A blood-based protein signature adds to genomic prognostication of survival in TP53-mutated Acute Myeloid Leukemia Free

Outcomes for patients (pts) with newly diagnosed acute myeloid leukemia (AML) have improved significantly with the advent of genomically targeted therapies. However, 5–10% of pts with de novo AML and up to 30% of those with therapy-related AML (tAML) harbor TP53 mutations (TP53^MUT), which are associated with a particularly poor prognosis and median overall survival (mOS) of 6-8 months. Stem cell transplantation remains the only potentially curative option, yet relapse rates approach 50% at one year. Prior studies suggest that TP53 mutation burden – as assessed by variant allele frequency (VAF) and/or allelic status – is independently associated with survival. We developed a blood-based protein biomarker signature to improve prognostic stratification in TP53^MUT AML.

Methods Serum samples from pts with newly diagnosed TP53^MUT AML were collected prior to administration of definitive induction therapy. Using the NULISA platform, 251 circulating inflammatory proteins were quantified. To identify the key proteins prognostic for survival, we implemented a rigorously validated and stress-tested machine learning pipeline. 70% of pts were randomly allocated to a training set while 30% was reserved for independent validation testing. Model development followed the Predictability, Computability, and Stability (PCS) framework. Specifically, within the training set, 1000 bootstrap resamplings were performed and in each iteration importance scores based on the univariate coefficients were estimated for all proteins. Features that consistently ranked among the top ten predictors across iterations were selected for final modeling. Next, a machine learning algorithm based on the Cox proportional hazards model was constructed using the most stable features. Model stability and prognostic utility were assessed across clinically relevant subgroups. An optimal cutoff for stratifying pts into high- and low-risk groups was derived using the minimum p-value method. Both the combination rule from the selected features and risk threshold were independently validated in the testing set. The resulting model was benchmarked against an established risk model from Senapati et al., Haematologica 2025 which only considers genomic factors for risk stratification. In this model, pts are considered low risk (TP53^LR) if they possess a single TP53 mutation with VAF <40% without concurrent 17/17p loss.

Results The panel of 251 inflammatory proteins was quantified from 108 pts with newly diagnosed TP53^MUT AML from 2012 to 2023 with a median follow up of 19.0 months. The median age at diagnosis was 69 years and 53% of pts were male. Sixty-seven pts (62%) had de novo AML, 31 (29%) tAML, and 10 (9%) secondary AML. Overall, 96 (89%) pts received a lower intensity induction regimen, with 69 (64%) including venetoclax. Using the Senapati genomic model, 27 (25%) had TP53^LR disease.

Thirteen circulating serum proteins —IL10RB, IL1R1, TNFRSF4, TNFSF8, TIMP1, CCL13, IL18BP, SIRPA, CD46, TNFRSF1B, NCR1, OSMR, and TNFRSF11A—consistently ranked among the top important features across 1000 iterations. Using the machine learning model, an optimal risk cutpoint from the 13 proteins was estimated. The fixed combination of the 13 serum inflammatory proteins with the derived cutpoint yielded a hazard ratio (HR) of 2.31 (95% CI: 0.97–5.54; p=0.03) for death in the high-risk vs low-risk cohort in the testing set. The mOS of pts with high-risk TP53^MUT AML based on this signature was 3 months (95% CI: 2-5) vs 10 months (95% CI: 7-15) in the low risk. In the full dataset, the 13-protein signature demonstrated significant additive prognostic value beyond the established Senapati genomic risk model based on likelihood ratio testing (p=0.001). When integrated with the Senapati model, the 13-protein panel with TP53 genomic aberrations achieved an HR of 3.02 (95% CI: 1.85-4.90) for risk stratification of survival, compared to an HR of 2.29 (95% CI: 1.05–5.00) from the genomic risk model alone.

Conclusions A blood-based 13-protein signature derived using machine-learning effectively stratifies high and lower risk TP53^MUT AML, with additive prognostic effect to a standard genomic-based model. In such a high-risk population, inflammatory protein modeling provides valuable information that could guide future therapeutic approaches. Future studies with longitudinal serum proteome quantification in TP53^MUT AML are ongoing.