGFRHIGH cells.
Principal components analysis of the enrichment scores indeed showed that these two categories of genes whose knockout affects resistance are diametrically opposed in their profile (Fig. 6). The gene set categories associated with the transcriptome differences for these knockouts are largely related to development, differentiation, morphogenesis, and migration (see Supplemental table 6). This fits with the fact that reduction of the master differentiation factors MITF and SOX10 leads to decreased differentiation and hence drug resistance, but the other cluster of knockouts points to a potentially novel means by which to increase drug resistance by instead increasing the expression of genes involved in differentiation (via knockout of DOT1L, LATS2, RUNX3, etc.).
The transcriptome analysis also revealed different categories of knockouts that resulted in a reduction of the number of resistant colonies. Some resistance reducing knockouts (BRD8 and PRKAA1) clustered with DOT1L, while another BRD2 clustered with MITF/SOX10. It is possible that these factors work in inverse ways to reduce drug resistance by either affecting differentiation or dedifferentiation. Meanwhile, the majority of resistance reducing knockouts appeared to cluster separately into two distinct clusters, generally through changes in the expression of a distinct set of genes. For one cluster (cluster 2), the set of genes whose expression was affected included several associated with metabolism (e.g. RNA transport and biosynthesis of amino acids), suggesting that modulation of metabolic processes may be a means of reducing drug resistance (Supplemental table 6). The other cluster did not show any coherent set of biological processes affected (e.g. SRC, IRF7, PKN2, among others), rendering that particular pathway or set of pathways rather mysterious.
Relative timing of targeting variability can affect drug resistance
Our two screens, one for variability in cellular state (pre-resistance) and the other for the ultimate cellular fate (stable drug resistance), isolated different factors, and our transcriptomic profiles suggest that different factors may act by different pathways. One possibility is that the effects of inhibiting these factors in concert with the standard BRAFV600E inhibitor may vary, in particular, with respect to the relative timing of drug administration. For instance, if a factor primarily affected the mapping between initial state and ultimate fate, then it may be that pre-administration of the drug would have no effect on resistance, but co-administration with the BRAFV600E inhibitor would have a synergistic effect. We previously observed this with the EGFR inhibitor, where inhibition of EGFR before the inhibition of BRAFV600E had no effect, but concurrent inhibition greatly reduced the number of resistant colonies (Shaffer et al. 2017). In principle, the opposite may also be true: if a factor affected primarily the initial state of the cell but did not greatly affect the mapping between state and fate, then inhibition of the factor before BRAFV600E inhibition may have a strong effect on ultimate resistance, but if applied concurrently with BRAFV600E inhibition, it may have a greatly reduced effect because the BRAFV600E inhibitor already selects which cells will survive before the internal state of the cell can change.
To test for such a possibility, we used the DOT1L inhibitor pimenostat (Daigle et al. 2011; Basavapathruni et al. 2012) (which increases the number of colonies resistant to vemurafenib over a range of doses; Supplemental fig. 8) to see if timing of DOT1L inhibition would affect the formation of resistant colonies. In addition to the standard vemurafenib treatment, we both pre-treated with the DOT1L inhibitor for seven days before adding vemurafenib and co-treated with the DOT1L inhibitor concurrently with vemurafenib (we tested both pre-treatment followed by vemurafenib alone and pre-treatment followed by concurrent treatment). We found that pre-inhibition of DOT1L resulted in three-fold more colonies than with BRAFV600E inhibition alone, but that co-treatment with the DOT1L and BRAFV600E inhibitors led to no change in the number of resistant colonies (Fig. 7), suggesting that DOT1L inhibition is altering the distribution of states of the cells, and consequently the number of cells that develop resistance to BRAFV600E inhibition. Our results demonstrate that the relative timing of inhibition of different pathways can have a profound effect on resistance.
