Artificial intelligence predicted natural killer cell fitness to guide daratumumab and transplant strategies in newly diagnosed multiple myeloma.

Abstract

7520 Background: Daratumumab (Dara) requires CD16-expressing natural killer (NK) cells for anti-myeloma activity. The use of bone marrow transplant (ASCT) results in prolonged immunosuppression, including lower levels of NK-cells. We evaluated if artificial intelligence (AI) predicted CD16 from routine hematoxylin and eosin (H&E) stained bone marrow biopsy slides could be used to guide Dara and transplant treatment strategies. Methods: We analyzed 212 newly diagnosed patients from the HealthTree registry treated with VRd (n=135; median follow-up 24.5 months (mo)) and D-VRd (n=77; 16.6 mo). H&E slides were processed using GigaTIME, a foundation model via zero-shot inference. Patients were dichotomized at median predicted CD16 and stratified by ASCT status. Subgroups: transplant-eligible (TE: age<65, ASCT; n=53), deferred (TD: age<70, no ASCT; n=34), and ineligible (TI: age≥70; n=17). Primary endpoint: time to next treatment (TTNT). Multivariate Cox regression adjusted for age, cytogenetic risk tested interaction between CD16, treatment and ASCT. Results: D-VRd achieved superior outcomes vs VRd (23.4% event rate vs 74.8%, p<0.001) despite lower ASCT use among D-VRd treated patients (50.6% vs 71.1%, p=0.005). In the VRd group, AI-predicted CD16 was highly prognostic in non-transplanted patients: High-CD16 achieved median TTNT 33.1 vs 5.1 mo for Low-CD16 (p=0.0053). ASCT rescued Low-CD16 patients (median 5.1 to 26.7 mo, p<0.001). Conversely, in D-VRd, CD16 did not stratify outcomes (p=0.48), indicating that Dara can rescue low-immune-fitness disease. Specifically, Dara rescued low CD16 non-transplanted patients from progression (86.8% 18-mo event-free survival for D-VRd versus 28.6% 18-mo event-free survival for VRd; p<0.001). High CD16 patients on D-VRd without ASCT achieved 86.3% event-free survival, non-inferior to ASCT (82.2%, p=0.21), supporting transplant deferral. The CD16×treatment×ASCT interaction was significant (HR 0.06, p=0.045). Among transplant-deferred patients, D-VRd improved 18-mo event-free rates vs VRd (70.5% vs 40%, p=0.029), with the largest benefit in Low CD16 (p=0.016). TI patients on D-VRd achieved 100% event-free at 18 mo vs 38.2% on VRd (p=0.011). Conclusions: We validated that AI-quantified CD16 from routine H&E guides Dara and transplant strategies in NDMM (n=212). The CD16×treatment×ASCT interaction (HR=0.06, p=0.045) confirms that CD16 predicts treatment benefit, not just prognosis. In VRd, low CD16 identifies a high-risk group (median TTNT 5.1 mo) needing intensification. D-VRd rescues low-immune-fitness patients (18-mo event-free survival: 86.8% vs 28.6% with VRd, p<0.001) and enables safe transplant deferral in high-fitness patients (82.2% with ASCT vs 86.3% without, p=0.21). This novel strategy enables precision medicine using standard H&E slides, with immediate clinical implications.