Machine learning uncovers invariant evolutionary molecular trajectories in MDS. Free

Distinct molecular underpinnings in diverse cases of MDS may result in similar prognosis and therefore outcome-supervised prognostic systems are unlikely to reflect common pathogenesis within such risk groups. We used unbiased autoencoder-based machine learning (ML) to resolve the genetic complexity of MDS by defining molecular clusters (MCs) based on combinatorial similarity of genetic features. The resultant classification is reflective of the convergence of leukemogenic pathways and combinatorial similarities with distinctive inter MC signature. Using such an algorithm referred to here as the “molecular nosology” of MDS encompasses 14 MCs¹. Because of clonal architecture dynamics, stationary molecular MCs may reflect not only the stage of the disease but also the subsequent direction and speed of progression. This notion inspired our hypothesis that serial MC analysis of patients may reveal invariant evolution patterns with clinical implications. Our algorithm was trained and validated in 3588 patients¹and now we applied the molecular nosology classification to an expanded cohort of 3810 patients, of whom 320 were studied serially (210 clinical progressors; 110 non progressors). Our expanded analysis yielded comparable MC composition and distribution as the original training set. For example, MC2,13 & 7 comprised 22, 20, and 11%, respectively, while MC4, 6, 8, 9 & 12 each represented approximately 6-7% of the cohort while MC1, 5, 11 & 14 ~12% of patients. We set to determine invariant trajectories of individual MCs including transitions (MC switching) and how these patterns are associated with progression dynamics. As expected, at progression the MC distribution was skewed towards MCs associated with advanced disease (MC13 & 7) and either by acquisition of new MC defining mutations, or within MC through non-MC-defining hits or clonal burden increases (intra MC progression). In contrast, in non-progressors, MC distribution did not change significantly with only 9% of cases exhibiting new MC redefining. Among the 210 serial cases with documented progression, 40% underwent MC reassignment, while 60% retained their original MC; this dynamic reorganization of clonal architecture was non-random and predictive of clinical behavior (e.g., MC13 emerged as the central node of malignant progression). Prevalence of MC13 as a central progression node increased from 20% (93% of whom remained in the same MC) to 34% at progression, MC7 also expanded (11→19%) and 21% of patients transitioned to MC13, suggesting that MC7 functions as a gateway MC in malignant acceleration. MC2, the most common MC at baseline declined (22→11%) with only 1/3 of patients retaining their original classification. Despite its association with a normal karyotype, DNMT3A and RAS mutations. MC2 frequently transitioned to MC13 & 7. MC10 (favorable karyotype, SF3B1, DNMT3A, TET2) disappeared entirely at progression. MC8, defined by del(5q), DNMT3A, and early TP53 was unstable with 54% of them transitioning to MC13. MC4 & MC5 had intermediate behavior; MC4, associated with SF3B1 and DNMT3A and a normal karyotype, often transitioned to MC7. Similarly, MC5, initially defined by del(20q), U2AF1, and ASXL1, frequently progressed to MC7,11 or 13. MC6 emerged as intrinsically unstable and associated with SRSF2 and RAS pathway but also shifted to higher risk states irrespective of progression status, suggesting a transient, unstable molecular configuration. In contrast, MC3, characterized by del(Y), TET2, ZRSR2, ASXL1 was stable, while MC9,11 & 12 (ASXL1, SRSF2, TET2, RUNX1) rarely transitioning to MC13. Of note, MC transitions showed invariant timing which can be applied in reverse (prior to diagnosis) or forward fashion. Our analysis also allowed conclusions as to the ontogenesis of nosologic MCs. Reverse analyses recapitulated the “founder MCs” from which advanced MCs must have evolved prior to the initial diagnostic testing. In sum, our study underscores the dynamic, but non-random nature of molecular MC nosology of MDS and sAML, significantly influencing leukemic evolution and the clinical course. MCs in progressed patients result from underlying invariant patterns from originator MCs acting as “evolutionary destinies”. Blast trajectory measures in the context of specific molecular progression patterns will offer further insights into disease kinetics, aiding the early identification of rapid progression patterns.

Ref.1. Nat Commun. 2023 May 30;14(1):3136.