Discussion
We have here demonstrated, using high-throughput screening, that there are genetic factors that can alter cellular plasticity in cancer cells, thereby affecting their resistance to targeted therapeutics. We identified several such factors, including known players in the biology of melanoma such as MITF and SOX10, but also a variety of new factors that appear to work by distinct, indeed potentially opposing pathways from these. These factors revealed new possible vulnerabilities that a conventional genetic screen targeting resistance did not uncover, thus demonstrating the potential for screens specifically designed to target single cell variability to reveal new biological mechanisms that may subsequently emerge as therapeutic opportunities. Drug screens targeting gene expression “noise” have also shown similar therapeutic potential (Dar et al. 2014).
While we isolated several new factors that specifically affected cellular variability, it is important to note that no single factor we isolated resulted in a change in cellular variability that was stronger than all the rest; i.e., no factor emerged as the “smoking gun”. This may be the result of the fact that our screen did not target all potential regulators. Alternatively, it may be that the biology of cellular variability is intrinsically multifactorial, with the coherent activity of many factors being required for cells to ultimately enter the highly deviated cellular state responsible for phenotypes like drug resistance (Shaffer et al. 2018). Larger scale screens may help reveal a more complete picture of the origins of rare cell behavior; however, the limitations imposed by the rarity of the pre-resistant cellular phenotype make this rather difficult. The raw numbers of cells required to properly sample these rare cell behaviors in a pooled genetic screening format remains a major technical challenge for the field of rare cell biology.
Indeed, it is the very difficulty of performing these screens at full depth that provides motivation for screening for variability rather than simply screening for resistance. If one is primarily interested in factors affecting resistance, then in principle such a screen, if carried to saturation, would reveal all such factors, including those that exert such an effect via modulation of cellular variability. However, the degree of overlap in the factors identified between our variability screen and our conventional resistance screen was relatively small. This lack of overlap suggests that distinct biological processes may dominate the results of these differently designed screens. That of course in turn raises the question of why one might want to perform variability screens at all, given that the phenotype of interest is resistance. Our results on timing of variability inhibition suggest that while the mechanisms governing rare cell variability may not appear as potent as those revealed by conventional resistance screens, the fact that they represent distinct mechanisms means that they may present an opportunity to be used in tandem. It is also possible that these mechanisms may be more dominant in other, more clinically relevant contexts.
In our validation studies, for several factors, we measured the effects of knocking out those factors on both the number of NGFRHIGH cells (which serves as a proxy for cellular state) and number of resistant colonies upon adding vemurafenib (which is our measurement for cellular fate). Interestingly, different knockouts affected both of these validation metrics differently, with some (e.g. LATS2) both increasing the frequency of NGFRHIGH cells as well as concomitantly increasing the number of resistant cells, and some (e.g. CSK) dramatically increasing the frequency of resistant cells without a proportional change in the frequency of NGFRHIGH cells. One possible way to conceptualize these distinct phenotypic outcomes is that the former category of knockout affects primarily cellular variability, i.e., cellular state, while the latter affects the mapping between these states and their fates upon addition of vemurafenib. In one simple model, one could imagine a distribution of cellular states in the initial population and a threshold whereby cells above the threshold are resistant and those below the threshold are not (Supplemental fig. 12). In this model, some knockouts may alter the distribution of cells in the initial population, thus rendering a different proportion of them above or below the threshold, or may alter the threshold itself, or potentially some combination of both. It is wise to caution against this simple interpretation, however. First, we note that NGFR expression is just a marker for the pre-resistant state, and it may be that factors may affect the frequency of pre-resistant cells without showing any effect on NGFR expression, thus giving the false appearance of a change in the mapping. (Arguing against this, however, is the fact that the transcriptomes of knockouts such as DOT1L that increase the frequency of NGFR and resistance appear to be similar to the profile of NGFRHIGH cells themselves; Fig. 5) Further molecular profiling of individual cells from these knockouts may help reveal the ways in which the molecular state of these cells changes. Secondly, it is also likely that the categorization of fates as “resistant” or “dead” is dramatically oversimplified, and that there may be a number of different types of resistant cells (anecdotally, we have noticed that the resistant cells in some of our knockout lines do appear morphologically different from those formed in the unperturbed cell line). Such results suggest that there is a mapping from a continuum of initial cellular states to multiple, canalized, or even potentially continuous cellular fates. An important future direction is to characterize this mapping and its regulation.
