P-370: CoMMsol: a machine learning model for outcome-specific Copy Number Alterations’ clonality cut-offs definition in Multiple Myeloma

Introduction: Multiple Myeloma (MM) genetic architecture is complex and heterogeneous and includes a vast amount of Copy Number Alterations (CNAs), whose presence and recurrence can be used to stratify patients (pts) into prognostic subgroups. Due to the sub-clonal composition of MM, CNAs call is not trivial, especially when small clonal frequencies should be discriminated from background. At the same time, the set-up of experimental cut-offs is crucial to assess the CNAs prognostic role. To date, a fixed 10% cut-off is commonly employed for all CNAs’ call.

Aims: 1)To employ a machine learning(ML) approach to define the best clonality cut-offs for each CNAs, able to predict for early relapse (ER)(i.e. within 18 months from the start of therapy); 2) to select baseline CNAs most significantly correlated to the achievement of ER.

Methods: 743 newly diagnosed MM pts with SNPs array-derived genomic data and clinical annotations were included in the study. Analyzes were performed by R Studio 4.1.2 scripts and packages.

Results: Overall, 155/743pts relapsed within 18 months from the start of therapy. By ML computational approach, we set up a model (named coMMsol) able to select, for each observed CNAs, the best clonality cut-off significantly predictive of ER. Among CNAs called according to this model, the one significantly correlated to ER were amp1q, amp2p, del2p, del12p, del17p, del19p.
By stratifying pts according to the presence of at least one of these CNAs, two groups of 249 and 494 pts, either carrying or not ≥1 ER-CNAs were identified, the first one with significantly shorter PFS and OS (PFS HR 2.15, CI 1.79-2.58 p< 0001, OS HR 2.37, CI 1.86-3.02 p< 0.0001). The worse prognostic impact was observed even in the absence of both amp1q and del17p, observed in 38/249 (15%) pts (PFS HR 2.57, CI 1.79-3.70 p< 00001, OS HR 2.49, CI 1.57-3.94 p=0.0001), compared to 494pts.
The risk of relapse defined by the presence of ≥1 ER-CNAs was independent from those conferred both by R-IIS 3 (HR=1.51 CI 0.17-2.39-2.2; p=0.01) and by low quality (< stable disease) first-line best clinical response (HR=2.59 CI0.33-2.84 p=0.004). Notably, the type of induction therapy was not descriptive in this multivariate model, suggesting that ER is strongly related to pts’ baseline genomic architecture, and not to induction therapy they were provided to.

Conclusions: ML approach allowed to define CNAs-specific dynamic clonality cut-offs, improving the CNAs calls’ accuracy to identify MM pts with the highest probability to ER. The use of these ER-related cut-offs pinpointed few CNAs, whose presence at baseline is highly predictive of ER, including amp2p, del2p, del12p, and del19p, whose biological role in MM needs further investigations.
As being outcome-dependent, the coMMsol method is dynamic and might be adjusted according to the selected outcome variable of interest (e.g. MRD-negativity) thus providing outcome-specific clonality cut-offs.
Thanks to AIRC IG2018_22059, RF-2016-02362532