Feasibility of a high-throughput small molecule screen of human AML cell responses

We performed an in vitro high-throughput small molecule screen (Fig. 1A) using KMT2A-AML and AMKL leukemia samples (n = 34; Supplementary Table 1) to identify compounds capable of blocking the growth of these cells while exhibiting minimal impact on healthy CD34+ cord blood cells used as controls (n = 3). Samples were screened against 11,142 compounds including commercial inhibitors from APExBIO (Supplementary Table 2, 3) and compounds selected for structural diversity. Based on these tests and prior screens of adult AML samples [20], we tested all molecules at a single dose (1uM for APExBIO, 2uM for all others) for 6 days before cell viability was assessed. Control wells containing cells and DMSO only (64/384-well plate; 2304/sample) served to normalize inhibition values and remove samples with high variability from hit identification (Supplementary Fig. 1).

Fig. 1: High-throughput screening of AML/AMKL leukemias.

A Workflow of the high-throughput chemical screening of 11,142 compounds at a single dose. After incubation, the measurement of luminescence by Cell Titer Glo (RLU gradient of the heatmap) is obtained. This was followed by a secondary dose-response screen to validate identified hits. Selected samples (n = 28) were profiled by RNA-seq. B The heatmap gradient representing the number of total hits. The histogram on the x-axis corresponds to the number of hit per samples. The histogram on the y-axis corresponds to the number of hit per plates. Plates with proprietary compounds are shown in light green, commercial inhibitors from APExBIO are colour in orange and commercial diversity compounds in dark green. C Heatmap of selected compounds show heterogeneous response between sample types. The C-score (corrected % inhibition) is annotated by the color gradient. D Dose response curves of proprietary compound UMxxxx808 showing responses of CB CD34+ cells (red lines) and AML patients’ cells (blue lines). The curves were obtained by a Bayesian inference method where the solid line depicts the median curve with a shaded 95% C.I. E Comparison of AUC from hits identified in common with Beat AML. 95 compounds in common between Beat AML and the commercial inhibitors with AUC values split between hit and non-hit. The difference between non-hit vs hit is calculated by a t-test independent with Bonferroni correction (P = 7.294 × 10-13 stat = −7.176).

The results showed a high level of heterogeneity in the number of hits depending on the sample and compound source (Fig. 1B-C). The impact of patient genetic heterogeneity on drug response has been previously described [20] and was not unanticipated, even with our biased use of KMT2A-AML samples (Supplementary Fig. 2). The screen showed that many current therapeutics were toxic to all samples at the dose tested (1 µM; Supplementary Fig. 3) reflecting their known lack of selectivity, however, numerous previously identified anti-leukemic compounds and novel proprietary molecules displayed similar activity (Supplementary Fig. 3). Both commercial and proprietary diversity collections exhibited similar hit rates (~0.99%) whereas the 1960 commercial (APExBIO) inhibitors and tool compounds showed a higher hit rate (~1.84%). Only 140 compounds showed 50% or greater inhibition across all samples (41 compounds at 97%) of which 89 were from the APExBIO library (Supplementary Table 4). In combination with a false-positive rate of 0.08% (58 hits in 69,120 DMSO control wells across all samples), our data suggest that any differential activity of compounds between normal and leukemic cells should have been detected. Indeed, a large number of response profiles of compounds across a range of cell samples demonstrated clear patterns of inhibition (Fig. 1C) despite the heterogeneity of the samples tested, and these hits were selected for further investigation.

