Risk Prediction in MDS: A Validation of the IPSS-M and Comparison to IPSS-R and to Two Other Personalized Prediction Tools

Background: Risk prediction in myelodysplastic neoplasms (MDS) has long been based on clinical variables and cytogenetic aberrations incorporated into the Revised-International Prognostic Scoring System (IPSS-R). A decade of sequencing data demonstrated the importance of molecular analysis. This is reflected in the Molecular International Prognostic Scoring System (IPSS-M), which includes mutations in 31 genes in addition to cytogenetics, bone marrow blasts, hemoglobin and platelet count (Bernard et al. NEJM Evid 2022). Recurrently mutated genes were also used in two personalized prediction models by Nazha et al. and Bersanelli et al. (both JCO 2021). Despite the seemingly analogous concepts of all three recent scores, our comparison uncovers differences in terms of their objective and predictive capabilities.

Aim: (1) Independently validate results of the IPSS-M in a cohort of 735 MDS cases (419 with <5% blasts) characterized in depth by whole genome sequencing (WGS); (2) compare the three recently published, molecular scoring systems.

Patients and Methods: Gold standard routine diagnosis was performed (morphology, chromosome banding analysis). In addition to panel sequencing, amplification-free WGS was performed for 735 patients (median coverage > 100x).

Results: The IPSS-M risk score was calculated for 735 cases with the following result: 15% (107) very low (VL), 39% (286) low (L), 12% (86) moderate low (ML), 7% (50) moderate high (MH), 13% (99) high (H), 15% (107) very high (VH). The IPSS-M includes TP53^multi (combination of mutations, deletion or loss of heterozygosity), which we identified in 41 of 735 patients (5.5%), of which 39/41 were classified as VH. We found FLT3^ITD+TKD in 8 patients (1%, all VH), and KMT2A^PTD in 8 cases (1%; 4 VH, 3 H, 1 ML). The IPSS-M risk categories showed strong prognostic separations for OS, LFS and leukemic transformation (LT), however ML and MH did not separate (Fig).

For the comparison of all 4 scores, we reduced the cohort to 439 patients with sufficient data for calculation of all models. Of this, 366 were part of the training cohort for the work by Nazha et al. and Bersanelli et al. While the application of IPSS-M and IPSS-R leads to one value or category, the other two models provide individualized predictions of OS and LT. All three novel scores include molecular markers, but only six genes overlap between all scores (ASXL1, RUNX1, SF3B1, SRSF2, STAG2, TP53). In total, Bersanelli et al. require 54, the IPSS-M 37 and the score of Nazha et al. 19 parameters (Fig.).

The IPSS-M does not consider age, but Bersanelli's score gives the second strongest weight to age, while age is of intermediate importance for OS in Nazha's model. The resulting differences are illustrated by the following: the youngest 10^th percentile of our cohort (median age 51 [23-56] years) includes 12/44 (27%) patients in the IPSS-M H/VH group, and have a median 60 months OS probability of 77% (21-97%) by Bersanelli et al. and 52% (13%-82%) by Nazha et al., respectively. On the other hand, the oldest 10^th percentile (median age 84 [82-93] years) also includes 10/44 (23%) patients from the H/VH class, but have a median 60 months OS probability of 27% (1-72%) by Bersanelli et al. and 36% [5%-57%] by Nazha et al.

We used the concordance index (Harrell's c-index) to assess predictions and real outcomes (median follow-up of 9.3 years), where 1 is the best degree of concordance. The c-index for IPSS-R was 0.67 (OS), 0.68 (LFS) and 0.77 (LT), and improved to 0.70 (OS), 0.72 (LFS) and 0.81 (LT) for the IPSS-M. This closely matches the values of the original IPSS-M publication. The personalized prediction models by Nazha et al. achieved 0.74 (OS) and 0.86 for (LT), and Bersanelli et al. resulted in 0.73 (OS) and 0.69 (LT); a direct comparison with the IPSS-M is hampered by a partial overlap with the respective discovery cohorts.

Conclusions: (1) With a cohort of 735 patients we confirm the increased predictive power by the addition of molecular parameters in the IPSS-M compared to IPSS-R. (2) The aims of the three recently published scores are different, reflected by the parameter selection (foremost age) and resulting in different strength to predict OS and LFS/LT. (3) We will need to translate molecular data into clinical decisions. Our goal should be a biological classification allowing selection of targeted therapies rather than prognostic scores, which do not differentiate between co-morbidities and biology driving progression.

Baer:MLL Munich Leukemia Laboratory: Current Employment. Huber:MLL Munich Leukemia Laboratory: Current Employment. Hutter:MLL Munich Leukemia Laboratory: Current Employment. Meggendorfer:MLL Munich Leukemia Laboratory: Current Employment. Nadarajah:MLL Munich Leukemia Laboratory: Current Employment. Walter:MLL Munich Leukemia Laboratory: Current Employment. Platzbecker:Takeda: Honoraria; Silence Therapeutics: Honoraria; Geron: Honoraria; Novartis: Honoraria; BMS/Celgene: Honoraria; Janssen: Honoraria; Jazz: Honoraria; Abbvie: Honoraria. Götze:Servier: Honoraria; Abbvie: Honoraria; BMS: Honoraria. Kern:MLL Munich Leukemia Laboratory: Current Employment, Other: Ownership. Haferlach:Munich Leukemia Laboratory: Current Employment, Other: Part ownership. Hoermann:MLL Munich Leukemia Laboratory: Current Employment. Haferlach:MLL Munich Leukemia Laboratory: Current Employment, Other: Ownership.