Patient cohort: Overview.
The cohort consisted of 6,043 adults with a documented or suspected myeloid neoplasm and at least one mutation identified by FoundationOne®Heme testing (Table 1). To investigate for sampling bias, we assessed the mutation profiles in individual genes in our four diagnosis categories. Logistic regression analysis confirmed significant genetic differences between AML, MDS, MPN, and MDS/MPN categories typical of each myeloid phenotype,23 supporting separation into these four groups for study purposes (Supplementary Fig. 1).
Table 1
Feature | AML | MDS | MDS/MPN | MPN | Total |
Total patients | 2,214 (37%) | 2,413 (40%) | 363 (6.0%) | 1,053 (17%) | 6,043 |
ASXL1mut | 408 (18%) | 530 (22%) | 193 (53%) | 283 (27%) | 1,414 (23%) |
c.1934dupG (p.G646Wfs) | 160 (39%) | 172 (32%) | 82 (42%) | 106 (37%) | 520 (37%) |
Others | 248 (61%) | 358 (68%) | 111 (58%) | 177 (63%) | 894 (63%) |
c.1900_1922del (p.E635Rfs) | 60 (15%) | 77 (15%) | 23 (12%) | 31 (11%) | 191 (14%) |
c.2077C > T (p.R693*) | 21 (5.1%) | 26 (4.9%) | 6 (3.1%) | 13 (4.6%) | 66 (4.7%) |
Cohesin-binding motif | 23 (5.6%) | 34 (6.4%) | 15 (5.3%) | 11 (5.7%) | 83 (5.9%) |
Median age (IQR) | 64 (51–73) | 71 (63–79) | 71 (64–78) | 67 (55–75) | 68 (58–76) |
ASXL1wt | 62 (48–72) | 71 (61–78) | 71 (63–78) | 64 (52–73) | 67 (55–75) |
ASXL1mut | 70 (63–77) | 74 (66–79) | 71 (65-77.5) | 71 (65–78) | 72 (65–78) |
P value (ASXL1mut vs. ASXL1wt) | < 0.0001 | < 0.0001 | 0.3995 | < 0.0001 | < 0.0001 |
Male:female | 1.34:1 | 1.52:1 | 1.81:1 | 1.18:1 | 1.40:1 |
ASXL1wt | 1.21:1 | 1.36:1 | 1.66:1 | 0.97:1 | 1.23:1 |
ASXL1mut | 2.19:1 | 2.35:1 | 1.97:1 | 2.08:1 | 2.19:1 |
P value (ASXL1mut vs. ASXL1wt) | < 0.0001 | < 0.0001 | 0.4437 | < 0.0001 | < 0.0001 |
European:non-European ancestry | 3.08:1 | 4.02:1 | 5.26:1 | 3.26:1 | 3.54:1 |
ASXL1wt | 2.97:1 | 3.82:1 | 5.54:1 | 3.03:1 | 3.36:1 |
ASXL1mut | 3.64:1 | 4.89:1 | 5.03:1 | 4.05:1 | 4.32:1 |
P value (ASXL1mut vs. ASXL1wt) | 0.1268 | 0.0564 | 0.7754 | 0.1007 | 0.0009 |
Mutations (median, IQR) | 4 (2–5) | 2 (1–4) | 4 (3–5) | 2 (1–4) | 3 (2–4) |
ASXL1wt | 3 (2–5) | 2 (1–3) | 3 (2–4) | 2 (1–3) | 2 (1–4) |
ASXL1mut † | 4 (3–5) | 3 (2–4) | 4 (2–5) | 3 (2–4) | 3 (2–5) |
P value (ASXL1mut vs. ASXL1wt) | < 0.0001 | < 0.0001 | 0.0003 | < 0.0001 | < 0.0001 |
† Excludes ASXL1 mutations. |
AML: acute myeloid leukemia; MDS: myelodysplastic syndrome; MPN: myeloproliferative neoplasm; IQR: interquartile range |
We also directly compared mutation frequencies in AML patients (n = 2,214) to a large published AML comparison cohort (n = 1,540).9 Patients in the study cohort were generally older, had more mutations in genes associated with secondary8 and adverse risk AML,24 and fewer NPM1, FLT3, and CEBPA mutations than comparison AML patients (Supplementary Fig. 2). We noted a more than two-fold higher frequency of TP53 mutations (17%) than reported in previous large MDS cohorts25,26 and a large publicly available database23 (~ 7–8%). These results supported a sampling bias towards patients with higher-risk genetics.
