Construction of a prognostic Editor/Inflammation-based score
An overview of the workflow is shown in Figure 1. In brief, we applied univariate and multivariate Cox regression analysis testing 163 selected features (Appendix Table 1-4 (online only)) to identify covariates with significant time-to-event outcomes that were independent from patient age and first-line treatment (hazard ratios and individual test statistics are shown in Figure 2 and Appendix Table 5 (online only)). To identify potentially interesting relationships, we computed pairwise Pearson’s correlation between all survival-associated parameters and observed a positive correlation between multiple APOBEC and cytokine transcript levels, which was most significant among APOBEC members 3C, -3D, -3F, and -3G (R=0.81-0.9, p<5*10-18, Appendix Figure 1A (online only)). Furthermore, APOBEC3A levels correlated significantly with a subset of interleukin receptors comprising the IL-8 receptors, CXCR1 and CXCR2 (R=0.87, p<2*10-24), CSF3R (R=0.92, p=4*10-11), as well as the IL10 receptor paralog IFNAR1 (R>0.89, p=2*10-21), further supporting the connection between APOBECs and inflammation in MM (Appendix Figure 1B, (online only)).
A fourteen-parameter model that reproducibly showed the best performance in predicting PFS and OS included ß2M, hemoglobin, LDH as well as RNA transcript levels of three APOBEC family members (APOBEC2, APOBEC3B, APOBEC3C), six pro/anti-inflammatory cytokines (IL10, IL11, IL27, IFNG, TGFB1, TGFB3) and all tested genome mutational signatures (SBS1, SBS2, SBS5, SBS13, SBS40) [33]. Interestingly, prediction of PFS in CoMMpass patients strongly relied on low APOBEC2 expression levels (PFS z=-4.34, p=0.00001 vs. OS z=-3.56, p=0.0004, (Appendix Table 6, (online only)) and was improved by inclusion of the SBS1 ageing- and SBS2 APOBEC-associated mutational signatures (PFS: p=0.004 and p=0.005, respectively vs OS: not significant). In contrast, OS prediction in this patient cohort was improved by the inclusion of creatinine blood levels and RNA levels of cytokines such as TGFB3 and IL11 (Appendix Table 6, (online only)). Notably, only one out of five cytogenetic features, being gain(1q), passed all of our selection criteria.
It should be noted that the aim of this study was not to establish a de novo transcription-based classifier for NDMM patients but rather to test whether incorporation of disease-associated and therefore pre-selected features would be able to predict patient outcome with comparable or greater accuracy than existing classifiers that heavily rely on cytogenetics. As a proof of principle, we therefore decided to further reduce our editor/inflammation-associated feature set from 16 to 7 variables, only retaining those that were most significant for both PFS and OS prediction. Mutational signatures were excluded as they are difficult to compute in clinical practice. Lastly, we applied a simplified weighting based on the rounded integer multivariate z-score of each parameter:
EI-score[OS]: (APOBEC2: ≤0.20=0, >0.20=3) + (APOBEC3B: ≤3.79=0, >3.79=3) + (IL11: ≤0.48=0, >0.48 =2.5) + (TGFB1: ≤0.11=0, >0.11 =1) + (TGFB3: ≤0.10=0, >0.10=2) + (ß2M: ≤4.22=0, >4.22=4) + (LDH: ≤3.18=0, >3.18=2)
EI-score[PFS]: (APOBEC2: ≤0.19=0, >0.19=4) + (APOBEC3B: ≤6.91=0, >6.91=3) + (IL11: ≤0.56=0, >0.56 =2) + (TGFB1: ≤0.43=0, >0.43 =1) + (TGFB3: ≤0.03=0, >0.03=1) + (ß2M: ≤3.33=0, >3.33=4) + (LDH: ≤3.08=0, >3.08=2)
A total of 599 CoMMpass patients had sufficient data to compute both EI-scores. Computed EI-score[OS] values applying the above formula in these 599 patients ranged from 0 to 17.5 and computed EI-score[PFS] values from 0 to 15 (Appendix Figure 2A and B, (online only)). The median PFS and OS was 688 (95% CI; 605 to 760) and 1052 (95% CI; 996 to 1094) days, respectively. Dichotomization of this patient subset into EI-score[OS] standard risk and high-risk patients groups allowed the classification of a standard-risk cohort with a 5-year OS rate of more than 50% and median OS of 2207 days (95% CI; 2207 to 2207) as well as a high-risk group with a median OS of 1500 days (95% CI; 934 to 1500; Figure 3A). Similarly, dichotomization into EI-score[PFS] standard- and high-risk patient groups allowed the classification of a standard-risk cohort with a median PFS of 1302 days (95% CI; 1176 to 1472) and a high-risk cohort with a median PFS of 485 days (95% CI; 390 to 604, Figure 3B). Overall, the EI-score accurately predicted outcome in the CoMMpass patient cohort and we observed a positive correlation between EI-score[PFS] and EI-score[OS] (R=0.78; P<0.0001, Appendix Figure 2C, (online only)). Applying the EI-score formulas to the independent IFM/DFCI patient cohort (n=263), we could confirm the above findings by dichotomization into two EI-score[OS] risk groups (47% and 53%, respectively) with a 5-year OS rate of 87% and 73% (p=0.0062; Figure 3C), respectively. Similarly, PFS was dichotomized into two EI-score[PFS] risk groups (45% and 55%, respectively) with 37% and 25% 5-year OS rates, respectively (p=0.0007; Figure 3D). Of note, the reported 5-year OS rate of 78% in the IFM/DFCI 2009 study was superior to only 68% in the CoMMpass cohort.