To validate our single-dose screen, we performed dose-response analysis using a different set of 40 leukemia samples from various sources (Supplementary Table 1) along with control HSPCs. For this, 133 hits from the primary screen with inhibition profiles showing the greatest specific anti-leukemic effects were re-tested in 10-point dose-response experiments (1.5 nM – 10uM). Within this group, 121 (93%) compounds displayed a sigmoidal dose-response relationship for at least two of the samples tested and revealed the same type of heterogeneity seen in the primary screen (Supplementary Fig. 4) with numerous compounds segregating responsive and non-responsive samples (Fig. 1D, blue curves, 3 log difference relative to CB CD34+ cells, red curves). Of the novel compounds validated by dose response, 12 were judged to be of high interest for further development and target identification efforts based on the extent of their anti-leukemic activity. Although detailed studies of the mechanisms of these molecules are beyond the scope of this report, given the KMT2A translocated samples included in the screen, we performed synergy tests with the Menin inhibitor revumenib (SNDX-5613) in combination with 2 of our novel molecules or navitoclax. These tests revealed no synergy between any of the combinations (Supplementary Fig. 5) however this is likely due to the fact that revumenib alone displayed little activity in short-term viability assays, an observation recently independently published [21]. We next compared our data with that of the Beat AML study in which 122 small molecules were tested against 409 AML primary patient samples [12]. Of the 119 commercial inhibitors tested in both Beat AML and our study (Supplementary Table 5), 95 had useful area under the curve (AUC) values that reflect activity in the Beat AML study (Fig. 1E). This comparison shows that our hit compounds had significantly higher AUC values than our non-hit compounds (p = 7.3 × 10−13), strongly supporting the hypothesis that our screen replicates the findings of the Beat AML study.

De novo model and patient cell responses are more similar than KMT2A cell lines

We compared the therapeutic vulnerabilities of cells from 11 primary AML patients, 6 human AML cell lines, and 10 separately generated human de novo AML models. The distribution of the ~800 total hits identified was broadly similar for each of the 3 sample types tested (Fig. 2A). Hierarchical clustering was then used to assess similarities in responses across the different sample types. This showed that the AML patient samples clustered closely with the human de novo AML models despite the greater level of genetic heterogeneity of the former (Fig. 2B) and the cell line samples segregated as a distinct group from the primary leukemias and human models with the exception of the OCI-AML-8227 cell line used as a model of an AML hierarchy [22].

Fig. 2: Differences in drug discovery using patient, model, and cell line AML cells.

A Number of hits per sample type where colouring represents: leukemia cell lines (red), human model leukemias (pink), patient leukemia samples (blue), with the number of samples analyzed shown in brackets. B Hierarchical clustering heatmap (Ward linkage) is shown based on the C-score values for all compounds that were hits across all samples. The heatmap is annotated along the top two rows with a colour code for both samples types and leukemia types. C Bar graphs are shown that depict the number of hits found to be unique to each sample types or hits in common between sample types. Sample types and the number analysed are shown along the x-axis with the number of hits indicated at the top of each bar graph. D Line graph showing the number of hits unique to each sample type but common to at least 2, 3, or 4 samples of that type. E Shared hits between patients and models (left) and patients and cell lines (right). The y-axis represents the number of AML patient samples and the x-axis is the number of AML human models (left) and AML cell lines (right). The number of common hits between each set of samples (e.g. between 3 models and 3 patients) is indicated by circles representing different numeric ranges. F Coloured bar graph segments indicate the source of unique hits across different sample types shown in C. G Columns show the unique hits from the APExBIO compound library identified for each sample class labelled on top. Colour coding legend on the right indicates the target/pathway affected by each individual compound based on the manual annotation of published activities.

Because of genetic differences within the AML and AMKL leukemia groups, we compared only the samples with KMT2A rearrangements including five cell lines, seven de novo models, and eight patient leukemias to perform a comparison of overlapping hits (Fig. 2C). This showed that the number of unique hits was largely similar, with 150–200 hits in each sample class, but the human models again showed 2-fold higher shared hits with the patients’ cells than with the cell lines. Analysis of all shared positive hits showed 147 compounds were unique for the primary patient samples tested (Fig. 2D) with only 41 hits (~28%) shared with more than one patient, and only four hits (3%) shared with three or more patients. A similar pattern was observed for compounds uniquely affecting cell lines, with a ~50% decrease in the number of shared compounds for each additional cell line. In contrast, hits unique to the de novo model AML cells showed a higher level of similarity, with 25% of hits shared between at least four samples. Because model AML generation requires only a KMT2A fusion [19], these results likely reflect the underlying genetic heterogeneity of the samples in each class which impacts compound responses.

