Placing purines in precision medicine: Targeting a metabolic reliance in KRAS-mutant tumors

Summary

Precision oncology workflows rely heavily on genomic identification of oncogenic driver mutations or the functional loss of tumor suppressors. These pipelines can identify single-agent treatments for patients, but monotherapy is often insufficient and can drive resistance. Recently, functional drug screening has been employed to evaluate tumor-specific drug sensitivities that complement molecular testing. We describe a resistance evaluation after first line exposure (REFLEX) multi-omic paradigm using drug-induced molecular changes to prioritize effective hits from combination screening. In KRAS-mutant cancer models, trametinib treatment caused dysregulation of the purine biosynthetic pathway driven by reductions in enzyme GART. This induced vulnerability nominated purine analog 6-thioguanine as a synergistic partner. Across diverse KRAS-mutant lineages, trametinib-induced GART loss predicts sensitivity to the combination. In vivo, the treatment significantly increases overall survival without systemic toxicity. Integrating drug-induced multi-omic changes with functional screening identifies therapeutic strategies, supporting the use of purine analogs with MEK inhibitors for KRAS-mutant tumors.

Graphical abstract

Highlights

• •Drug-induced multi-omic changes help identify combination therapies
• •MEK inhibition reduces GART and disrupts purine biosynthesis in KRAS-mutant cells
• •Trametinib and 6-thioguanine are synergistic in tumors with induced GART loss

Oncology; Precision medicine; Pharmacology; Human metabolism; Systems biology

Introduction

The goal of precision oncology is to identify effective therapeutic strategies for individual patients. While the end goal of these pipelines remains largely focused on improving patient outcomes, the methods employed—particularly the use of genomic and functional modeling—are highly variable and are still being optimized.^1,2,3 While genomic-based strategies can identify many of the drivers of oncogenic transformation and progression, this information is often insufficient for the identification of effective treatment strategies for patients with advanced disease.^4,5 Conversely, functional modeling approaches like drug screening can identify tumor-specific drug sensitivities, but often the monotherapies they identify are limited due to acquired and adaptive resistance mechanisms that allow tumors to escape treatment.

This is exemplified in the first iteration of our precision medicine pipeline where we identified trametinib as the top hit in a single-agent drug screen of a patient with metastatic colorectal cancer (patient C) harboring a KRAS^G13D mutation. Despite promising in vitro results, inhibition of the MAPK pathway for KRAS-mutant colorectal cancer patients has shown limited efficacy in altering overall patient survival both as a mono and as a combination therapy.^6,7,8,9 Furthermore, even with the development of the first KRAS inhibitors, sotorasib and adagrasib, we still only see modest improvements in survival due to mechanisms of acquired resistance.¹⁰ Moreover, the complexity of optimizing precision oncology workflows is compounded by the need to identify rational combination therapies as current approaches often lack the capacity to differentiate and prioritize between the identified combinations. This is evidenced by our initial combination drug screen in patient C’s organoid, where we identified a myriad of potentially effective combination therapies, such as trametinib and celecoxib; however, they were already known to be broadly toxic.⁴ These observations underscore the necessity to identify and prioritize effective combination treatment therapies for patients.

To overcome these limitations and find functional combination therapies, we expanded our pipeline by incorporating a multi-omic assessment of drug-induced changes, thereby enabling us to focus on combinations that exploit drug-induced tumor-specific sensitivities. Our initial pipeline incorporated whole exome sequencing, transcriptomics, and high throughput drug screening of patient-derived organoids (PDOs) in combination with in vivo xenograft validations to identify tumor-specific drug sensitivities.⁴ We refined this pipeline by adding a secondary multi-omic analysis, including proteomics, metabolomics, and transcriptomics, of patient-derived xenografts (PDXs) that were treated with a clinically relevant dose of drug chosen in our initial single-agent drug screen. In this paper, we employed this method in PDXs of our colorectal cancer patient C whose tumor was found to be sensitive to trametinib in our single-agent screen. The xenografts were treated with or without trametinib as a means by which to focus on physiologically relevant changes that occur within the tumor in response to MEK inhibition. By incorporating these multi-omic changes, we could filter drug combinations from the putative hits in our combination drug screen that target the tumor-specific detriments we see in the patient’s tumor. This allows us to probe drug-induced signaling axes and interrogate resistance mechanisms in tumors in innovative ways (Figure 1).

Figure 1.

Precision medicine pipeline workflow

High throughput drug screening is performed on patient derived tumor organoids to identify top single-agent hits. Xenografts derived from these organoids are then treated with initial hits from the pipeline in vivo. Tumors are excised and sent for resistance evaluation after first line exposure (REFLEX) multi-omics analysis while concurrently combination drug screening is conducted using a sensitizing model specific IC30 of the drug from the initial screen with every drug in our library. This allows us to focus on agents and combinations that target treatment induced vulnerabilities. Selected combination therapies are then validated using a 9-point grid screen and finally in vivo to evaluate toxicity and efficacy. Created with biorender.com.

Using our multi-omic pipeline on trametinib-treated xenografts, we discovered a convergence of alterations in purine metabolism in KRAS-mutant tumors driven by a reduction in the purine biosynthetic enzyme GART (glycinamide ribonucleotide transformylase). GART is a trifunctional enzyme with enzymatic activities including glycinamide ribonucleotide synthetase (GARS), glycinamide ribonucleotide formyltransferase (GARFT), and aminoimidazole ribonucleotide synthetase (AIRS) that has been implicated in a variety of cancers, including hepatocellular carcinoma and colorectal cancer.^11,12 These data build on a growing body of evidence that oncogenic KRAS rewires a cell’s metabolism via purine (adenine and guanine) and purine nucleotide (adenosine monophosphate and guanosine monophosphate for example) levels to support hyperproliferation. It is known that oncogenic KRAS metabolically reprograms glucose metabolism to allow for increased glucose uptake and glycolysis.^13,14 The glycolytic intermediates can be shunted into the non-oxidative pentose phosphate pathway thus supplying the cancerous growth with ribose bases via ribose-5-phosphate for nucleotide synthesis.^15,16,17 In addition, KRAS-mutant cells require large amounts of glutamine as a necessary purine precursor. To meet this biosynthetic demand for glutamine, such cells utilize macropinocytosis to scavenge proteins to make glutamine as a fuel source for central carbon metabolism and nucleotide formation.¹⁸ Prior studies also show that KRAS-driven NSCLC cells increase their dependence on folate metabolism via MYC to support de novo purine biosynthesis.¹⁹ This pathway supplies 10-formyl-THF for GART to catalyze the assembly of the imidazole ring of purines. Previous literature has also reported that oncogenic KRAS can decouple glycolysis and glutamine metabolism to support hyperproliferation by reducing reactive oxygen species thus maintaining the cellular redox state.^20,21 Our observed decrease in GART from our multi-omics and this documented tumor-specific purine dependence in KRAS-mutant tumors further informed us on how to best choose a therapy that augments trametinib treatment from our combination drug screen.

Building on our data and known changes in purine biosynthesis in KRAS-mutant tumors, we revisited our initial combination drug screen to leverage this metabolic vulnerability. Taking a non-biased approach, we integrated our trametinib-induced dysregulation of purine metabolism from our multi-omics and focused on the combination of trametinib with the purine analog thioguanine. Mechanistically, thioguanine is a prodrug that exerts its antiproliferative activity via its active metabolites thioguanosine monophosphate (TGMP), methylated TGMP (meTGMP), and thio-deoxyguanosine triphosphate (Thio-dGTP).²² These antimetabolites contribute to the cytotoxic effects of thioguanine by being actively incorporated into DNA and RNA thus inhibiting DNA replication and protein synthesis, respectively, and by blocking de novo purine biosynthesis.^23,24,25 Thioguanine was a rational choice as it has safely been used in the clinic, in particular in heme malignancies, for over 70 years with doses upwards of 100 mg/m^2.26 Given thioguanine’s pharmacological effects, it could be therapeutically leveraged to target the observed detriment in GART and in purine metabolism. This would have been otherwise overlooked as a rational therapy with trametinib if not for the multi-omics. Our findings suggest that multi-omic data can refine precision oncology workflows by elucidating tumor-specific detriments and guide the rational selection of drug candidates that target these perturbations. It also raises the possibility that reductions in GART in KRAS-mutant tumors could be a clinical biomarker for combination thioguanine and MAPK inhibitor therapy including targeting MEK1/2 and ERK1/2.

Results

Precision medicine pipeline identifies targetable purine biosynthesis detriment

To identify novel therapeutic combinations that might synergize with trametinib, we established a precision oncology workflow (Figure 1) whereby xenografts of a patient’s KRAS-mutant colorectal cancer cells (patient C) were treated with trametinib or vehicle control for five days in FOXN1^nu mice. Tumors were excised and sent for metabolomics, transcriptomics, and proteomics to identify drug-induced changes in the cancer cells.

When compared to control, we found that PDX tumor cells treated with trametinib significantly altered levels of the metabolites riboflavin, cysteine and guanine after false discovery rate adjustment (log₂ FC = 1.78, p_adjusted = 0.029, log₂ FC = −2.99, p_adjusted = 0.000046, and log₂ FC = −2.26, p_adjusted = 0.019, respectively) (Figure 2A and Table S1). The guanine results were confirmed via an in vitro guanine colorimetric assay that showed decreased guanine levels after the addition of increasing concentrations of trametinib in four KRAS-mutant cell lines including Mia PaCa-2, A549, HCT116, and DLD-1 (Figure S1A). These data are consistent with and build upon data in the literature that MEK inhibition leads to a reduction in purine metabolite levels such as adenosine monophosphate and inosine monophosphate (the precursor to both adenosine monophosphate and guanosine monophosphate).¹⁵ In parallel we ran transcriptomics and proteomics on the same PDX samples to further investigate the trametinib-induced changes in the cancer cells.