In this study, we have focused on cellular variability in the context of drug resistance in cancer. However, we have observed similar rare-cell variability in primary melanocytes (Shaffer et al. 2017), raising the possibility that the same variability may play a role in normal biological processes as well. It is thus possible that the factors we have isolated may play a role in regulating variability in these normal biological contexts, and it remains to be seen whether such factors act primarily in melanocytes or act more generally across different cell types in various tissues.
Methods
Cell Culture
We obtained patient-derived melanoma cells (WM989 and 451Lu, female and male, respectively) from the lab of Meenhard Herlyn. For WM989 we derived a single cell subclone (A6-G3) in our lab (Shaffer et al. 2017). We grew these cells at 37°C in Tu2% media (78% MCDB, 20% Leibovitz’s L-15 media, 2% FBS, and 1.68mM CaCl2). We authenticated all cell lines via Human STR profiling. We periodically tested all cell lines for mycoplasma infections.
Plasmid Construction and single guide RNA Cloning
All the Cas9 positive melanoma cell lines in this study were derived by lentiviral transduction with a Cas9 expression vector (EFS-Cas9-P2A-Puro, Addgene: 108100). All the single guide RNAs were cloned into a lentiviral expression vector LRG2.1(Addgene: #108098), which contains an optimized single guide RNA backbone. The annealed single guide RNA oligos were T4 ligated to the BsmB1-digested LRG2.1 vector. To improve U6 promoter transcription efficiency, an additional 5’ G nucleotide was added to all single guide RNA oligo designs that did not already start with a 5’ G.
Construction of Domain-Focused single guide RNA Pooled Library
Gene lists of transcription factors (TF), kinases, and epigenetic regulators in the human genome were manually curated based on the presence of DNA binding domain(s), kinase domains, and epigenetic enzymatic/reader domains. The protein domain sequence information was retrieved from NCBI Conserved Domains Database. Approximately 6 independent single guide RNAs were designed against individual DNA binding domains (Supplementary tables 1-3).(Huang et al. 2018; Tarumoto et al. 2018; Brien et al. 2018) The design principle of single guide RNA was based on previous reports and the single guide RNAs with the predicted high off-target effect were excluded (Hsu et al., 2013). For the initial pooled CRISPR screen, all of the single guide RNAs oligos including positive and negative control single guide RNAs were synthesized in a pooled format (Twist Bioscience) and then amplified by PCR. PCR amplified products were cloned into BsmB1-digested LRG2.1 vector using Gibson Assembly kit (NEB#E2611). For the targeted pooled validation screen, individual single guide RNAs were synthesized, cloned, and verified via Sanger sequencing in a 96-well array platform (Supplementary table 5). Individual single guide RNAs were pooled together in an equal molar ratio. To verify the identity and relative representation of single guide RNAs in the pooled plasmids, a deep-sequencing analysis was performed on a MiSeq instrument (Illumina) and confirmed that 100% of the designed single guide RNAs were cloned in the LRG2.1 vector and the abundance of >95% of individual single guide RNA constructs was within 5-fold of the mean (data not shown).
Lentivirus preparation
We produced lentivirus containing single guide RNAs using HEK293T cells cultured in DMEM supplemented with 10% Fetal Bovine Serum and 1% penicillin/streptomycin. When the cells reached 90-100% confluency, we mixed the single guide RNA vectors with the packaging vector psPAX2 and envelope vector pVSV-G in a 4:3:2 ratio in OPTI-MEM (ThermoFisher Scientific: #31985070) and polyethylenimine (PEI, Polysciences: #23966). We collected viral supernatants for up to 72 hours twice daily.
Transduction of spCas9
We introduced the stable expression of spCas9 via spinfection of lentivirus along with 5ug/ml polybrene for 25 minutes at 1750 rpm. We exchanged the media ~6 hours post-transduction and selected for cells expressing spCas9 via puromycin selection (1-2μg/ml, 1 week). For WM989-A6-G3, we generated two cell lines, WM989-A6-G3-Cas9 and WM989-A6-G3-Cas9-5a3, the later being a single cell isolate of the bulk Cas9-expressing population. We generated a 451Lu-Cas9 cell line from 451Lu cells.