To better understand the differences in the unique hits, hits for each combination of sample pairs for either patients’ cells and models or patient leukemia and cell lines were identified and compared (Fig. 2E). The data was grouped according to shared hits for each combination of sample pairs for either patients and models or patients and cell lines (Fig. 2E). This analysis showed that most of the hits shared between patient and cell line samples were shared between just one patient and multiple cell lines or between just one cell line and multiple patient samples. In contrast, most of the hits that were shared between the patient and human model cell samples were correlated with larger numbers of each respective sample type (e.g. along the diagonal). These findings suggest that the similarity between models and patients is not the result of an individual sample but rather a broader similarity in their responses. This hypothesis is further supported by similarities in the sources of hit compounds for each response category (Fig. 2F). Specifically, nearly half of all hits unique to cell lines were from the APExBIO collection, whereas this source accounted for <15% of the hits for patient and model samples.

Manual curation of the biological pathways and targets of the unique hits affecting each type of sample further demonstrated a profound difference between the cell lines and other sample classes. For example, the KMT2A-AML cell lines had a large number of unique hits in a range of signalling pathways, that were not detected in the model or patient leukemias, or the combination (Fig. 2G). Conversely, the latter two shared responses to classes of compounds that did not affect the AML cell lines (Supplementary Fig. 6). Because the primary and human model AMLs were grown in media that differed from AML cell lines (i.e. serum-free media with growth factors) we tested whether this could explain the biased responses exhibited by the cell lines. To evaluate this we selected 104 compounds that were unique hits in 1-3 AML cell lines and retested them in both media types. As expected, all cell lines showed poor growth in the serum-free medium and ~20% of the compounds lost their inhibitory effects. However, 60% of the hits remained shared between the cell lines (Supplementary Fig. 7) and importantly, these compounds did not show any change in the targeted biological pathways. Together, these data show that cells from the human KMT2A-AML de novo models displayed greater similarity to primary patient AMLs than the established AML cell lines tested, despite their lower level of concurrent genetic abnormalities. Critically, they also show that even under optimal growth conditions, large-scale in vitro screening of leukemia cell lines identifies a biased selection of compounds that lack activity against many primary patient cells.

Use of drug response data to identify specific targets

To investigate whether gene expression data of the tested samples could be used to understand sensitivity to the compounds, we performed RNA-seq on the screened samples and then globally correlated these data with the drug response levels. Pearson correlations between the inhibition values of the 840 hits identified and the expression levels of the 500 most variable genes were computed (Fig. 3A). Hierarchical clustering revealed large groups with similar expression signatures across a range of different compounds, both correlated and anti-correlated with the compound activity. To gain more insight into these relationships, we identified groups of compounds along the y-axis that showed similar strong correlation patterns (Fig. 3A, purple/red boxes) and then performed gene expression enrichment analysis (GSEA) [23] of the genes within these groups. Interestingly, transcripts identified as a signature of KMT2A-rearranged leukemia showed a strong correlation with the activity of some compounds (red box), although 7 out of 30 samples lacked KMT2A translocations (Fig. 3B). In contrast, other leukemia-relevant signatures were anti-correlated with a separate group of drugs (purple box). Similarly, genes involved in protein translation were either strongly correlated or anti-correlated with the activities of the specific sets of drugs examined.

Fig. 3: Correlation screen response and expression data.