Figure 2.

Trametinib dysregulates purine biosynthesis and cooperates with Thioguanine in drug screening of a patient-derived model of colorectal cancer

(A) Differential metabolite analysis of trametinib-treated tumors revealed significant dysregulation of three metabolites (riboflavin, cysteine, and guanine) after false discovery rate adjustment (q < 0.05) when compared to DMSO-treated tumors. n = 3 for each treatment.

(B) Volcano plot of the significantly disrupted proteome after trametinib treatment (red differentially increased and blue differentially decreased). Differential abundance was defined by a p_adjusted < 0.05 and a log₂ Fold Change >1 or < −1. n = 5 for each treatment.

(C) Ingenuity pathway analysis of proteomics dataset for pathway enrichment based on gene ontology (GO) terms. Proteins deemed statistically significant (ANOVA p < 0.05) and possessing an intensity ratio of <0.75 (trametinib/DMSO control) were only submitted. Pathways with a p < 0.05 were deemed statistically enriched for downregulated proteins in response to trametinib.

(D) Differential expression of proteins involved in de novo purine biosynthesis with GART highlighted in blue as the only significantly downregulated protein. These data were also plotted in (B).

(E) Analysis of combination high throughput drug screen scatterplot highlighting thioguanine as cooperative with trametinib. y axis indicates the log2 fold change of the AUC of the combination agent along with trametinib (thioguanine+trametinib for example) divided by the AUC of the single agent alone (thioguanine for example). A log2 AUC fold change <0 indicates that the combination of trametinib and thioguanine reduced cell viability more than single agent thioguanine alone. x axis indicates how the patient’s tumor responded to every single-agent in the 120 agent drug screen as a Z score.

(F) Patient C ex vivo organoid validation demonstrates that the addition of thioguanine to trametinib treatment enhances the anti-tumor effect. Red line is the single agent thioguanine, blue line is the single agent trametinib, and the green line represents an IC₅₀ of trametinib of ∼5.7 nM plus a dose response of thioguanine. Dose response curves demonstrate means ± SD.

In our proteomics dataset we performed Ingenuity Pathway Analysis on trametinib treated versus control treated cells. Trametinib treatment resulted in a dysregulation of over 941 proteins involved in major cell signaling pathways such as EIF2 signaling and mTOR signaling suggesting that the drug’s effect on the MAPK pathway is heavily involved in cellular growth and proliferation, as expected (Figures 2B and 2C, Table S2). Notably, we also observed that the purine biosynthesis pathway was significantly enriched, which corroborated our metabolomics finding that guanine levels were decreased. Given this observation, we further investigated whether there was a disruption in a purine biosynthesis enzyme that would contribute to this observed effect. Within the pathway, we observed a significant dysregulation of only one of the enzymes, GART (log₂ FC = −1.26, p_adjusted = 0.0059) (Figure 2D). In predominantly KRAS-mutant tumors like pancreatic adenocarcinoma and lung adenocarcinoma, GART expression has been correlated with highly unfavorable survival (Figure S1B). Furthermore, using DepMap we found that GART’s protein expression in KRAS-mutant cell lines is highly correlated with that of both MEK 1 and MEK 2 suggesting that this axis could be targetable via an inhibitor like trametinib (Figure S1C). In addition, we investigated other metabolic pathways including tetrahydrofolate salvage and pyrimidine biosynthesis. We found that thymidylate synthase (TYMS) and methylenetetrahydrofolate dehydrogenase 1 (MTHFD1) were statistically significantly decreased (log2 FC = −1.57, p_adjusted = 0.027 and log2 FC = −0.69, p_adjusted = 0.026, respectively) in the tetrahydrofolate pathway and that the pyrimidine biosynthesis enzymes CAD and Uridine 5′-monophosphate synthase (UMPS) were also significantly decreased post-trametinib treatment (log2 FC = −0.93, p_adjusted = 0.023 and log2 FC = −1.13, p_adjusted = 0.008) (Figure S1D and Table S2). We also looked into members of the pentose phosphate pathway which provides ribose-5-phosphate as a metabolite for the creation of de novo purines and found that ribose-5-phosphate and glyceraldehyde-3-phosphate were decreased but were found to not be statistically significant (log2 FC = −0.61, p_adjusted = 0.78 and log2 FC = −0.76, p_adjusted = 0.66). Together, these data highlight the importance of metabolism, and in particular de novo purine biosynthesis, in maintaining the proliferative state in KRAS-mutant cells. It also highlights a potential axis for targeting this GART-induced metabolic detriment after trametinib treatment.

In our transcriptional dataset, trametinib treatment significantly decreased the expression of genes targeted by MYC and E2F within gene sets of Hallmark “MYC Targets V1” and “E2F Targets” (normalized enrichment score [NES] = −1.93, p_adjusted < 0.005 and [NES] = −1.69, p_adjusted < 0.05, respectively) (Figure S1E and Table S3). Both MYC and E2F are transcription factors that have been implicated in regulating genes involved in nucleotide biosynthesis and purine metabolism, including GART.^27,28,29 There was a trend for GART’s mRNA level to decrease, but this decrease was found to not be statistically significant after multiple testing correction (log₂ FC = −0.56, p_adjusted > 0.05). Furthermore, RT-PCR was performed on trametinib treated MIA PaCa-2 cells treated with 50 nM trametinib and a similar trend of decreases in GART mRNA observed but was found to not be statistically significant (Mann Whitney, p = 0.10) (Figure S1F). As a result, the transcriptional control of GART may not be related to its loss of protein after MEK inhibition, and the mechanism by which is not fully understood. A similar trend was also seen in the protein set enrichment, although it was not significant after multiple testing corrections (Figure S1G and Table S2). Taken together, these data indicate that trametinib induces a tumor-specific detriment in purine metabolism and biosynthesis.

In conjunction with the multi-omics approaches and to nominate drugs that may synergize with trametinib, cells of patient C were subjected to high throughput combination drug screening using a patient-model specific IC₃₀ concentration of trametinib against a library composed of 120 compounds.⁴ Using this method, we identified combinations that could potentially improve patient responses. Unlike single-agent drug screening, in combination screening, it is difficult to identify and isolate tumor-specific sensitivities as opposed to combinations that are universally toxic due to the infeasibility of running combination screens using each single-agent drug as a sensitizing agent for each cell line. As a result, the necessary points of comparison across multiple models to identify tumor-specific effects are not available. As exemplified by the initial hits from this combination screen pursued without multi-omic consideration, many nominated drug combinations with trametinib were generally toxic (such as trametinib and celecoxib in green or trametinib and erlotinib in blue), and thus, did not supply a sufficient therapeutic window (Figure 2E). In the absence of multiple points of comparison afforded in single-agent screens, we integrated the observation of purine metabolic and biosynthetic dysregulation from our omics datasets into our decision making. Of note, there were multiple inhibitors of purinergic signaling including pemetrexed, methotrexate, cladribine, fludarabine and thioguanine. We found that the antifolates, pemetrexed and methotrexate, performed antagonistically in this patient’s tumor with positive log2 area under the curve (AUC). On the other hand, the purine analogs cladribine, fludarabine and thioguanine all had negative log2 AUCs suggesting that purine analogs as a class perform quite well with trametinib (Figure S1H). We selected the nucleotide analog thioguanine, a guanine mimetic, as a potentially synergistic therapeutic agent with trametinib due to its safe use in the clinic for decades and due to its particular ability to mimic guanine in the cell.²⁶ We thought that a guanine analog could perform particularly well with trametinib due to its ability to take advantage of the observed intracellular guanine deficit we saw in our metabolomics as it would be incorporated into ribonucleotides just as guanine would be. Thioguanine appeared to be a poor single-agent therapy in these patients’ tumor with a positive Z score, but demonstrated a large decrease in log2 AUC when in combination with trametinib (Figure 2E). The combination of trametinib and thioguanine was validated in an ex vivo organoid assay, showing a marked sensitivity to the combination treatment compared to either of the single agents alone (Figure 2F). We find that at even low nanomolar doses of trametinib given with an increasing dose response of thioguanine that there is substantial cell death in this organoid model with reductions in cell viability to roughly 10% which is greater than either single agent alone (trametinib at ∼70% and thioguanine at 100% viability respectively).

Pharmacological and genetic perturbation of MEK1/2 decreases GART levels

Our proteomics findings showed that trametinib induced a dysregulation of the purine biosynthetic enzyme GART in a patient’s KRAS-mutant colorectal tumor (Figure 2D). To validate that decreases in GART levels occurred after trametinib inhibitor treatment, we ran western blots probing for GART on a panel of ten KRAS-mutant cell lines, two KRAS-mutant organoids and four KRAS WT cell lines from different tissue lineages, including colon, pancreas, lung, kidney and breast. Cell lines were plated and dosed with increasing concentrations of trametinib from 10 nM to 10 μM for 48 h 80% (8/10) of the KRAS-mutant cell lines showed dose-dependent decreases in GART levels upon trametinib treatment (Figure 3A). In the KRAS WT cell lines including HEK 293T, NCI-H508, BxPC-3, and RKO we found either no loss or slight loss of GART at higher doses of drug (Figure S2A). Furthermore, we performed a similar experiment on two KRAS-mutant organoids whereby we dosed cells with increasing concentrations of trametinib for five days and found similar reductions in GART (Figure S2B). GART loss appeared in a time-dependent manner and was noticeable at 48 h in two of the trametinib-sensitive lines, K8484 and MIA PaCa-2 (Figure S2C). Interestingly, in our panel there were three cell lines (MDA-MB-231, H647, and HEK 293T) that maintained GART protein levels after trametinib treatment indicating that there may be mechanisms to sustain GART levels in these cells that circumvent MEK inhibition.