Transduction of lentivirus containing single guide RNAs
For transfection of melanoma cells, we infected cells with lentivirus and 5ug/ml polybrene for 25 minutes at 1750 rpm. We exchanged the media ~6 hours post-transfection. We quantified the percent of the population transfected by measuring the number of GFP-positive cells at day 5 post-transfection. For the screens, we aimed to transfect 30% of the population. For all other experiments, we aimed to transfect >95% of the population.
Primary pooled CRISPR screens
We worked with three main pooled single guide RNA libraries in WM989-A6-G3-Cas9-5a3 cells. These libraries targeted ~2,000 different kinases, transcription factors, and proteins involved in epigenetic regulation. In total, the libraries contained ~13,000 different single guide RNAs including non-targeting and cell-viability editing controls (Supplementary tables 1-3). We aimed to transfect > 1,000 cells per single guide RNA and isolated ~1,000 cells per single guide RNA about a week post-transfection and prior to any selection. These baselines allowed us to validate the efficiency of our screen by single guide RNA enrichment/depletion of non-targeting controls and of controls that affect cell viability (Supplemental fig. 2). Additionally, these baselines helped us identify single guide RNAs with lethal effects in our cells. Given that we were interested in rare cell phenotypes that exist in 1:2000 cells or less, throughout our screens we significantly expanded the population of cells to 50,000-250,000 cells per single guide RNA, often surpassing a billion cells per screen. This scale allowed us to observe the rare cell phenotypes dozens-to-hundreds of times in each of our controls (and in each of our single guide RNAs).
The state screen aimed to identify perturbations that altered the frequency of NGFRHIGH/EGFRHIGH cells. To this end, one month after we transfected and expanded the cells, we isolated the NGFRHIGH/EGFRHIGH cells via magnetic cell sorting (MACS) followed by fluorescence-activated cell sorting (FACS) (see below). We also collected an additional ~1,000 cells per single guide RNA, without any selection, for comparison. Then, we isolated DNA from the cells and built sequencing libraries (see below) to quantify the representation of each single guide RNA in the NGFRHIGH/EGFRHIGH population and compare it to the unsorted baseline.
In the fate screen we aimed to identify proteins important for the development of resistance to vemurafenib. Here, we treated the cells as above, except that instead of isolating NGFRHIGH/EGFRHIGH cells we grew cells resistant to vemurafenib (see below) by exposing the cells to vemurafenib for three weeks. As above, we isolated DNA from the resulting population of cells and built sequencing libraries to quantify the representation of each single guide RNA. The raw output of all screens was reads per single guide RNA.
To select hits in our screens, we first normalized the output of our screens to reads per million, and then calculated the fold change in single guide RNA representation between different samples. For our state screen, we focused on the fold change in single guide RNA representation between NGFRHIGH/EGFRHIGH cells and the bulk population of melanoma cells. For the fate screen, we focused on the fold change in single guide RNA representation between cells treated for three weeks with 1μM vemurafenib (a BRAFV600E inhibitor) and cells never exposed to the drug. After normalizing the change in single guide RNA representation of each single guide RNA by the median change across all single guide RNAs, we organized our hits into tiers (one through four) based on the percent of single guide RNAs against the target exhibiting at least a two-fold change in representation. We considered hits those targets where (1) ≥ 66% of its single guide RNAs showed at least a two-fold enrichment enrichment/depletion throughout the screen, and (2) no two single guide RNAs showed a significant change (two-fold change) in opposing directions (i.e. one single guide RNA is significantly enriched in the selected population while another one is significantly depleted). Note that we excluded from analysis any single guide RNA with less than 10 raw reads in all samples.