Males were more represented than females (1.4:1), and most patients were of European ancestry (3.5:1). MDS (40%) and AML (37%) were the most common diagnoses. Collectively, ASXL1 mutations occurred in 1,414 patients—the most of any gene on the test panel (Supplementary Table 3)—and occurred in 18% of AML and 26% of chronic myeloid neoplasms. All ASXL1 mutations were nonsense, frameshift, truncating insertions/deletions, or splice site variants (Supplementary Spreadsheet S1), and most occurred in the final exon (Fig. 1A). Twenty-eight (2.0%) patients had multiple ASXL1 mutations, and ASXL1 was the sole mutated gene in only 52 patients (3.7%). The most common ASXL1 mutation was c.1934dupG (p.G646Xfs*12), and this was the sole or dominant ASXL1 mutation in 520 cases (collectively referred to as ASXL1c.1934dupG hereafter). The remaining 894 patients had one or more mutations at other sites in the ASXL1 gene (ASXL1other).
Myeloid neoplasms with ASXL1c.1934dupG versus ASXL1other.
There were no significant differences in age, sex, or genomic ancestry between ASXL1c.1934dupG and ASXL1other (Supplementary Table 4). We noted slightly fewer ASXL1c.1934dupG mutations in patients with MDS (ASXL1c.1934dupG:ASXL1other 0.48:1) compared to AML (0.65:1, p = 0.03) and MPN (0.60:1, p = 0.01) and those in whom ASXL1 was the sole mutation (Supplementary Fig. 3A). However, these trends may be due to our VAF-based reporting thresholds, as ASXL1 VAFs were lower in singly mutated patients and those with MDS (Supplementary Fig. 3B).
Across all patients (n = 1,414), STAG2 mutations were more likely to be seen with ASXL1c.1934dupG (21% vs. 16%, p = 0.02), whereas SETBP1 mutations were more commonly co-mutated with ASXL1other (15% vs. 10%, p = 0.01). In general, the absolute differences in co-mutation frequencies between ASXL1c.1934dupG versus ASXL1other were small (see Supplementary Table 5), but more prominent differences emerged within phenotypic subsets (Fig. 1B). For instance, KMT2A rearrangements and STAG2 mutations were strongly associated with ASXL1c.1934dupG in MDS/MPN and MPN, with ASXL1c.1934dupG:ASXL1other ratios of 5:1 (p = 0.03) and 9:1 (p < 0.01), respectively.
In contrast, two genes in AML, TP53 and SETBP1, had a significantly higher co-mutation rate with ASXL1other (TP53: 11% vs. 3% in ASXL1c.1934dupG, p < 0.01); SETBP1: 14% vs. 7%, p = 0.04). As such, we sought to determine whether other specific ASXL1 mutations were associated with TP53 or SETBP1 co-mutation in AML (Supplementary Fig. 4). A single mutation, ASXL1 p.R693*, was significantly more frequent in cases of ASXL1mutTP53mut AML (22% vs. 3.5% of ASXL1mutTP53wt AML, p < 0.01). Similarly, the ASXL1 p.R404* variant was more often seen in ASXL1mutSETBP1mut AML (7.3% vs. 0.05% of ASXL1mutSETBP1wt AML, p = 0.01). These collective results confirmed that the c.1934dupG variant occurs in a similar patient population to other ASXL1 variants. However, subset analysis supports that ASXL1c.1934dupG and ASXL1other may be associated with certain phenotypic and co-mutational tendencies.
The ASXL1 cohesin binding motif and its relationship to cohesin gene mutations.
As our analysis of ASXL1c.1934dupG patients supported that ASXL1 mutation site may play a role in co-mutation profiles, we also examined mutations in another functional region of the ASXL1 gene. Recent in vitro analysis has demonstrated that the ASXL1 protein directly interacts with core cohesin proteins between amino acids 401 and 587,6 a region denoted the cohesin-binding motif (CBM).27 As such, we sought to characterize ASXL1 CBM mutations to investigate for potential relationships to cohesin gene mutations.