Figure 3.

Targeting of MEK 1/2 reduces GART protein levels

(A) Western blots of KRAS-mutant pancreas, colon, lung, and breast cell lines show a trametinib dependent loss of GART after 48 h of treatment. GART Retainers were defined as those not losing more than 20% of GART proteins levels after the highest dose of trametinib. Quantification was determined by normalizing each sample to its own actin control.

(B) Representative western blot analysis of MIA PaCA 2 cells treated with pimasertib, cobimetinib and trametinib for 48 h (doses are control, 10 nM, 100 nM, 250 nM, 500 nM, 1 μM).

(C) Log2 fold change quantification of western blots from B. GART protein levels were normalized to each sample’s actin control and then to the “control” group. n = 4 biological replicates were performed for each inhibitor.

(D) Representative western blot analysis of genetic knockdown of MEK 1/2 using siRNA after 48 h.

(E) Log2 fold change quantification of western blots from (D). GART protein levels were normalized to each sample’s tubulin control and then to the “control” group. n = 5 biological replicates were performed for each siRNA. All bar graphs display a mean ± SD. Two-tailed Mann Whitney tests were performed for significance (no asterisk p ≥ 0.05, ∗p < 0.05, ∗∗p < 0.01).

To confirm that this effect was not due to the drug trametinib alone, we profiled two other MEK 1/2 inhibitors, pimasertib and cobimetinib, in the pancreatic KRAS^G12C-mutant MIA PaCa-2 cell line. Consistent with the trametinib data, both pimasertib and cobimetinib decreased GART levels after 48 h of drug treatment with log2 fold changes in GART levels less than −1.0 (Figures 3B and 3C). To further evaluate the specificity of these effects, we used siRNAs targeting both MEK1 and MEK2 and transfected MIA PaCa-2 cells with 25 nM of each siRNA both separately and in combination.³⁰ As shown by western blot, knockdown of MEK2 (Mann Whitney, log₂ FC = −0.54, p = 0.0079), but not MEK1 (Mann Whitney, log₂ FC = −0.19, p = 0.68), contributed to a reduction in GART levels (Figure 3D). A further reduction in GART level was achieved when siRNA of both MEK1 and MEK2 were used, exhibiting an even greater log2 fold change of GART levels (Mann Whitney, log₂ FC = −0.92, p = 0.0079) (Figure 3E). This loss of GART was concomitant with loss of phosphorylation of ERK indicating that the siRNAs were working as expected, and rather, we found no changes in phosphorylated or total protein levels of ribosomal protein S6, suggesting that under these conditions, protein synthesis may not be affected. Together, these data suggest that GART protein levels in KRAS-mutant tumors are responsive to changes in MEK1/2.

Given the selectivity of ERK1/2 as a substrate for MEK 1/2, we next investigated whether loss of GART could also be mediated by pharmacological inhibition and genetic abrogation of ERK1/2 levels.³¹ Inhibition of ERK1/2 was performed with two differing inhibitors, ravoxertinib and temuterkib. In concordance with pharmacological inhibition of MEK1/2, both ravoxertinib and temuterkib reduced GART protein levels (Figure S3A). We observed that temuterkib decreased GART levels, as measured by log2 fold change, starting at a concentration of 500 nM (Mann Whitney, log₂ FC = −1.06, p = 0.0286 for 500 nM and log₂ FC = −1.20, p = 0.0286 for 1 μM respectively) whereas ravoxertinib only reduced GART levels at the highest dose of 1 μM (Mann Whitney, log₂ FC = −0.75, p = 0.0286) (Figure S3B). This result suggests that there is a MAPK pathway-level effect that contributes to loss of GART. To test whether genetic loss of ERK1/2 would also contribute to the GART loss phenotype, we designed siRNA against both isoforms of the protein.³² At a dose of 25 nM of either ERK 1, ERK 2 or both ERK1/2 siRNA, we found that there was no significant reduction in GART levels even though ERK1/2 levels were reduced via western blot (Figures S3C and S3D). Given these data, we surmised that ERK1/2 may influence GART levels to a lesser extent than reductions in MEK1/2. To compare the MAPK pathway level effect of MEK1/2 and ERK1/2 inhibition on GART levels, we dosed a panel of KRAS-mutant cell lines including ASPC-1, A549, HCT116, and MIA PaCa-2 cells with either trametinib or temuterkib and used DUSP6 levels as a readout for pathway activity. DUSP6 has been implicated as a key protein in MAPK pathway signatures with its protein levels decreasing after MEK and ERK inhibitor treatment.^33,34,35 In both the MEK1/2 and ERK1/2 inhibitor-treated cells, we observed a concomitant loss of GART with decreases of DUSP6, indicating that the MAPK pathway was turned off (Figure S3E). These results indicate that both MEK1/2 and ERK1/2 inhibitors can modulate GART levels suggesting that the effects may be mediated at the pathway level.

Concurrent trametinib and thioguanine therapy decreases tumor growth in vitro and extends survival in vivo

To further validate the therapeutic efficacy of trametinib and thioguanine outside of our n = 1 patient sample, we dosed a cohort of KRAS-mutant and KRAS WT cells using a 9-point dose response of both drugs in an 81 point grid matrix, as previously described (Figures S4A and S4B).³⁶ We observed enhanced cell death in cell lines that lost GART (i.e., GART reducers) when they were given a constant 1.11 μM dose of thioguanine (below the maximum serum concentration of drug seen in patients) in conjunction with a dose response of trametinib (Figures 4A and S4C).³⁷ In the cell lines that did not lose GART after trametinib treatment (i.e., GART retainers), we did not detect an added cell killing effect in the combination dose responses as the cell viabilities were no less than that of 1.11 μM thioguanine alone (Figures 4B and S4C). We also found that the AUC of the combination dose response was significantly lower than that of the AUC of single-agent thioguanine in the cell lines that lost GART (Mann Whitney, p < 0.0001) (Figure 4C) which we did not see in the cell lines that retained GART (Mann Whitney, p = 0.10). These data demonstrate that cell lines that lose GART are more sensitive to the combination of trametinib and thioguanine. Given these observed decreases in GART and guanine in our western blots and multi-omics datasets we proceeded to correlate how much of the observed decrease in guanine is due to losses of GART. Plotting our guanine colorimetric assay against calculated loses of GART from our western blots in MIA PaCa-2 cells we found that there was a positive correlation between GART levels and guanine level which indicated that 41% of the variance in guanine levels is explained by this linear relationship with GART (Figure S5A).

Figure 4.

In vitro validation of the combination of trametinib and thioguanine

(A) Dose-response curves of various KRAS-mutant cell lines that lose GART after trametinib treatment. Cells were treated for 72 h with either single agent trametinib (red), single agent thioguanine (blue), or a combination of 1.11 μM thioguanine with a dose response of trametinib (green).

(B) Dose-response curves of two KRAS-mutant cell lines that do not lose GART after trametinib treatment. In (A) and (B), each point is the average of three technical replicates, error bars indicate ±SD, and dose-response curves were generated via a four-parameter logistic model fit and normalized to DMSO control.

(C) Comparison of area under the curves (AUCs) in the single agent thioguanine group versus the combination therapy in cell lines that lose GART and in cell lines that retain GART showing marked reduction in the combination (Two-tailed Mann Whitney, p < 0.0001, n = 3) when you lose GART but no change when you retain GART (Two-tailed Mann Whitney, p = 0.10, n = 11).

(D) Log (natural log) Loewe additivity model combination index (CI) of each cell line in (A) and (B) along with the KRAS WT cell lines for the combination of trametinib and thioguanine. Values were calculating by averaging each dose combination across an 81-point grid screen. Values below −0.1 were determined to be synergistic. GART retainers are shown in blue and GART Reducers are shown in red.

(E) MuSyC synergy framework of all 14 cell lines tested. Observed synergistic efficacy (β_obs, x axis) versus how trametinib’s potency is affected by thioguanine (log α₂₁, y axis). Log(α) and β_obs levels >0 are synergistically potent and synergistically efficacious, respectively, and log(α) and β_obs levels values <0 are antagonistically potent and antagonistically efficacious, respectively. H647, MDA-MB-231, and 293T cells have no observed synergistic efficacy or potency.

(F) Representative western blots of GART protein levels in HCT 116 cells after given 50 nM of siRNA targeting GART or a scramble control for two or five days.

(G) Dose-response curves of thioguanine in either siGART- or scramble control-treated HCT 116 cells after 72 h of dosing (day 5 of siGART treatment in panel F).

(H) Western blots of GART protein levels in MIA PaCa 2 cells treated for two or five days with 50 nM of siRNA targeting GART or a scramble control.

(I) Dose-response curves of thioguanine in either siGART or scramble control-treated MIA PaCa 2 cells after 72 h of dosing (day 5 of siGART treatment in panel H).

(J) Western blot of 293T GART knockdown treated for 5 days with 100 nM of siRNA targeting GART or scramble control.

(K) Dose response of 293T cells treated with thioguanine and given either siGART or scramble control after 72 h of dosing. In G, I, and K, each point is the average of five technical replicate measurements, error bars indicate ±SD, and dose-response curves were generated via a four-parameter logistic model fit. Conditions were normalized to DMSO of scramble control. All graphs with error bars display a mean ± SD.

(L) AUC of combination trametinib/thioguanine divided by single agent AUC of thioguanine on the y axis for every cell line at varying constant dosages of trametinib shown on the x axis. GART retainers in blue have no change in the ratio of combination AUC to single Agent AUC indicating insensitivity to the combination, but GART reducers have decreases in the AUC ratio indicating sensitivity to the combination. All graphs display a mean ± SD. Two-tailed Mann Whitney tests were performed for significance against each mean of the GART retainers versus the GART reducers for each constant dose of trametinib (no asterisk p ≥ 0.05, ∗p < 0.05, ∗∗p < 0.01).