A Pearson correlation between C-score hits and FPKM values of the 500 most variable genes across all samples. Clustering on compounds (row) and gene (columns) is represented by Ward linkage and correlations are coloured from negatively (blue) to positively correlated (red). Two clusters of compounds are identified by a purple and red box were used for GSEA analysis. B The average C-score values for all 500 genes for the compounds in the coloured boxes in A were calculated and used to generate a ranked list for GSEA analysis. Selected pathways with statistical enrichment highlighted by GSEA are shown. C Scatterplot showing RG2833 C-score values plotted against NPM1 expression (FPKM) with a linear regression model fit with the 95% CI also shown. D A volcano plot from an acetylome analysis of NOMO1 cells treated for 6 h with RG2833 showing peptides which are significantly decreased (green) or increased (red) in acetylation. E boxplots show the IC50 values for RG2833 and TC-H106 for primary AML specimens with either a wt NPM1 gene (NPM1 wt, n = 5) or a tetranucleotide insertion in exon 12 (NPM1 mut, n = 5). All AML samples had blast percentages >90% except two NPM1 mutated samples (50% & 60%) that also exhibited higher IC50 values F Scatterplot showing Sabutoclax C-scores plotted against SRFS5 expression (FPKM) with a linear regression model fit with the 95% CI shown. G Sabutoclax activity versus BCL2/MCL1 expression and isoform abundance. All leukemia samples were binned into Sabutoclax responsive samples (red; C-score >median response across all samples) and unresponsive samples (blue; C-score < median response across all samples) for top row and bottom right panels. Bottom left panel shows all samples binned based on SRSF5 expression (red or blue for < median or > median across all samples respectively). The panels in the first column show the expression level of BCL2 family members and the panels in the second column, show the expression level of BCL2 (top) or MCL1 (bottom) isoforms. *=0.05 where P-values were calculated using students t-test.

Given the relevant global correlations observed, we used a similar approach to calculate the most positively or negatively correlated compounds on a per-gene basis. Two of the resulting individual gene correlations, which were potentially relevant in the context of known leukemia biology, were further investigated (Fig. 3C). First, expression of the nucleophosmin gene (NPM1) was strongly inversely correlated with several histone deacetylase inhibitors (HDACi) including RG2833 along with TC-H106, CI994, and Chidamide (Supplementary Fig. 8). NPM1 is involved in a wide range of cellular functions [24] and is frequently mutated in adult AML, where the most common mutation consists of a 4-bp insertion into exon 12, resulting in the relocalization of NPM1 to the cytoplasm from the nucleolus [25]. NPM1 is, however, rarely mutated in KMT2Ar AML [4], although acetylation of NPM1 is required for its function [26]. Mutations in numerous lysine residues in NPM1, which are targets for acetylation, do not alter its subcellular localization [27]. The mechanisms of action of HDAC inhibitors (HDACi) are extremely broad [28] and while mutated NPM1, but not the wild-type form, has been described as a target of a different HDACi (panobinostat) [29], the mechanism remains unclear. To investigate whether NPM1 expression was associated with sensitivity to HDAC inhibition in the present study, we examined the degree to which NPM1 expression levels were correlated with responses to all histone deacetylases and the inhibition levels of six HDACi (Supplementary Fig. 8). Only HDAC2 was strongly anti-correlated with all of the HDACi tested except KD5170, similar to NPM1, and the expression of HDAC2 and NPM1 also showed the strongest correlation across all samples (Supplementary Fig. 8). Interestingly, a recent study demonstrated that a specific lncRNA can mediate the interactions between HDAC2 and NPM1 in the nucleolus [30], suggesting a functional connection between these two proteins. To directly assess the effects of HDACi treatment in AML, acetylome analysis was performed on NOMO1, an AML cell line sensitive to HDACi, after a 6 hour exposure to the HDACi RG2833. The results showed significant changes in acetylation of 229 peptides from 181 proteins, including an increase in NPM1 acetylation (Supplementary Tables 7 and 8). Analysis of proteins with acetylation changes (Supplementary Fig. 9) showed enrichment in functions consistent with HDACi activity [31, 32]. Taken together with published observations, these findings suggest that HDACi treatment disrupts the activity of NPM1 in the nucleus without altering its localization and that AML sensitivity to HDACi is directly related to NPM1 protein levels in the nucleus. The observation that NPM1-mutated AML samples have been reported to be more sensitive to HDACi [29], (also seen here; Fig. 3E) could be explained by a ~ 50% reduction of NPM1 in the nucleus caused by cytoplasmic relocalization.