Secondary, targeted pooled CRISPR screen
To validate the replicability and generality of our hits, we designed a pool of single guide RNAs for targeted screening that targeted proteins that either emerged as hits in our primary screens or did not pass our hit-selection criteria but changed the frequency of NGFRHIGH/EGFRHIGH cells or the frequency of cells resistant to vemurafenib (Supplemental table 5). In this pool, we included ~3 single guide RNAs per protein target, and carried out the screen in WM989-A6-G3-Cas9-5a3 cells as well as in another BRAFV600E melanoma cell line, 451Lu-Cas9. As before, we conducted a state screen where we isolated NGFRHIGH/EGFRHIGH cells as well as a fate screen where we exposed cells to 1μM vemurafenib for three weeks. Here too, we first normalized the output of our screens to reads per million, and then calculated the fold change in single guide RNA representation between different samples. Unlike on our primary screens, here we normalized the change in single guide RNA representation to the median change in representation of the ten non-targeting single guide RNAs controls included in the screen.
Immunostains
For NGFR stain of fixed cells, after fixation and permeabilization, we washed the cells for 10 min with 0.1% BSA-PBS, and then stained the cells for 10 min with 1:500 anti-NGFR APC-labelled clone ME20.4 (Biolegend, 345107). After two final washes with PBS we kept the cells in PBS. For EGFR and NGFR stains of live cells, we incubated melanoma cells in suspension for 1 hour at 4C with 1:200 mouse anti-EGFR antibody, clone 225 (Millipore, MABF120) in 0.1% BSA PBS. We then washed twice with 0.1% PBS-BSA and then incubated for 30 minutes at 4C with 1:500 donkey anti-mouse IgG-Alexa Cy3 (Jackson Laboratories, 715-545-150). We washed the cells again (twice) with 0.1% BSA-PBA and incubated for 10 minutes with 1:500 anti-NGFR APC-labelled clone ME20.4 (Biolegend, 345107). We again washed the cells twice with 0.1% BSA-PBS and finally re-suspended them in 1%BSA-PBS.
Isolation of pre-resistant cells (MACS + FACS)
To enrich for NGFRHIGH/EGFRHIGH cells we first immunostained melanoma cells as detailed above. Then, we used a Manual Separator for Magnetic Cell Isolation (MACS, with LS columns and Anti-APC microbeads). In short, following the manufacturer’s instructions, we incubated cells and microbeads at 4C for 15 min, then washed and pelleted the cells via centrifugation. After resuspending the cells, we passed them through LS magnetic columns. After enriching for NGFRHIGH cells, we proceeded to select only the cells expressing both NGFR and EGFR via Fluorescent-Activated Cell Sorting (FACS, MoFlo Astrios EQ).
Growth of resistant colonies
To grow melanoma cells resistant to BRAFV600E inhibition, we exposed melanoma cells to 1μM vemurafenib (PLX4032, Selleckchem S1267) for 2-3 weeks. For the BRAFV600E and MEK co-inhibition assays, we also used dabrafenib at 500nM and 100nM (GSK2118436, Selleckchem S2807), trametinib at 5nM and 1nM (GSK1120212, Selleckchem S2673), and cobimetinib at 10nM and 1nM (GDC-0973, Selleckchem S8041).
Inhibition of DOT1L via small molecule inhibitor
For all assays involving pharmacological inhibition of DOT1L we used pinometostat at concentrations ranging from 1μM to 5μM (EPZ5676, Selleckchem S.7062).
MiSeq library construction and sequencing
In order to quantify the single guide RNA representation following selection in our screen we sequenced the single guide RNAs as per (Shi et al. 2015). In short, we isolated genomic DNA using the Quick-DNA Midiprep Plus Kit (Zymo Research: #D4075) per manufacturer specifications. We then PCR-amplified the single guide RNAs using Phusion Flash High Fidelity Master Mix Polymerase (Thermo Scientific: #F-548L) and primers that incorporate a barcode and a sequencing adaptor to the amplicon. Our amplification strategy consisted of an initial round of parallel PCRs (23-29 cycles of up to 200 parallel reactions per sample. We then pooled the PCR reactions and purified them using the NucleoSpin® Gel and PCR Clean-up kit (Macherey-Nagel: #740609.250). We continued with eight PCR cycles using Phusion Flash High Fidelity Master Mix Polymerase, followed by column purification with the QIAquick PCR Purification Kit (QIAGEN: #28106). We quantified the single guide RNA libraries with the DNA 1000 Kit (Agilent: #5067-1504) on a 2100 Bioanalyzer Instrument (Agilent: #G2939BA). We pooled the barcoded single guide RNA libraries and sequenced via 150-cycle paired-end sequencing (MiSeq Reagent Kit v3, Illumina: #MS-102-3001). We then mapped the resulting sequences to our reference single guide RNA library and proceeded to select hits.