To assess the extent of cooperativity of the combination, synergy scores were calculated across the cell lines using both the Loewe synergy model and the MuSyC synergy framework.^38,39 A log Loewe synergy score < −0.1 was determined to be synergistic, values >0.1 were labeled antagonistic, and scores between −0.1 and 0.1 were considered additive. Using this method, we determined that the combination was synergistic in cell lines that lost GART (red) but was either slightly additive or antagonistic in cell lines that did not (blue) (Figure 4D). We also observed synergy across the entire 81-point dose response grid screen using synergy heatmaps of the Loewe synergy framework. Using the loewe model, cell lines that lost GART were found to have moderate synergy with values all above three whereas cell lines that did not lose GART showed either antagonistic or additive effects (Figure S4E). The bliss synergy framework was also tested but was found to have little change across the cell lines tested with values hovering around zero for most cell lines tested (Figure S4F). Given the Loewe model assumes drug dependency whereas the bliss model assumes drug independency, we find that the Loewe framework is better suited for this drug combination as thioguanine takes advantage of the purine detriment that trametinib is inducing. Using an orthogonal approach, we also tested whether the combination was synergistically potent or synergistically efficacious using the MuSyC synergy model. We found that in all of the GART reducing cell lines, the combination was either synergistically potent or synergistically efficacious. This is in contrast to the three GART retaining cell lines, where the combination was antagonistically potent and/or antagonistically efficacious (Figure 4E), which corroborates our Loewe synergy results. Furthermore, to determine if purine supplementation can rescue the observed synergistic effect, we dosed K8484, ASPC-1, and 293T cells with combination trametinib and thioguanine but with the addition of 1 μM of adenosine and guanosine. We find that purine supplementation can rescue the observed synergy by reducing the Loewe synergy value in the cell lines that reduce GART, K8484, and ASPC-1, but not in a cell line that retains GART, 293T (Figure S5B). To discern whether it is the loss of GART or other effects of MEK inhibition driving sensitivity to thioguanine therapy, we knocked down GART via siRNA in two cell lines that lost GART after trametinib treatment (HCT116 and MIA PaCa-2) and performed a dose response of thioguanine. At the endpoint of the experiment on day 5 we confirmed that GART levels were diminished in both cell lines (Figures 4F and 4H). We found that the addition of thioguanine to cells treated with GART siRNA had a diminished capacity to grow with reductions in cell viability (Figures 4G and 4I). As an orthogonal approach, we also knocked down GART in a cell line that did not lose GART after trametinib treatment (293T) and found that we can sensitize these cells to thioguanine therapy in the siGART treated cells when compared to the scramble control (Figures 4J and 4K). Finally, we compared the AUCs of every constant dose of trametinib plus thioguanine to the AUC of single agent thioguanine alone in both the GART retainers and the GART reducers. We found that the addition of trametinib to thioguanine therapy dramatically decreased the proportional AUC in cell lines that reduced GART whereas in cell lines that retained GART there was no change in proportional AUC indicating that it is loss of GART that is driving the sensitivity to the combination therapy (Figure 4L). These data further support our observations that the addition of thioguanine to trametinib treatment is particularly effective in KRAS-mutant cell models due to the loss of GART and that the exogenous addition of purines can rescue this observed synergy.

As a translational approach to confirm efficacy and assess the toxicity of the trametinib and thioguanine combination, we used both ex vivo organoid models from colorectal and pancreatic patients and an in vivo patient-derived xenograft model from our initial colorectal cancer patient. In our four KRAS-mutant organoid models, we found that there was a substantial increase in cell death when the PDO-specific IC₃₀ of trametinib was given with a dose response of thioguanine (green lines) compared to either single-agent drug alone (Figure 5A). Using the MuSyC synergy model, we also calculated log α₂₁ and β_obs values for each of the organoid models and found that the drug combination was synergistically potent in all lines (log α₂₁ > 0) and synergistically efficacious (β_obs > 0) in two of the four lines (WCM773 and WCM389) (Figure 5B). Furthermore, we also observed in the organoids a large decrease in the AUC of the combination treated group than when compared to the single-agent AUC of thioguanine (Mann Whitney, p = 0.0286) (Figure 5C). In addition, KRAS WT organoids were also tested using the combination and were found to not display an added cell killing effect greater than the single agent IC₃₀ of trametinib indicating that in these cell lines the combination is not efficacious. The combination was also tested in a non-tumorigenic cell line in human foreskin fibroblasts and was found to not display an added effect (Figure S4D). For our in vivo model we initially tested a dose schedule optimization using a syngeneic allograft of the pancreatic cell line K8484 in BL/6 mice where we found that treating the mice three times a week could reduce tumor volume over time (Figure S6A) as compared to once or twice a week. Furthermore, when K8484 mice vehicle treated mice were challenged with trametinib/thioguanine after 10 days of growth, we observed a reduction in tumor volumes (Figure S6B). In our in vivo PDX model, animals were treated orally with trametinib (1 mg/kg) and thioguanine (1 mg/kg) alone or in combination three days a week (n = 5 for all treatment arms). Tumor growth was tracked over time to assess overall and progression-free survival of the animals. Animals treated with trametinib and thioguanine had a statistically significant survival benefit compared to animals treated with either trametinib or thioguanine as a single-agents (Mantel-Cox, p < 0.05), with all the combination mice living after 100+ days of treatment (Figure 5D). Furthermore, the combination delayed the growth of these tumors (Figure 5E). We found that these mice tolerated the combination dosing regimen well with no significant reduction of body mass over the experiment’s time range, indicating the combination regimen at these doses is safe and worthy of further investigation as a potential therapeutic option for KRAS-mutant tumors (Figure 5F).

Figure 5.

Ex vivo organoids and in vivo PDX validation of the combination of trametinib and thioguanine

(A) Dose-response curves of four KRAS-mutant, patient-derived organoid models given either single agent trametinib (blue), single agent thioguanine (red) or a combination of the IC₃₀ of trametinib (WCM039-01 = 2.5 nM, WCM715 = 5 nM, WCM733 = 2.5 nM, and WCM389 = 50 nM) and a dose response of thioguanine (green). Each point is the average of a triplicate measurement and dose-response curves were generated via a four-parameter logistic model fit normalized to DMSO.

(B) MuSyC synergy framework of all four organoids highlighting synergistic potency in all organoid lines. Observed synergistic efficacy (β_obs, x axis) versus how trametinib’s potency is affected by thioguanine (log α₂₁, y axis). Log(α) and β_obs levels >0 are synergistically potent and synergistically efficacious, respectively, and log(α) and β_obs levels values <0 are antagonistically potent and antagonistically efficacious, respectively.

(C) Area under the curves (AUCs) in the single agent thioguanine group compared to the combination therapy in all of the organoid lines indicating a large decrease in AUC in the combination group (Two-tailed Mann Whitney, p = 0.0286, n= 4).

(D) Overall survival of patient C PDX mice treated per oral (PO) with 10% DMSO (control), 1 mg/kg trametinib, 1 mg/kg thioguanine, or combination of 1 mg/kg of trametinib and 1 mg/kg of thioguanine (combo). There was an increased in percent survival in the combination treatment group when compared to single-agent trametinib alone (p = 0.049, df = 1, Mantel-Cox log rank, n = 5 for all treatment arms). Days are indicative of days since the start of drug treatment.

(E) Fold change of tumor volume of each mouse normalized to size of tumor at start of treatment showing a delay of tumor growth in the combination trametinib/thioguanine treatment arm.

(F) Mouse weight over time indicated no change in mass in any treatment arm. All graphs with error bars display a mean ± SD.

All graphs with error bars indicate a mean ± SD. Two-tailed Mann Whitney tests were performed for significance against the mean of Thioguanine AUC in each cell line versus the mean of Thioguanine + Trametinib AUC in each cell line (no asterisk p ≥ 0.05, ∗p < 0.05, ∗∗p < 0.01).

Discussion

The dawn of precision medicine ushered in an era of targeted therapies against single genomic alterations across tumor types and was exemplified in the completed NCI-MATCH trial (NCT02465060).⁴⁰ Although the trial showed that targeting some genetic alterations was efficacious, the overall response rate (ORR) to a previously FDA-approved targeted therapy was roughly 10%. The low ORR may be partially due to single-agent inhibitors being limited by adaptive and acquired resistance mechanisms. For example, the use of RAF and MEK inhibitors can cause the transcriptional upregulation of the receptor tyrosine kinases (RTKs) EGFR and HER3 in BRAF-mutant colorectal cancer and KRAS-mutant lung cancer, respectively.^41,42 This subsequent upregulation reactivates the pathway, thus allowing the cancer cell to become resistant to treatment. Furthermore, we also see acquired resistance mechanisms, including secondary mutations in the MAPK pathway, amplifications, splicing variants, and drug pocket mutants that diminish the efficacy of these targeted drugs as single agents.^10,43,44 Due to these resistance mechanisms, targeting a single enzyme is often insufficient to cure a patient’s cancer. This highlights the need to find combination therapies that either target these resistance mechanisms or target tumor-intrinsic growth and signaling pathways to improve treatment efficacy. This need has spurred the creation of the next generation of the NCI-MATCH trial in ComboMATCH (NCT05564377), which aims to target cancer with combination therapies identified via precision medicine workflows.