The vast majority of cohesin gene mutations in ASXL1mut patients were in STAG2 (n = 235), with few mutations in RAD21 (n = 9) and SMC1A (n = 4); four cases had multiple cohesin gene mutations (Supplementary Table 6). The proportion of CBM mutations was lower in ASXL1/cohesin gene co-mutated cases (n = 7/244, 2.9%) than in ASXL1-mutated patients overall (n = 83/1,414, 5.9%), though this difference did not quite meet statistical significance (p = 0.06, see Table 1). The few CBM/cohesin co-mutated cases contained mutations in either STAG2 (n = 6) or RAD21 (n = 1) and occurred in all disease phenotypes (Fig. 2A). Notably, in 5 of 6 CBM/STAG2 co-mutated cases, the CBM mutation clustered within a 21-codon region between amino acids 491 and 512. Although cohesin gene co-mutations were uncommon in CBM-mutated patients, they resurfaced in those with mutations just beyond the CBM (Supplementary Tables 7–8), and the ASXL1/cohesin co-mutation rate in codons 588–594 (15%) were similar to those in the remainder of the ASXL1 gene (18%, p = 0.61).
We next investigated ASXL1mut patients for differences in co-mutation frequencies of non-cohesin genes based on CBM status (Fig. 2B, Supplementary Table 9). CBM mutations were significantly enriched for SETBP1, EZH2, or CUX1 co-mutation (each p < 0.05). Thus, co-mutation data supported that CBM mutations are usually exclusive of cohesin mutations, and that mutations in other genes may be selectively favored over cohesin mutations when the CBM is compromised.
Clinical and mutational overview of ASXL1mut versus ASXL1wt patients.
We next performed demographic, phenotypic, and co-mutation comparisons between ASXL1mut and ASXL1wt patients. We hypothesized that comparing features based on ASXL1 mutational status using a large, diverse dataset could a) shed light on biologic differences between these groups, and b) allow for more focused subset analysis of patients with co-mutations in other genes associated with presence or absence of ASXL1 mutation. Overall, ASXL1mut patients were significantly older, more often male, and more likely to have a European genomic ancestry than ASXL1wt patients (all p < 0.01, see Table 1). Total burden of non-ASXL1 mutations (Supplementary Fig. 5A) was higher in the presence of an ASXL1 mutation (median 3, IQR 2–5) compared to an ASXL1wt setting (median 2, IQR 1–4, p < 0.01). Nearly all ASXL1mut patients (96%) harbored one or more additional mutations, with a median and mode of 4 total mutations. In contrast, an ASXL1wt genotype was associated with fewer total mutations (median = 2, mode = 1; p < 0.01).
SRSF2 was the most frequently co-mutated gene with ASXL1 (34%), and as a gene class, spliceosome mutations were found in 57% of ASXL1mut cases. SETBP1 and EZH2 mutations exhibited the greatest enrichment in ASXL1mut cases (each 9-fold more than ASXL1wt frequency, Supplementary Fig. 5B-C and 6). Mutations in STAG2, RUNX1, and the signaling factors CSF3R and CBL were also far more common in the setting of ASXL1 mutation. In contrast, and consistent with prior studies, 8,9,28 NPM1, WT1, TP53, and DNMT3A mutations were more common in ASXL1wt patients.
Given our large dataset, we performed gene-gene mutation co-occurrence and mutual exclusivity (COME) analysis using a cancer-specific independence test (DISCOVER, see Materials and methods for details).19 Mutations in nine genes—SRSF2, U2AF1, RUNX1, SETBP1, EZH2, STAG2, CUX1, CSF3R, and CBL—showed strong co-occurrence (q < 0.01) with ASXL1 mutation (Fig. 3, Supplementary Spreadsheets S2-S3). The most striking distinction we appreciated within a single gene class was the mutual exclusivity and co-occurrence of ASXL1 mutation with SF3B1 and non-SF3B1 spliceosome gene mutations (SRSF2, U2AF1, ZRSR2), respectively. Thus, we investigated for potential relationships between ASXL1 and spliceosome gene mutation status.
Different spliceosome gene mutations occur based on ASXL1 mutational status.