We next examined the strong negative correlation observed between the compound sabutoclax, a pan-BCL2 inhibitor, and the Serine and Arginine Rich Splicing Factor 5 (SRSF5) gene. Apoptosis is regulated by the BCL2 (BCL2 Apoptosis Regulator) family of proteins including BCL2, BCL-XL, and MCL1, where many members have both pro- and anti-apoptotic isoforms. SRSF5 has been described as an oncogenic splicing factor, and its loss of expression leads to an increase in the pro-apoptotic short MCL1 isoform [33]. Examination of the expression of all BCL2 family members relative to both SRSF5 expression and sabutoclax activity identified 10 BCL2 family members were expressed in the AML cells tested (log(FPKM) + 1 > 0), (Fig. 3G). However, only BCL2 expression was significantly lower in samples with higher sabutoclax activity and only MCL1 expression positively correlated with SRSF5 expression. In addition, the sabutoclax-sensitive cells showed reduced expression of all BCL2 isoforms including the shorter BCL2-203 isoform whose function remains poorly defined. In contrast, despite a higher level of expression of the anti-apoptotic MCL1 isoform (MCL1-202) in samples with elevated SRSF5 expression, this gene was expressed at lower levels in samples sensitive to sabutoclax and was associated with an increase in the non-coding MCL1 isoform (MCL1-204). Taken together, these data suggest that elevated SRSF5 expression alters MCL1 isoforms that, in combination with differences in BCL2 expression, increase sensitivity to proapoptotic compounds.

Inducers of apoptosis reveal AML subtype-specific vulnerabilities

Given the clinical relevance of pro-apoptotic compounds in recent years as well as our observed correlation of expression and splicing of BCL2-family members with respect to sabutoclax sensitivity, we also looked for potential differences in apoptosis regulation, particularly between the different subtypes of AML included in our screen. Among 28 compounds that were defined as hits against AMKL (≥75% inhibition against AMKL and ≤30% inhibition against normal CB CD34+ cells, Fig. 4A), we identified a group of seven compounds that were classified as inducers of apoptosis. We previously showed that AML of megakaryocytic lineage (AMKL) are sensitive to inhibition of pro-survival factor BCL-XL [34]. In this large-scale and unbiased screen, we find that AMKL cells are sensitive to BCL-XL inhibitors navitoclax (targeting BCL2, BCL-XL, BCL-W) and A-1155463 (BCL-XL), but resistant to BCL-2 inhibitor venetoclax at 1 µM, pointing to BCL-XL as a dependency in this subgroup of leukemia (Fig. 4B) as previously reported [34]. Sensitivity to pro-survival factor inhibitors, measured as IC50 values, was compared between 9 AMKL and 22 AML subtypes according to cell type (FAB classification), Fig. 4B. Analyses confirmed that AMKL and monocytic AML (FAB M5A/B subtypes) are rather resistant to BCL-2 inhibition with venetoclax, with IC50 values well above 1uM for AMKL subtypes. In contrast, AMKL are sensitive to BCL-XL inhibition with either navitoclax or A-1155463 (Fig. 4B). Of note, some non-megakaryocytic AML are also sensitive to BCL-XL inhibitors, such as NUP98::NSD1 and a subset of KMT2Ar AML. AML cells (n = 9) were found to be more sensitive to MCL1 inhibitor S63845 (lower IC50 values as compared to AMKL (n = 9), Fig. 4C).

Fig. 4: Subtype-specific sensitivities of AML cells to inducers of apoptosis.