In this paper, we addressed the need for identifying combination therapies in precision oncology workflows by complementing traditional functional screening with multi-omic analysis of drug-induced changes. We observed that a colorectal cancer patient with a KRAS^G13D mutation exhibited a high sensitivity to trametinib via high-throughput drug screening. Integration of transcriptomic, metabolomic, and proteomic data of trametinib-treated PDX tumors revealed that trametinib induced a dysregulation of purine metabolism and purine biosynthesis, specifically through the enzyme GART. Due to GART’s central role in proliferation, inhibitors were designed—such as pelitrexol—to prevent its function; however, they have failed in phase 2 clinical trials of metastatic colorectal cancer (NCT00078468) due to poor single-agent efficacy. This monotherapy failure highlights the continued necessity for finding combination therapies that target metabolic axes like purine metabolism in KRAS-mutant tumors. More recently, targeting of KRAS-mutant tumors via GART was performed using multitargeted antifolates, like pemetrexed, that inhibit its catalytic function by competing with 10-formyl-THF binding KRAS-mutant NSCLC patients in a phase 2 clinical trial.

(NCT02185690) that received the MEK inhibitor binimetinib with carboplatin and pemetrexed had a higher ORR (62.5%) than KRAS WT patients (25%) alone in a secondary analysis.⁴⁵ This study shows that by adopting strategies that target both the signaling related changes and the metabolic related changes in KRAS-mutant tumors you can increase antitumor efficacy. The incorporation of the multi-omic datasets with our original combination high throughput drug screen allowed us to identify and target this drug-induced metabolic detriment in KRAS-mutant tumors and informed our choice of thioguanine as a potentially potent augmenter of trametinib therapy. Our study demonstrates that, instead of the broad GART inhibitory strategy employed by drugs like pelitrexol and pemetrexed, targeting of the purine deficiency through thioguanine incorporation into the cell represents a more mechanistic driven approach. Understanding the role that purine metabolism plays in these cancers could therefore serve as a future therapeutic strategy for maximizing anti-tumor effects.

Building on purine biosynthetic changes known to develop in KRAS-mutant tumors, western blot analysis of cell models from various tissue histologies and mutational subtypes indicated that there is a trend for KRAS-mutant tumors to lose GART in response to trametinib in a dose-dependent manner. This effect was also observed with other MEK and ERK inhibitors, suggesting a broader MAPK pathway-dependent mechanism. Further grid screening showed that synergy of the combination was dependent upon loss of GART in the KRAS-mutant lines. We acknowledge that this trend doesn’t necessarily hold across all models. For example, H647 and MDA-MB-231 are both KRAS-mutant lines that did not lose GART after trametinib treatment, and the combination of trametinib and thioguanine did not exhibit a cooperative effect in these two models, perhaps due to their continued expression of GART that precluded the decreases in guanine levels that drives thioguanine sensitivity. Lastly, given that a single driving mutation in KRAS is not enough to explain the heterogeneity of response to trametinib/thioguanine combination treatment; other passenger mutations could play a role in sensitivity to the combination treatment across tumor types.

Altogether, our functional genomics pipeline can be used to leverage multi-omic datasets to elucidate the molecular impacts of known targeted agents and it informs us of potential combination therapies. The development of this pipeline and the identification of combination therapies are useful as it can be a means by which to overcome the innate tumor resistances and dependencies that develop from single-agent therapy. These data presented here suggest that sensitivity to the combination of trametinib/thioguanine is predicated on the loss of the purine biosynthetic enzyme GART caused by trametinib. Moving forward, it might be possible to utilize trametinib-induced GART loss as a clinical biomarker of dual targeting of MEK and purine biosynthesis. For example, to identify patients most likely to benefit from trametinib/thioguanine treatment, patient’s PDXs could be treated with trametinib and immunohistochemistry stained for GART levels. If the tumors show a reduction in GART, thioguanine could be added to augment therapy. Taken together, these results show how the addition of multi-omics into precision medicine pipelines can identify tumor-specific drug combinations and facilitate the translation of mechanistic drug combinations into the clinic.

Limitations of the study

There are several limitations of this study that should be noted. For one, the initial patient PDX data were performed in an n = 1 mouse study. Although the study uses other preclinical models, such as organoids and cell lines, further studies are needed to determine whether these findings are generalizable to other patients. Second, while this study identified that KRAS-mutant tumors are particularly sensitive to the combination of trametinib and thioguanine, the specific molecular subtypes and mutational profiles most likely to benefit from this therapy remain to be defined. Furthermore, the optimal dosing regimen of trametinib and thioguanine to achieve the maximal decrease in GART that allows for large therapeutic effects of thioguanine warrants further investigation. Addressing these issues through additional studies will be essential for developing an optimized regimen that yields maximal therapeutic benefit for individual patients.

Resource availability

Lead contact

Any requests for further information or resources and reagents should be directed to and will be fulfilled by the lead contact, Benjamin D. Hopkins (beh2020@med.cornell.edu).

Materials availability

This study did not generate any new reagents.

Data and code availability

• •There is no original code to report in this paper.
• •Any data that is required to reanalyze data presented in this article are in the supplemental and can be found publicly at Zenodo: http://www.doi.org/10.5281/zenodo.18471666.
• •Any additional information required to reanalyze the data is available from the lead contact upon request.

STAR★Methods

Key resources table

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Antibodies
GART	Abcam	Cat#ab169550; RRID:AB_3720855
Phospho-p44/42 MAPK (Erk1/2)	CST	Cat#9106S; RRID:AB_331768)
p44/42 MAPK (Erk1/2)	CST	Cat#9102S; RRID:AB_330744
DUSP6	Abcam	Cat#ab76310; RRID:AB_1523517
Phospho-S6 ribosomal protein	CST	Cat#5364S; RRID:AB_10694233
S6 Ribosomal Protein	CST	Cat#2317S; RRID:AB_2238583
alpha-Tubulin	CST	Cat#3873S; RRID:AB_1904178
Beta-actin	CST	Cat#3700S; RRID:AB_2242334
IRDye® 680RD Goat anti-Rabbit IgG Secondary Antibody	Licor	Catalog# 926-68071; RRID:AB_10956166
IRDye® 680RD Goat anti-Mouse IgG Secondary Antibody	Licor	Catalog# 926-68070; RRID:AB_10956588
IRDye® 800CW Donkey anti-Mouse IgG Secondary Antibody	Licor	Catalog# 926-32212; RRID:AB_621847
IRDye® 800CW Donkey anti-Rabbit IgG Secondary Antibody	Licor	Catalog# 926-32213; RRID:AB_621848
Anti-rabbit IgG, HRP-linked Antibody	CST	Catalog# 7074S; RRID:AB_2099233
MEK2	Abcam	Catalog #ab776277; RRID:AB_776277
MEK1	CST	Catalog # 2352s; RRID:AB_10693788
Chemicals, peptides, and recombinant proteins
Trametinib	Selleckchem	Cat# S2673
Cobimetinib	Selleckchem	Cat# S8041
Pimasertib	Selleckchem	Cat# S1475
6-Thioguanine	Sigma-Aldrich	Cat#A4882
DMSO	Sigma-Aldrich	Cat#D2650
Ravoxertinib	Cayman Chemical Company	Cat#21107
Temuterkib	Cayman Chemical Company	Cat#27936
Matrigel	Corning	Cat#354234
Nicotinamide [10mM]	MilliporeSigma	Cat#N0636
N-acytlecysteine [1.25mM]	MilliporeSigma	Cat#A9165
Recombinant Human FGF-10 [20 ng/mL]	PeproTech	Cat#100-26
Recombinant Human FGF-basic [1 ng/mL]	PeproTech	Cat#100-18B
PGE2 [1 μM]	R&D Systems	Cat#2296/10
SB202190 [10 μM]	MilliporeSigma	Cat#S7067
Recombinant Mouse EGF [50ng/mL]	ThermoFisher Scientific	Cat#PMG8043
Y-27632 [10μM]	Selleckchem	Cat#S1049
A-83-01 [500nM]	Tocris	Cat#2939
NRG [10ng/mL]	PeproTech	Cat#100-03
Penicillin/Streptomycin	ThermoFisher Scientific	Cat#15140-122
Fetal Bovine Serum (FBS)	Life Technologies	Cat#16000044
Bovine Serum Albumin (BSA)	GeminiBio	Cat#700-100P
Adenosine	Sigma-Alrich	Cat#A9251
Guanosine	Sigma-Alrich	Cat#G6264
Critical commercial assays
Guanine assay kit	Cell Biolabs	Cat#MET-5147
CellTiter-Glo Luminescent Cell Viability	Promega	Cat#G7573
RNeasy mini kit	QIAGEN	Cat#74104
Superscript First-Strang Synthesis III	Invitrogen	Cat#18080051
Deposited data
Multi-omics Datasets	Zenodo	10.5281/zenodo.18471666
Experimental models: Cell lines
K8484	Lab of Kenneth P. Olive	N/A
DLD-1	ATCC	Cat# CCL-221
HCT116	ATCC	Cat# CCL-247
MIA PaCa-2	ATCC	Cat# CRL-1420
PANC-1	ATCC	Cat# CRL-1469
AsPC-1	ATCC	Cat# CRL-1682
A549	ATCC	Cat# CCL-185
MDA-MB-231	ATCC	Cat# HTB-26
RKO	ATCC	Cat# CRL-2577
BxCP-3	ATCC	Cat# CRL-1687
293T	ATCC	Cat# CRL-3216
NCI-H508	ATCC	Cat# CCL-253
H23	Lab of Fred Hirsch	N/A
H647	Lab of Fred Hirsch	N/A
H358	Lab of Fred Hirsch	N/A
Patient C Organoid	Weill Cornell Medicine	N/A
WCM039-01	Weill Cornell Medicine	N/A
WCM715	Weill Cornell Medicine	N/A
WCM773	Weill Cornell Medicine	N/A
WCM389	Weill Cornell Medicine	N/A
WCM601	Weill Cornell Medicine	N/A
WCM526	Weill Cornell Medicine	N/A
WCM236	Weill Cornell Medicine	N/A
Human Foreskin Fibroblast (HFF-1)	ATCC	Cat# SCRC-1041
Experimental models: Organisms/strains
Mouse: Athymic nude mice ((Foxn1nu/Foxn1nu)	Jackson Laboratory	Strain #:007850; RRID:IMSR_JAX:007850
Mouse: C57BL/6J	Jackson Laboratory	Strain #:000664; RRID:IMSR_JAX:000664
Oligonucleotides
MEK1 siRNA: 5’-AAGCAACUCAUGGUUCAUGCUUU-3′ (sense), 5’-AAAGCAUGAACCAUGAGUUGCUU-3′ (antisense)	Ussar et al.30 MilliporeSigma	N/A
MEK2 siRNA: 5’-AAGAAGGAGAGCCUCACAGCA-3′ (sense), 5’-UGCUGUGAGGCUCUCCUUCUU-3′ (antisense)	Ussar et al.30 MilliporeSigma	N/A
MEK Scramble siRNA: 5’-AAGGGUCGUCUAUAGGGAUCGAU-3′ (sense), 5’-AUCGAUCCCUAUAGACGACCCUU-3′ (antisense)	Ussar et al.30 MilliporeSigma	N/A
ERK1 siRNA: 5’-CCCUGACCCGUCUAAUAUAdTdT-3′ (sense), 5’-UAUAUUAGACGGGUCAGGGdAdG-3′ (antisense)	Dimitri et al.32 MilliporeSigma	N/A
ERK2 siRNA: 5’-CAUGGUAGUCACUAACAUAdTdT-3′ (sense), 5’-UAUGUUAGUGACUACCAUGdAdT-3′ (antisense)	Dimitri et al.32 MilliporeSigma	N/A
ERK Scramble siRNA: 5’-CACUCGUAUUCUCAACCGA-3′ (sense), 5’-UCGGUUGAGAAUACGAGUG-3′ (antisense)	Dimitri et al.32 MilliporeSigma	N/A
GART siRNA: 5’-GCUGGAGAAACAAUUGUCA[dT][dT-3′ (sense), 5’-UGACAAUUGUUUCUCCAGC[dT][dT]-3′ (antisense)	Liu et al.46 MilliporeSigma	N/A
GART Scramble siRNA: 5’-GTAAGTAACCCGGAGTATA-3′ (sense), 5’-UAUACUCCGGGUUACUUAC-3′ (antisense)	Liu et al.46 MilliporeSigma	N/A
Software and algorithms
GraphPad Prism 10	Dotmatics	https://www.graphpad.com/
ImageStudio	Licor	https://www.licorbio.com/image-studio
RStudio	R Studio	RStudio 2026.01.0+392
R	R Project	https://www.r-project.org/
Synergyfinder	Zheng et al.47	Synergyfinder.org
MuSyC	Wooten et al.48 Meyer et al.39	N/A
SiCoDEA	SiCoDEA project	(https://sicodea.shinyapps.io/shiny/)
Other
DepMap	Arafeh et al.49	DepMap Public 25Q2 (https://depmap.org/portal)
Kaplan Meier Curves	Proteinatlas Uhlen et al.50	(proteinatlas.org)
Integrated Proteomics Pipeline (IP2) software	Brucker Xu et al.51	N/A