SRSF2 and U2AF1 mutations were strongly associated with the presence of ASXL1 co-mutation (Fig. 3, q-values < 10− 10 and < 0.01, respectively). ZRSR2 mutations were also common in this setting (6.4% in ASXL1mut vs. 2.9% in ASXL1wt, q-value = 0.06). In contrast, SF3B1 mutations were strongly mutually exclusive with ASXL1 mutations (q-value < 10− 12). This SF3B1mut/ASXL1wt association was observed in all myeloid phenotypes (Fig. 4A). MDS/MPN patients showed the strongest trend towards a SF3B1mut/ASXL1wt genotype, while mutations in non-SF3B1 spliceosome genes were generally enriched in ASXL1mut AML, MDS, and MPN (fold difference in mutational frequencies ranging from 1.7 to 4.9). SRSF2, U2AF1, and SF3B1 mutations were mutually exclusive of one another independent of ASXL1 status, while ZRSR2 was commonly co-mutated with other spliceosome genes (Supplementary Spreadsheets S3, S5, Supplementary Table 10).
We examined VAFs of ASXL1 and spliceosome genes significantly co-mutated with ASXL1 (Fig. 4B) as a surrogate marker for mutational acquisition. Median VAFs of mutations in SRSF2 (43%, p < 0.01), U2AF1 (40%, p < 0.01), and ZRSR2 (38%, p = 0.06) were higher than ASXL1 (34%) when co-mutated, suggesting that these spliceosome mutations more often occur in the dominant neoplastic clone in the presence of ASXL1 mutation. However, SF3B1 VAFs were similar regardless of ASXL1 mutation status (median VAF of both ASXL1mut and ASXL1wt cohorts = 35%; p = 0.43). ASXL1mut patients harbored significantly fewer total mutations when co-mutated with SF3B1 compared to those with non-SF3B1 spliceosome mutations (Fig. 4C). ASXL1/SF3B1-mutated patients also displayed a significantly different co-mutation profile with the myeloid transcription factors CEBPA and RUNX1. Thus, these collective findings raise the possibility that ASXL1mut myeloid neoplasms with SF3B1 co-mutation may be biologically distinct from those with other spliceosome mutations.
Pairwise mutations and genotype-phenotype analyses reveal ontogenetic insights into ASXL1mut myeloid neoplasms.
We next tabulated co-mutation frequencies of other commonly mutated genes in ASXL1mut patients. Patients with either STAG2, SETBP1, or IDH2 co-mutation were most likely to harbor an additional mutation in a spliceosome gene, occurring in > 70% of cases each (Supplementary Fig. 7). We thus performed a separate COME analysis to identify any gene-gene pairs with a tendency for co-mutation in an ASXL1mut background and found several mutational co-occurrences in gene-gene pairs (Fig. 5A, Supplementary Fig. 8). However, these associations were not dependent on the presence of ASXL1 mutation and were also found to occur in an ASXL1wt setting (see Supplementary Spreadsheets S4-S5). Thus, we focused our remaining analysis on only those pairs where both genes were among the nine genes significantly co-mutated with ASXL1 in the entire cohort we discovered previously (see Fig. 3).
SRSF2 mutations tended to co-occur with STAG2 (q < 0.01), IDH2 (q < 0.01) and SETBP1 (p = 0.03) mutations. However, SETBP1 mutations were strongly mutually exclusive with both STAG2 and IDH2 mutations (each pair q < 0.01, Fig. 5B and Supplementary Fig. 8). We also performed a COME analysis within the smaller ASXL1/SRSF2 co-mutated cohort (n = 480, Supplementary Table 11) and again found SETBP1 mutations to be exclusive of STAG2 and IDH2 mutations (each q < 0.01). These findings suggested some functional coordination between certain genes (e.g., SRSF2-STAG2-IDH2), and potentially divergent pathways of mutation acquisition among others (e.g., SETBP1 and STAG2).
To explore genetic ontogeny in these multi-mutant subsets, we performed multivariate logistic regression analysis to determine if co-mutation could predict AML, MDS, or chronic proliferative (either MPN or MDS/MPN) phenotypes (Table 2, Supplementary Table 12). Of the genes significantly co-mutated with ASXL1, STAG2 (OR 2.0, p < 0.01) and RUNX1 (OR 1.7, p < 0.01) co-mutations were predictive of AML. Among chronic myeloid neoplasms, STAG2 was also predictive of MDS (OR 1.6, p = 0.03) over a chronic proliferative phenotype (OR 0.2, p < 0.01). Co-mutation with the spliceosome genes SF3B1, U2AF1, and ZRSR2 were each strongly predictive of MDS (all OR > 2.5, p ≤ 0.01); however, no specific phenotypes were significantly favored with SRSF2 co-mutation. SETBP1 co-mutation was associated with a chronic proliferative phenotype (OR 1.9, p = 0.01), while being significantly less likely to be seen in MDS (OR 0.6, p = 0.02).