A An Overview of selected compounds that fulfilled threshold criteria for CB-CD34+ and were considered a hit for at least one AMKL sample. Drugs were grouped by pathway. Drug targets or modes of action are unknown for proprietary compounds IRIC. B IC50 values were inferred from dose-response experiments for venetoclax, navitoclax and A-1155463 with three independent CB-CD34+ pools together with nine models of AMKL as well as 22 models of AML. Samples were grouped by FAB classification where available. C IC50 values were inferred from dose-response experiments for S63845 with three independent CB-CD34+ pools together with nine models of AMKL as well as eight models of AML. D Schematic overview of experimental design of combinatory treatments with navitoclax and cytarabine (AraC) of mice that were transplanted with a pdx of NNSD1. E Percentage of leukemic blasts in peripheral blood (%hCD45+) before (left graph) or after completion of drug treatment (right). F Representative flow cytometry profiles of spleens of vehicle-treated mice versus mice treated with the combination of navitoclax and cytarabine (Navito+AraC). G Spleen weight, H infiltration of bone marrow (BM) and I spleen was assessed in transplanted mice after 3 weeks of indicated treatments. J Gene expression of AML and AMKL samples that were treated with selected compounds was correlated with drug response (IC50 values). K Bliss synergy scores of indicated drug combinations in AMKL and AML xenografts, one synthetic model with a KMT2A::AFDN fusion (MLLT4), as well as in cell lines (ML-2, KMT2A::AFDN fusion; SHI1, KMT2A::AFDN fusion; THP1, KMT2A::MLLT3 fusion). The synergy scores were calculated using the BLISS reference model in SynergyFinder. Values are the average of two experiments for synthetic models and PDX and n = 1 for cell lines. [60]. P-values: * <0.05, ** <0.005, *** <0.001, **** <0.0001. CB-CD34+, CD34+ cord blood; AMKL, acute megakaryoblastic leukemia; AML, acute myeloid leukemia; NNSD1, NUP98::NSD1; AraC, cytarabine; CG2, CBFA2T3::GLIS2; NTF, NUP98::BPTF; N5A, NUP98::KDM5A.

To exploit the potential vulnerability of some non-AMKL leukemia to BCL-XL inhibition, we tested the in vivo anti-leukemic activity of navitoclax against a NUP98::NSD1 PDX-AML model as a single agent or in combination with standard-of-care cytarabine (Fig. 4D). Engraftment of AML cells was detected in peripheral blood at similar levels for all recipient mice prior to drug treatment (Fig. 4E, left panel). At sacrifice after a 3-week treatment course, leukemic cell infiltration was decreased at a similar level between cytarabine and navitoclax-treated groups, and further reduced in the combination treatment arm (Fig. 4E, right panel). The combination of cytarabine and navitoclax was most effective at reducing leukemic burden of the spleen and bone marrow (Fig. 4F-I). Correlative gene expression studies uncovered that combined high BCL-XL (BCL2L1) and low MCL1 RNA levels predicted sensitivity to navitoclax (Fig. 4J). Of clinical relevance, in vitro drug synergy studies showed that combined pro-survival factor dependencies are also AML subtype-specific (Fig. 4K). For instance, BCL-XL and MCL1 inhibition was most effective against GLIS2r AMKL (CG2, BLISS score of 38), whereas BCL-XL and BCL-2 inhibition was synergistic for CG2 AMKL, for a NUP98::NSD1 (NSD1) PDX model as validated in vivo, as well as the ML2 and THP1 cell lines (with KMT2A fusions) [35,36,37]. Overall, these results underscore the complexity of the cell- and regulation of pro-survival factors and highlight combinations of BH3 mimetics could overcome resistance to monotherapy and represent an improved option for the treatment of a broad range of AML subtypes.