Experimental model and study participant details

Cell culture

K8484 (Mouse, Pancreas) was a gift from Kenneth P. Olive, Columbia University.⁵² Organoids patient C [Human, colon], WCM039-01 [Human, pancreas], WCM715 [Human, colon], WCM773 [Human, pancreas], WCM389 [Human, pancreas], WCM601 [Human, endometrial], WCM526 [Human, clear cell renal carcinoma], and WCM236 [Human, malignant mixed müllerian tumor] were derived by the Englander Institute of Precision Medicine at Weill Cornell under tissue procurement protocols and approved by an institutional review board as previously described.⁴ All WCM organoids were grown as previously described, except Patient C’s organoid were grown with the addition of NRG (neuregulin 1).^4,53,54 Within each organoid line, replicate cultures were distributed across treatment conditions into separate wells for each study. DLD-1 (Human, Colorectal, CCL-221), HCT116 (Human, Colorectal, CCL-247), MIA PaCa-2 (Human, Pancreas, CRL-1420), PANC-1 (Human, Pancreas, CRL-1469), AsPC-1 (Human, Pancreas, CRL-1682), A549 (Human, Lung, CCL-185) and MDA-MB-231 (Human, Breast, HTB-26), RKO (Human, Colorectal, CRL-2577), BxCP-3 (Human, Pancreas, CRL-1687), 293T (Human, Kidney, CRL-3216), Human Foreskin Fibroblast (Human, Foreskin, SCRC-1041) and NCI-H508 (Human, Colon, CCL-253) were bought from the American Type Culture Collection (ATCC). H23 (Human, Lung), H647 (Human, Lung), and H358 (Human, Lung), were gifts from Fred Hirsch, Icahn School of Medicine at Mount Sinai. All 2D cell lines besides AsPC-1 and BxPC-3 (cultured in RPMI [Thermofisher, 11875093]) were grown in DMEM with 10% heat-inactivated fetal bovine serum (FBS, Thermofisher, SH3007903R0) and 100 U/mL penicillin-100 ug/mL streptomycin (Penstrep, Thermofisher, 15140163) at 37°C with 5% CO2. Cell lines were authenticated from ATCC and a mycoplasma test was performed every 2 weeks using the MycoAlert mycoplasma detection test kit (Lonza). This study did not include a clinical trial and was not registered.

In vivo mouse model

Animal studies were conducted following approved Institutional Animal Care and Use Committee.

(IACUC) protocols 2013-0016 and 2018-0018 at the Englander Institute for Precision Medicine at Weill Cornell and the Icahn School of Medicine at Mount Sinai, respectively. Animal studies followed ethical guidelines as outlined by the institutions. Mice were housed in cages under pathogen-free conditions, fed standard food and given water ad libitum on a 12 h light/dark cycle.

For the PDX trial, female athymic nude mice were purchased at 6-8 weeks of age from Jackson Laboratories (Bar Harbour, ME). For each treatment arm, 5 athymic nude mice (n = 5) were injected subcutaneously (100μL volume) around the fourth mammary fat pad with 7.5x10⁴ Patient C’s cells respectively in a 1:1 ratio of DMEM and Matrigel (Corning, #354230). Xenograft take rate was 4 weeks. Tumors were grown until a diameter of 0.6cm was reached before treatment start. Tumor bearing mice with tumors ≤ 0.6cm were excluded from the study. Mice were treated by oral gavage with 100 μL of either 10% DMSO (vehicle control), 1mg/kg trametinib, 1mg/kg thioguanine or both 1mg/kg trametinib and 1mg/kg thioguanine (combination therapy) three times per week with a drug holiday every weekend. This differs from the original PDX dosing for the multi-omics studies that were dosed 5 days a week for a single week.

Similarly, for the K8484 syngeneic allograft studies 100kcells were injected subcutaneous into the mammary fat pad of female BL/6 mice of 6-8weeks of age acquired from Jackson Laboratories (Bar Harbour, ME). For the dose optimization study, n = 5 for each treatment arm and for the regression study n = 10 for each treatment arm. Drug initiation began when tumors had a diameter of 0.6cm in one direction. Vehicle (10% DMSO) and trametinib/thioguanine were given intraperitoneally with doses of trametinib being 1mg/kg and thioguanine being 10mg/kg. Treatment with trametinib/thioguanine began at day 10 for the regression study.

For all studies, all drugs were diluted in DMSO as solvent. Mouse weight and tumor dimensions were measured three times per week. Survival endpoint was met when the mice died, when they met the criteria for euthanasia under the IACUC protocol or when tumors reached a size greater than 1.5cm in one dimension. We acknowledge that female-only mice were used for this study and that sex-based differences might differ between female and male mice including such differences such as the estrous cycle that can affect metabolism and treatment biology.

Tumor volume was calculated using a modified ellipsoid formula given by:

$$V=\frac{4}{3}\pi {\left(\frac{\sqrt{LW}}{2}\right)}^3$$

Where the longest axis was determined to be length, L, and the width, W, 90 degrees perpendicular from L.

Method details

Compounds

Trametinib (GSK1120212), cobimetinib, and pimasertib were purchased from Selleckchem (S2673, S8041 and S1475 respectively). 6-Thioguanine (A4882) and dimethyl sulfoxide(D2650) were purchased from Sigma-Aldrich. Ravoxertinib and temuterkib were purchased from Cayman Chemical Company (21107 and 27936 respectively).

Transcriptomics

Nine athymic nude mice (n = 5 vehicle treated; n = 4 trametinib treated) were injected with patient C’s PDOs and dosed orally with either vehicle (10% DMSO) or trametinib (1 mg/kg) for five days. Tumors were excised and snap frozen. Extraction and sequencing were performed as previously described.⁵⁵ Differential expression analysis was conducted using DESeq2 (v1.42.1), and p-values were adjusted using the Benjamini-Hochberg method. Gene set enrichment analysis (GSEA) was conducted using fgsea (v1.28.0) and Hallmark gene sets loaded using the MSigDB R package (v24.1.0).