The EI-score outperforms clinically established MM risk classifiers
To compare the prognostic accuracy of the EI-score to ISS, R-ISS, and mSMARTcyto (a reduced version of the mSMART score based exclusively on presence of t(4;14), t(14;16), gain 1q and/or del 17p), we computed EI-score performance metrics including the multivariate CoxPH Concordance index (Ci) and ROC-AUC in MMRF patients applying a standardized machine learning pipeline (Table 1). The EI-score achieved the best performance ratings for OS and PFS prediction (n=599; Ci 0.7 and 0.69, respectively; Table 1) followed by ISS (n=1113; Ci 0.66 and 0.6), R-ISS (n=690; Ci 0.64 and 0.6), and mSMARTcyto (n=817; Ci 0.58 and 0.54). Even after exclusion of patients allocated to the ISS, R-ISS, and mSMARTcyto intermediate risk classes, the EI-score showed superior classification accuracy, especially in predicting unfavourable outcomes (Table 1).
We then assessed which “score classes'' (blood markers, editors, or cytokines) had the most weight in clinical risk prediction. Although the best stratification was achieved when applying a combination of all three score classes, both APOBEC genes and blood parameters showed the most significant differential presence in high compared to low clinically defined patient risk groups (Appendix Figure 3A-B, online only). Next, we asked whether addition of gene expression information can improve the performance of established risk classifiers. Therefore, we added the individual EI-score components of each score class to the risk classifiers ISS, R-ISS and mSMARTcyto, which significantly improved their performance of PFS and OS prediction. The best results were achieved when transcript levels of both editors and inflammatory cytokines were added simultaneously to the established scores (Table 1). Overall, classifiers that included routinely assessed blood parameters, which are already known to be predictive of MM disease outcome (β2M and LDH), had the highest classification accuracy and scores based on cytogenetics performed poorest among all tested combinations (Appendix Table 6 online only). In line with our observation that adding gene expression parameters of the EI-score to ISS, R-ISS, and mSMARTcyto significantly improved their accuracy, we observed further improvement of these classifiers when cytogenetic information was removed completely (Table 1, see R-ISSnocyto), underscoring the potential of gene expression classifiers such as the EI-score for MM outcome prediction.
Applying the EI-score, we next aimed to identify MM patients with very high and very low risk of progression or death, similar to the three-risk-group stratifiers ISS and R-ISS. Based on natural valleys in the score distribution (Appendix Figure 2A,B, online only), we selected two cut-offs (EI-score 9 and 3), yielding EI-score low-, intermediate- and high-risk groups (comprising 31%, 56% and 13% of the investigated MMRF patient cohort (n=599); Figure 3E-F). The 5-year OS rate for patients classified into the individual groups were 85% (EI-score[OS] low-risk), 65% (intermediate-risk) and 35% (high-risk). Overall, hazard ratios for patients in the EI-score[OS] high category were four times higher than those of patients categorized as EI-score[OS] low (16.9 versus 69.9, 95% CI). A detailed comparison between the EI-score, ISS and R-ISS risk group survival probabilities are shown in Appendix Table 7 and Appendix Figure 4 (online only). Eventually, we applied the EI-score to already classified ISS and R-ISS risk groups which allowed for a further sub-stratification of these patients, revealing previously unrecognized risk subgroups (Figure 4A-B).
To adjust for the heterogeneous treatment protocols of patients included in the CoMMpass dataset, we performed a sub analysis for MM patients receiving CyBorD or VRD ± autologous stem cell transplantation. Also in these treatment groups, the EI-score outperformed ISS and R-ISS in predicting patient outcome (Figures 5A and B).
The EI-score identifies novel MM risk subgroups
Recent publications point towards the necessity of a more refined classification of high-risk MM patients. To test the suitability of the EI-score for sub-stratifying adverse risk MM patients, we applied the EI-score to patient subgroups that carried either del(17p), gain(1q) or t(4;14). With this approach, we were able to identify patients with poor and favorable outcomes in the CoMMpass (Figure 4C) and the IFM/DFCI cohorts (Figure 4D). In the CoMMpass dataset, the EI-score was able to subclassify del(17p) MM patients into three main risk subgroups: a very good prognosis group (0% with additional TP53mut) with 5-year OS of 100%, an intermediate group (30% with additional TP53mut) with 5-year OS rate of 75%, and a very poor prognosis group (40% with additional TP53mut) with 5-year OS of 0% (2-year OS: 40%) (CoMMpass data, Figure 4C). These finding could be reproduced in the IFM/DCFI cohort, where application of the EI-score allowed the subclassification of del(17p) MM patients into a standard (5-year OS rate: 73%), and a high-risk group (5-year OS of 35%) (Figure 4D). Similarly, the EI-score allowed us to identify t(4;14) MM patients with a very poor prognosis (median OS: 20 months) (Figure 4D). In line with the reported pathophysiological relevance of nucleotide editors for MM, we found that CoMMpass study patients who carried either del(17p), gain(1q) or t(4;14) and also achieved a high EI-score, displayed an enrichment of APOBEC-induced genomic mutations compared to intermediate and low EI-score patients (Appendix Figure 5, online only). Taken together, our data demonstrates that a simple gene expression-based score, composed to reflect disease biology, is able to identify previously unrecognized subgroups of MM patients who display adverse risk cytogenetics but experience favorable outcomes.