Cell fixation and permeabilization
For our imaging assays we fixed cells for 10 min with 4% formaldehyde and permeabilized them with 70% ethanol overnight.
Colony formation assays
For each condition tested, we first plated cells in duplicate (~10-50,000 cells per well of a 6-well plate). We fixed and permeabilized one of the duplicates to use as a baseline and exposed the second duplicate to the test condition. At the endpoint, we fixed and permeabilized the second duplicate.
Image analysis of NGFR immunostains
We developed a custom MATLAB pipeline for counting cells and quantifying immunofluorescence signal of DAPI-stained and NGFR-stained cells (https://bitbucket.org/arjunrajlaboratory/rajlabscreentools/src/default/). The software stitches together a large tiled image, then uses DAPI to identify cells. Using the nuclear area, it then looks at a set of pixels near the nucleus to quantify fluorescence intensity of the NGFR staining. After quantifying the expression level of NGFR following knockout of select screen targets and of non-targeting controls, we quantified the minimum expression level needed to be considered an NGFRHIGH cell. First, we selected the top one percent highest expressors of NGFR in each of our non-targeting negative controls. Then, within that top one percent we obtained the median expression level of the lowest expressor across all controls, and used that as a threshold to quantify the frequency of NGFRHIGH cells in each of our knockout samples. Then, we calculated the change in frequency of NGFRHIGH cells in each test condition compared to controls and obtained a median fold change and standard deviation across all samples with knockout of one same protein (~3 different biological samples per protein). In total, we targeted ~86 different proteins across ~258 different knockout biological samples.
Image analysis of colony formation
We developed a custom MATLAB pipeline for counting cells and colonies in tiled images of DAPI-stained cells (https://bitbucket.org/arjunrajlaboratory/colonycounting_v2/src/default/). First, the software stitches the individual image tiles into one large image by automatically (or with user input) determining the amount of overlap between each individual image. Then, the software identifies the location of each cell in the stitched image by searching for local maxima. We then manually identify the colony boundaries and quantify the number of colonies in each sample. We then calculate the frequency of resistant colonies by dividing the number of colonies by the total number of cells present in culture prior to BRAFV600E inhibition. Finally, we scale the frequency of colonies to colonies per 10,000 cells and calculate the change in frequency between each sample and the median change across controls.
RNA-sequencing and identification of differential expression
We sequenced mRNA in bulk from WM989-A6-G3 and WM989-A6-G3-Cas9 populations as per Shaffer et. al. In addition to quantifying the transcriptome of EGFRHIGHcells, NGFRHIGH, NGFRHIGH/EGFRHIGH cells and vemurafenib-resistant cells, we quantified the transcriptional changes following the knockout of many tier 1 and tier 2 hits from both the state and fate screens. In addition to hits from our screens, we also quantified the transcriptome of targets that were not tier1 or tier2 hits, but show a change in the frequency of NGFRHIGH/EGFRHIGH cells or of cells resistant to vemurafenib. In total, we targeted ~85 different proteins, each in triplicate (using different single guide RNAs) for a total of 280+ RNA sequencing samples. For each sample, we isolated mRNA and built sequencing libraries using the NEBNext Poly(A) mRNA Magnetic Isolation Module and NEBNext Ultra RNA Library Prep Kit for Illumina per manufacturer instructions. We then sequenced the libraries via paired-end sequencing (36×2 cycles) on a NextSeq 500. We aligned reads to hg19 and quantified reads per gene using STAR and HTSeq. We then used DEseq2 to identify differentially expressed genes.