Metabolomic analysis

PDX tumors were processed and sent for metabolomic analysis as previously described.⁵⁶ Briefly, tissue lysates of Patient C’s PDX were homogenized on a TissueLyser II (Qiagen) and extracted in 80% methanol/water. Samples were analyzed using a targeted, polar metabolomics approach using hydrophilic interaction chromatography on a Prominence UFLC HPLC system interfaced with a QTRAP 6500 triple-quadrupole mass spectrometer in selected reaction monitoring (SRM) mode with SRM libraries adapted from previously validated libraries.⁵⁷ The quantified metabolite values represent integrated SRM ion counts obtained using MultiQuant (AB SCIEX) and normalized to total protein content of matched lysates. Differential abundances of metabolites were evaluated using edgeR (v4.0.16) and differential expression analysis was conducted using DESeq2 (v1.42.1). P-values were adjusted using the Benjamini-Hochberg method.

Proteomic analysis

Data generated was searched using the ProLuCID algorithm in the Integrated Proteomics Pipeline (IP2) software platform (Integrated Proteomics Applications, Inc., San Diego, CA).⁵¹ Human proteome data was searched using concatenated target/decoy UniProt databases (reviewed entries only; release date: 05-05-2016). Basic searches were performed with the following search parameters: HCD fragmentation method; monoisotopic precursor ions; high resolution mode (3 isotopic peaks); precursor mass range 600-6,000 and initial fragment tolerance at 600 p.p.m.; enzyme cleavage specificity at C-terminal lysine and arginine residues with 3 missed cleavage sites permitted; static modification of +57.02146 on cysteine (carboxyamidomethylation); 4 total differential modification sites per peptide, including oxidized methionine (+15.9949), and phosphorylation (+79.9663) on serine, threonine, and tyrosine; primary scoring type by XCorr and secondary by Z-score; minimum peptide length of six residues with a candidate peptide threshold of 500. A minimum of one peptide per protein and half-tryptic peptide specificity were required. Starting statistics were performed with a Δmass cutoff = 10 p.p.m. with modstat, and trypstat settings. False-discovery rates of protein (pfp) were set to 1% for total proteomic datasets. Label-free quantification was run using the Census label free algorithm on an n = 5 biological replicates across conditions using a summed ion intensity ratio from MS1 scans for all peptides of identified proteins.^58,59,60 A correct linear regression (ln) of +0.85 was applied globally to achieve a normalized Gaussian global distribution of total ion intensity assumed to be equal across conditions. Normalized intensity values for identified representative proteins within the De Novo Purine synthesis pathway were extracted from the Census label free analysis (post global ln regression correction) and presented as a waterfall plot. Proteins deemed statistically significant (ANOVA p < 0.05) and possessing an intensity ratio of < 0.75 (Inhibitor/vehicle) were submitted to Ingenuity Pathway Analysis (QIAGEN Inc., https://digitalinsights.qiagen.com/IPA) for pathway enrichment based on Gene Ontology (GO) terms.^61,62,63 Pathways and proteins containing a p < 0.05 were deemed statistically significant in representation in down regulated pathways within the dataset in response to MEK inhibition. Protein set enrichment analysis (PSEA) was conducted using fgsea (v1.28.0) and Hallmark gene sets loaded using the MSigDB R package (v24.1.0).

High throughput drug screening

As previously described, 3D organoid drug screening was performed using our automated high throughput drug screening platform.^4,36

Organoid drug dose response assay

Organoids (patient C [colon], WCM039-01 [pancreas], WCM715 [colon], WCM773 [pancreas], WCM389 [pancreas], WCM601 [endometrial], WCM526 [clear cell renal carcinoma], WCM236 [malignant mixed müllerian tumor] and Human Foreskin Fibroblast) were plated and dosed in a 7-point dose curve as previously described.⁴ Analysis and dose response curves were computed using a 4-parameter logistic model in GraphPad Prism 10 for Mac OS X (GraphPad Software, San Diego, California USA). Growth metrics (area under the dose response curve [AUC] and the 30% inhibitory concentration [IC30]) were computed using a four-parameter logistic model in GraphPad Prism 10.

Immunoblotting

Cells were lysed using NP-40 lysis buffer (Thermofisher, FNN0021) with added 1X protease inhibitor (Thermofisher, 78442) and pelleted by centrifugation (10 x g for 10min). Protein levels were measured by BCA assay (Pierce, 23227) and 20-30 μg were ran on a precast, 4-12% Bis-Tris gel (Invitrogen, NP0323BOX) for 2.5 h at 120V. Proteins were dry transferred to nitrocellulose membranes using an Invitrogen iBlot2 gel transfer device with a modified protocol of 20V for 1.5min, 23V for 4.5min and 25V for 3min. The membrane was blocked using 5% bovine serum albumin (BSA, GeminiBio, 700-100P) in 1X tris-buffered saline with Tween 20 (TBST) for 30min. Blots were incubated overnight at 4°C in 5% BSA (GeminiBio, 700-100P) with the following antibodies: GART (Abcam, ab169550) 1:500, Phospho-ERK1/2 (Cell Signaling Technology, 9106S) 1:500, ERK1/2 (Cell Signaling Technology, 9102S) 1:1000, DUSP6 (Abcam, ab76310) 1:1000, p-S6 ribosomal protein [p-S6](Cell Signaling Technology, 5364S) 1:1000, S6 Ribosomal Protein (Cell Signaling Technology, 2317S) 1:1000, MEK1 (Cell Signaling Technology, 2352S) 1:1000, MEK2 (Abcam, ab776277) 1:1000, alpha-Tubulin (Cell Signaling Technology, 3873) 1:10000 or β-Actin (Cell Signaling Technology, 3700S) 1:10000. The next day, blots were washed 3 times with 1X TBST for 10min and the membranes were visualized with fluorescently tagged, anti-mouse or anti-rabbit secondary antibodies (680nm or 800nm) from LI-COR (LI-COR, 92668071, 92632212, 92632213, 92668070) at 1:10000 in 5% BSA using a LI-COR Odyssey CLx. All antibodies besides the DUSP6 antibody were visualized using the LI-COR. The DUSP6 antibody was conjugated to a 1:10000 dilution of secondary antibody of HRP conjugated Anti-Rabbit IgG Goat (Cell Signaling Technologie, 7074S) and was visualized using Immobilon Forte Western HRP substrate (Millipore Sigma, WBLUF0500) on a Chemidoc. Multiple independent blots were used as well as probed repeatedly with the antibodies above. Blot images were processed using Image Studio Lite software v5.2.5 (LI-COR). GART Intensity bands were normalized to their lane’s respective actin or tubulin control and then normalized to the DMSO control of each experiment. A log2 fold change was then calculated for each dose of inhibitor or siRNA.

KRAS-mutant cell lines/organoids western blot assay

Cells were seeded in 6 well plates at varying densities depending on cell line. Patient C (200kcells/well), DLD-1 (250kcells/well), HCT116 (250kcells/well), PANC-1 (600kcells/well), ASPC-1 (600kcells/well), K8484 (250kcells/well), MIA PaCA-2 (250kcells/well), MDA-MB-231 (250kcells/well), H23 (250kcells/well), H647 (500kcells/well), H2030 (150kcells/well), NCI-H508 (750kcells/well), BxPC-3 (250kcells/well), RKO (250kcells/well), 293T (250kcells/well) and A549 (150kcells/well). After 24 h, cells were treated with trametinib at a starting concentration of 10 μM down to 10 nM for 48hrs. Control wells were treated with 0.1% DMSO. Once control wells reached 90% confluency they were harvested for immunoblotting.

For the organoids, WCM 389 and WCM715 were plated at 200k/well in a 2:1 matrigel to media ratio into a 6 well plate and left to grow organoids for 5 days. After 5 days, trametinib was added into each well at varying concentrations (1nM to 10 μM) and a DMSO control well was included with 0.1% DMSO. Organoids were dosed for 5 days then collected and harvested as depicted in the Immunoblotting section above.

Guanine colorimetric assay

MIA PaCa-2, A549, HCT116 and DLD-1 cells were plated at 1million cells/well into six, 10cm plates. The next day, trametinib was added with a range of concentrations including 1 μM, 500nM, 250nM, 100nM and 10nM including a 0.01% DMSO control. 48hrs later cells were harvested using RIPA buffer with 1X protease inhibitor. Guanine levels were measured by colorimetric assay using a guanine assay kit (Cell Biolabs, MET-5147) according to the standard protocol with the exception of using a RIPA dilution series for the standard as that’s what the samples were diluted in.

MIA PaCa-2 MEK and ERK inhibitor assay

Cells were seeded in 6 well plates at 150kcells/well. After 24 h, cells were treated with either trametinib, pimasertib, cobimetinib, ravoxertinib or temuterkib at a starting concentration of 10 μM down to 10 nM for 48hrs. Control wells were treated with either 0.1% DMSO or 0.01% DMSO corresponding to the highest dilution of drug (0.1% for 10μM studies and 0.01% for 1μM studies). Once control wells reached 90% confluency they were harvested for immunoblotting.

MEK and ERK knockdown assay in MIA PaCa-2

MIA PaCA-2 cells were seeded in 6 well plates at 200kcells/well. 24hrs later cells were transfected with either 25nM of MEK 1, MEK 2 or both MEK1/2 siRNA using Lipofectamine RNAiMAX (ThermoFisher, 13778030) by following the manufacturer’s instructions (Protocol No. MAN0007825 Rev. 1.0).³⁰ The same was done using 25nM of ERK 1, ERK 2 and ERK 1/2 siRNA for those respective experiments.³² On the next day media was switched to DMEM containing 1% penstrep and 10% FBS and was left on for another 24hrs until the plates were harvested.