Table 2
Phenotype-genotype analysis
Phenotype | Mutation* | OR [95% CI] | P |
AML | FLT3 | 2.7 [1.4–5.2] | 0.0046 |
| IDH1 | 5.5 [2.8–10.5] | < 0.0001 |
| IDH2 | 3.0 [1.9–4.8] | < 0.0001 |
| NPM1 | 14.4 [2.8–78.6] | 0.0015 |
| RUNX1 | 1.7 [1.2–2.4] | 0.0025 |
| STAG2 | 2.0 [1.3–2.9] | 0.0016 |
MDS | SF3B1 | 8.2 [4.2–16.0] | < 0.0001 |
| U2AF1 | 2.8 [1.9–4.3] | < 0.0001 |
| ZRSR2 | 2.7 [1.5–4.7] | 0.0013 |
Chronic proliferative† | CALR | 11.4 [4.0-32.6] | < 0.0001 |
JAK2 | 7.1 [2.8–18.1] | < 0.0001 |
| KIT | 4.4 [1.5–12.3] | 0.0054 |
| MPL | 6.8 [2.7–17.5] | < 0.0001 |
| NRAS | 1.9 [1.2–2.8] | 0.0041 |
| SETBP1 | 1.9 [1.2-3.0] | 0.0077 |
Results of multivariate logistic regression analysis for AML, MDS, or chronic proliferative phenotype and the presence of co-mutation within the ASXL1mut cohort (n = 1,414). |
* Only genes that were a) mutated in > 1% of all myeloid neoplasms and b) had strongly statistically significant differences in odds ratios between phenotypes (p < 0.01) are shown. |
† Chronic proliferative: either MPN or MDS/MPN phenotype |
In AML, ASXL1mut patients (n = 408) harbored additional co-mutations in genes associated with secondary AML 70% of the time (Supplementary Figs. 9–10).8 We further analyzed ASXL1/SRSF2 co-mutated AML patients due to the high frequency and strong prognostic significance of this genotype.9,29,30 As supported by our prior COME analyses, ASXL1/SRSF2 co-mutated AML had very frequent STAG2 (42%) and SETBP1 (16%) mutations (Fig. 5C). Notably, the presence of either additional mutation was dependent on the presence of both ASXL1 and SRSF2 mutations (each p < 0.05). STAG2 and SETBP1 were the only mutations with this feature in ASXL1/SRSF2-mutated AML (Supplementary Fig. 11). STAG2 and SETBP1 mutations were again mutually exclusive in this setting (p < 0.05), and cases with these genotypes displayed co-mutational differences in several genes, especially those in signaling/kinase pathways (Supplementary Fig. 12). Additional mutations in SRSF2 and IDH2 were also seen in ASXL1/STAG2 co-mutated AML and dependent on both ASXL1 and STAG2 mutations (Fig. 5C, Supplementary Fig. 13), supporting the ASXL1-SRSF2-STAG2-IDH2 multi-mutant tendency we identified by COME analysis.
Given the mutational signatures we identified in AML, we also analyzed chronic myeloid phenotypes in cases with these genotypes (Fig. 5D). AML and MDS were the dominant phenotypes in all ASXL1-SRSF2-STAG2-mutated patients, including those with additional IDH2 mutations. In contrast, the ASXL1-SRSF2-SETBP1-mutated genotype consisted mostly of chronic proliferative (56%) and AML (27%) phenotypes. Mutation rates of STAG2 and SETBP1 in the setting of ASXL1/SRSF2 co-mutation were starkly different in AML/MDS and chronic proliferative phenotypes, respectively (all p ≤ 0.001, Supplementary Fig. 14). In addition, the VAF structure of these co-mutated cases did not significantly differ by phenotype (Supplementary Fig. 15), as STAG2, SETBP1, and IDH2 mutation were only occasionally subclonal to ASXL1 and SRSF2 mutations. Thus, the mutual exclusivity of STAG2 and SETBP1 mutations we identified by COME analysis in both ASXL1mut and ASXL1/SRSF2-co-mutated patients appeared to be associated with divergent myeloid phenotypes.