Complex polypharmacology complicates the definition of mechanisms of action of drugs

To test the possibility of repurposing clinically accepted drugs, we conducted dose-response studies of shikonin, which was identified as a potent inhibitor in the primary screen. This natural naphthoquinone (Fig. 5A) has been reported to inhibit inosine-5’-monophosphate dehydrogenase 2 (IMPDH2) in breast cancer [38] and pyruvate kinase 2 (PKM2) in various cancers [39, 40]. The inhibitory effect on purified IMPDH enzymes relative to two known IMPDH inhibitors, MPA and MFF, did not reveal the previously reported selectivity between IMPDH1/2 (Fig. 5A) [38] although MPA showed a more potent inhibition of IMPDH2 (Fig. 5B). The effects of shikonin on the viability of AML cell lines and patient samples showed a significantly better anti-leukemic effect than MPA, with a limited effect on the growth of CD34+ CB cells (Fig. 5C). Cell cycle analysis confirmed that AML cells treated with shikonin were blocked in G1 phase and showed increased levels of apoptosis (Supplementary Fig. 9). To investigate whether an activity against PKM2 might simultaneously contribute to the higher cell-based anti-leukemic activity of shikonin, we compared it to a published PKM2 inhibitor (PKM2-IN-1 [41]) in dose-response assays (Fig. 5D) and pyruvate kinase assays (Fig. 5E). Remarkably, shikonin effectively impaired pyruvate kinase activity in AML cell lines as well as PKM2-IN-1, but remained less toxic to healthy CB CD34+ cells than PKM2-IN-1 (Fig. 5F). RNA-seq was performed on two AML cell lines (SHI1 and THP1) and CB CD34+ cells treated with shikonin for a short period (16 h). Interestingly, both cell lines showed a rapid and profound downregulation of MYB, along with other potentially relevant genes (BCL2, IMPDH2, GFI1, and MYC) (Fig. 5G) with virtually no change in CB CD34+ cells (Supplementary Fig. 11). MYB is an essential gene in AML [42, 43], and its transcription is directly regulated by oncogenic KMT2A fusions. However, the molecular mechanisms underlying this rapidly downregulated response to a metabolic inhibitor remain unclear. GSEA of downregulated genes highlighted both E2F and MYC as likely target genes in both cell lines, but with significant MYC downregulation only in SHI-1 cells (Supplementary Fig. 12). Shikonin has been implicated in impaired mitochondrial function in AML cells, leading to an increased reliance on glycolysis in AML cells [44, 45], mediated by direct binding of shikonin to the mitochondrial electron transport chain complex II (ETC-CII) [44, 46]. These observations are concordant with our own metabolic profiling of AML cells after shikonin treatment showing increased glucose uptake and lactate secretion (Supplementary Fig. 13).

Fig. 5: The complex polypharmacology of shikonin.

A Molecular structures of two known IMPDH inhibitors, Mycophenolic acid (MPA), Mycophenolate mofetil (MFF) along with Shikonin. B Dose-response curves obtained for IMPDH1 (blue) and IMPDH2 (red) using purified enzymes and the three inhibitors in A are shown along with corresponding IC50 values for each enzyme. C Cell-based dose-response assays for MPA (top panels) and shikonin (bottom panels) were performed using AML cell lines (left panels) or AML patient samples (right panels). For both cell lines and patient samples, those containing either KMT2A::MLLT3 (KM3) or KMT2A::MLLT4 (KM4) fusions are coloured in orange/green respectively. Cord blood CD34+ cells were used as controls to show selectivity and are shown in red. D Molecular structure of PKM2-in-1 inhibitor (top) and cell-based dose-response assays (bottom) are shown with colouring as in C, using cell lines (top panel) and AML patient samples (bottom panel). E The inhibition of pyruvate kinase activity was assessed in a cell-based assay where two AML cell lines were treated with either inhibitor (shikonin or PKM2-in-1) or DMSO (control) prior to cell lysis and pyruvate kinase assay measurements performed in triplicate at the indicated time points. **** indicates a p-value of 0.0001 for difference between curves for treatment vs controls at 36 h (3rd time point). F A summary of IC50 values for MPA, shikonin and PKM2-in-1 inhibitors from all cell-based dose-response assays are shown. Identical colouring of dots indicates the IC50 values from replicates of each sample tested. G Volcano plots of differentially expressed genes as determined by DESeq2 from RNA-seq experiments performed on THP-1 cells (left) or SHI-1 cells (right) after 16 h of exposure to shikonin are shown.