MEK1:

5′-AAGCAACUCAUGGUUCAUGCUUU-3′ (sense),

5′-AAAGCAUGAACCAUGAGUUGCUU-3′ (antisense)

MEK2:

5′-AAGAAGGAGAGCCUCACAGCA-3′ (sense),

5′-UGCUGUGAGGCUCUCCUUCUU-3′ (antisense)

MEK Scramble:

5′-AAGGGUCGUCUAUAGGGAUCGAU-3′ (sense),

5′-AUCGAUCCCUAUAGACGACCCUU-3′ (antisense)

ERK1:

5′-CCCUGACCCGUCUAAUAUAdTdT-3′ (sense),

5′-UAUAUUAGACGGGUCAGGGdAdG-3′ (antisense)

ERK2:

5′-CAUGGUAGUCACUAACAUAdTdT-3′ (sense),

5′-UAUGUUAGUGACUACCAUGdAdT-3′ (antisense)

ERK scramble:

5′-CACUCGUAUUCUCAACCGA-3′ (sense),

5′-UCGGUUGAGAAUACGAGUG-3′ (antisense)

GART knockdown with thioguanine dose response assay

MIA PaCa-2 cells, HCT116 and 293T cells were plated at 175kcells/well, 250kcells/well and 250kcells/well respectively in a 6 well plate. 24hrs later cells were dosed with 50 nM of GART siRNA (100 nM siRNA for the 293Ts) or scramble control.⁴⁶ The next day media was switched to DMEM containing 1% penstrep and 10% FBS and was left on for another 24hrs. The following day, cells were trypsinized using 1mL of trypsin (Thermo Scientific, 25300062) and then cell counted. For MIA PaCa-2’s 4kcells/well and for HCT116’s 6kcells/well were then plated into a 96well plate for both siGART and scramble control cells. Roughly 250kcells were taken for each experimental condition for both MIA PaCa-2’s and HCT116s (400kcells were taken for the 293Ts) and plated in a 10cm plate for harvesting to confirm GART was knocked down at endpoint (day 5). 500k cells were immediately lysed using NP-40 buffer to confirm GART was knocked down at day 2. Cells in the 96well plates were dosed on the following day starting at a concentration of 10 μM serially diluted down 1:2 to 39nM and the experiment ended 3 days later when cells were roughly at 95% confluence. Images were taken at time of dosing and at time of endpoint and a fold change was calculated for both scramble control and siGART treated cells. Dose response curves were calculated using a 4-parameter logistic model in GraphPad Prism 10, and AUC and IC30 metrics were computed as described above.

GART:

5′-GCUGGAGAAACAAUUGUCA[dT][dT-3′ (sense),

5′-UGACAAUUGUUUCUCCAGC[dT][dT]-3′ (antisense)

GART Scramble:

5′-GTAAGTAACCCGGAGTATA-3′ (sense), 5′-UAUACUCCGGGUUACUUAC-3′ (antisense)

Grid screens

Grid screens were conducted as previously described.³⁶ Briefly, all 2D cell lines were plated in triplicate in 384 well plates with 1000cells/well. 3D cells (Patient C) were plated in triplicates of 5 μL dots (1,000 cells/μL) in a 1:1 ratio of growth media to Matrigel (Corning, cb40230) within 384 well plates and were left to form organoids for one week before drugging. The next day, cells were dosed in triplicate with either 0.1% DMSO, trametinib, thioguanine, or the combination of trametinib and thioguanine in a 9-point dose response starting with a starting concentration of 10μM down to a final concentration of 1.5 nM using the Echo sampler. Vehicle treated wells were considered to be “0 μM.” 72 hours after drug dosing, cells were treated according to the standard protocol with CellTiter-Glo® assay (Promega, G7573) to determine viability. The viability of each corresponding treated well was averaged and normalized to the average viability of the untreated control well. Heatmaps were made using the ComplexHeatmap v2.10.0 R package.⁶⁴ To calculate AUCs for the dose response curves for each individual single agent and also for trametinib/thioguanine combination wells, the GRID screen cell viability values were plotted in GraphPad Prism 10 and an AUC was calculated for single agent thioguanine, single agent trametinib and combination trametinib and thioguanine depending on the constant dose of trametinib or thioguanine.

Purine supplementation GRID screen

Cells of ASPC-1, K8484 and 293T were plated into 384well plates as triplicates with 4k cells/well and dosed with combination trametinib and thioguanine as described in the Grid Screen section above. Cells in the nucleoside rescue group were also dosed with 1 μM of both adenosine (Sigma-Alrich, A9251) and guanosine (Sigma-Alrich, G6264) on the day of drug dosing. Cell viability was normalized to that of the average viability of the DMSO control treated well.

Calculating the synergy CI values using loewe and MuSyC

Log Loewe Combination index (CI) values were calculated for each well of the 9-point dose response (81wells in total) using the SiCoDEA application (https://sicodea.shinyapps.io/shiny/), which uses the Loewe additivity synergy method to determine synergy.³⁸ The Log CI (natural log) was calculated for each well and was averaged across the plate. An orthogonal approach using the MuSyC synergy framework was deployed to determine synergistic potency and efficacy of the drug combination using each well of the 9-point dose response.^39,48 For the bliss and Loewe synergy maps, SynergyFinder was utilized (synergyfinder.org) to calculate the Bliss and Loewe synergy values across the entire GRID screen.⁴⁷

DepMap/KaplanMeier curve data generation

Correlation analysis of MEK1 and MEK2 protein levels to GART protein level in KRAS-mutant cell lines was performed using DepMap, Broad (2025). DepMap Public 25Q2 and in particular the harmonized_MS_CCLE_Gygi dataset from the Gygi lab (Harmonized Public Proteomics 24Q4). (https://depmap.org/portal).⁴⁹ Kaplan Meier curves of high/low GART mRNA expression in patients was adopted from Human Protein Atlas (proteinatlas.org).⁵⁰ Image credit is from Human Protein Atlas v24.1 with direct links to the GART expression images at https://www.proteinatlas.org/ENSG00000159131-GART/cancer/pancreatic+cancer#PAAD_TCGA and https://www.proteinatlas.org/ENSG00000159131-GART/cancer/lung+cancer#cptac_lung_ac.

RT-PCR

MIA PaCa-2 cells were plated at 200kcells/well and dosed with either DMSO or 50nM trametinib in triplicate for 48hours. Cells were harvested using RIPA buffer with protease inhibitor and RNA was extracted using an RNeasy mini kit (QIAGEN, #74104). RT-PCR was performed as previously described.⁶⁵ Briefly, reverse transcription was carried out using Superscript First-Strang Synthesis III (Invitrogen, 18080051) following Invitrogen’s protocol. cDNA was used as a template for RT-PCR on a ViiA 7 using SYBR Green Mix (Thermo Fisher, S7563) with the following primers:

GART:

Forward: CACCCGGTGTCGGTTTCA

Reverse: TTCCAGGCCAGCGTATGTTC

Ubiquitin:

Forward: CCGGGATTTGGGTCGCAG

Reverse: TCACGAAGATCTGCATTGTCAAG

The extension steps are as follows:

50C 2 min, 1 cycle

95C 10 min, 1 cycle

95C 15 s -> 60C 30 s -> 72C 30 s, 40 cycles

72C 10 min, 1 cycle

Quantification and statistical analysis

Statistical analyses were performed using R (v4.5.2) and GraphPad Prism 10. Differential proteins, transcripts and metabolites was performed via Wald test and corrected using the Benjamini–Hochberg method. Ingenuity pathway analysis was performed using two way ANOVA with p values < 0.5 deemed statistically significant. For comparisons of two groups in the western blot log2 fold change analysis, two-tailed Mann Whitney tests were performed for significance. Likewise, for determining the significance of two groups using Area Under the Curve (AUC), two tailed Mann Whitney tests were used. For mouse survival studies, Mantel-Cox log rank test was used comparing trametinib treated versus trametinib plus thioguanine treated mice with significance determined below <0.05. Any information on statistical details including, n, replicate numbers and test type used, are found in the figure legend and in the results.

Acknowledgments

We would like to acknowledge Fred Hirsch (Icahn School of Medicine at Mount Sinai, New York, NY, United States) for providing us with the cell lines H23, H647 and H358. We gratefully acknowledge the Department of Genetics and Genomic Sciences and the Department of Oncological Sciences at the Icahn School of Medicine, and the Englander Institute for Precision Medicine at Weill Cornell Medicine, for their generous support and access to institutional resources.

This work was supported by CA230384-03 from the NCI/NIH (B.D. Hopkins), the Foundation ISREC (C. Pauli) and Zurich Cancer League (C. Pauli).

Author contributions

L.D.B. and B.D.H. have conceived of and developed the project. S.C., S.N., E.P., C.P., and B.D.H., all contributed to the development and enhancement of the precision medicine pipeline including conceptualization, data curation, investigation, methodology, software, validation, and visualization. B.D.H., O.E., and C.P. provided resources and feedback for the execution of the project. L.D.B., S.N., K.T., and S.S. performed experiments including western blotting, dose responses, and high throughput drug screening. L.D.B. and E.P. helped facilitate the high throughput drug screening and large scale GRID screens of trametinib and thioguanine using the precision medicine pipeline. L.D.B., C.P., and B.D.H. contributed to and helped conduct murine trials. B.D.S. and S.C. helped perform and analyze transcriptomic, proteomic and metabolomic datasets using R. L.D.B. and B.D.H. analyzed datasets and contributed to the writing of the manuscript. S.C. S.N., K.T., B.D.S., E.P., S.S., O.E., and C.P. helped with reviews and revisions of the manuscript. All authors have reviewed and approved the final version of the manuscript for submission.

Declaration of interests

B.D.H. is Co-Founder of and consultant for Faeth Therapeutics. O.E. owns equity in Volastra Therapeutics and Owkin, but this had no relation to this project.

Declaration of generative AI and AI-assisted technologies in the writing process

During the preparation of this work the authors used chatGPT O3 (OpenAI) in order to improve upon grammar and readability of the manuscript. After using this tool/service, the authors reviewed and edited the content as needed and take full responsibility for the content of the publication.

Published: February 7, 2026