Gene set enrichment analysis
To identify “biological signatures” enriched or depleted following the knockout of a given target we used the GSEA software (http://software.broadinstitute.org/gsea/index.jsp) in combination with several of the MSigDB gene sets. In particular, we focused in the Biological Process ontology of the Gene Ontology gene sets (http://geneontology.org) to obtain enrichment scores.
Grouping of targets based on transcriptomic analysis
To group targets into classes based on their transcriptional effects, we clustered all RNA-seq samples (hierarchical clustering via pheatmap in R) based on the change in expression (as obtained by DEseq2) of any gene differentially expressed (two-fold change over control, with an adjusted p value ≤ 0.05) in at least one of the 85+ knockouts. We also grouped targets via pheatmap based on the enrichment scores obtained via GSEA. Additionally, we identified the genes in each knockout driving the enrichment score (leading edges) and identified the KEGG pathways and Gene Ontology terms enriched for those genes via cluster profiler. Finally, to identify the axes that account for the variability between each knockout we also performed principal component analysis based on the gene set enrichment scores of each knockout. Note that in the aforementioned analysis we included the transcriptomes of pre-resistant cells (marked by the expression of EGFR alone, NGFR alone, and NGFR and EGFR in combination) and of cells resistant to vemurafenib.
Supplemental Tables
All tables can be found at: https://www.dropbox.com/sh/t08558cl4mepfm6/AABBvbtlTPSNNPoMC9NTro-9a?dl=0
-Supplemental table 1. - sgRNA sequences of primary screens - Epigenetic targets
-Supplemental table 2. - sgRNA sequences of primary screens - Kinase targets
-Supplemental table 3. - sgRNA sequences of primary screens - Transcription factor targets
-Supplemental table 4. - Master table with results from genetic screens and their follow up experiments
-Supplemental table 5. - Targets included secondary screen
-Supplemental table 6. - Enrichment analysis of leading edges
Software and data availability
All data and code used for the analysis can be found at https://www.dropbox.com/sh/t08558cl4mepfm6/AABBvbtlTPSNNPoMC9NTro-9a?dl=0 The software used for colony growth image analysis can be found at: https://bitbucket.org/arjunrajlaboratory/colonycounting_v2/src/default/. The software used for analysis of immunofluorescence images can be found at: https://bitbucket.org/arjunrajlaboratory/rajlabscreentools/src/default/
Declaration of interests
A.R. and S.M.S. receives patent royalty income from LGC/Biosearch Technologies related to Stellaris RNA FISH probes. All other authors declare no competing interests
Author contributions
E.T., J.S., and A.R. designed and supervised the study. E.T. performed the experiments and analysis. E.A., K.B. assisted with CRISPR screens. S.B. assisted with tissue culture, image acquisition, and analysis. L.B. designed image analysis software. B.E, S.S. assisted with acquisition of transcriptomic data. B.E, S.S., I.M. assisted with data analysis. A.W. provided guidance on interpretation of the data.
Acknowledgements
We want to thank Dr. Meenhard Herlyn for always providing excellent advice and guidance. We also thank the Flow Cytometry core, especially Florin Tuluc, at CHOP for all their advice and help. We also thank all members of the Raj Lab as well as John Murray for their comments and suggestions. We thank C. Vakoc for providing the TF, epigenetic regulator, and kinase domain-focused sgRNA library. J.S acknowledges support from Linda Pechenik Montague Investigator Award and Cold Spring Harbor Laboratory sponsored research. AR acknowledges NIH/NCI PSOC award number U54 CA193417, NSF CAREER 1350601, P30 CA016520, SPORE P50 CA174523, NIH U01 CA227550, NIH 4DN U01 HL129998, NIH Center for Photogenomics (RM1 HG007743), and the Tara Miller Foundation. AW acknowledges support from CA207935 and CA174746. AW acknowledge support from CCSG P30CA010815 and NIH U01 CA227550.
Footnotes
↵*** Lead contact