The standard first treatment for acute myeloid leukemia (AML) is called induction chemotherapy - an intensive course of drugs (typically cytarabine and an anthracycline like daunorubicin) designed to rapidly destroy leukemia cells and induce a complete remission. For many patients it works. But for a significant minority, the leukemia is resistant to induction therapy - the cancer cells survive the treatment and the patient fails to achieve remission.
The scale of this problem is substantial: approximately 20-30% of younger adults with AML are resistant to induction treatment, and up to 50% of older adults fail to respond. These patients face an extremely poor prognosis. Their only potential cure is an allogeneic stem cell transplant - and even reaching that option requires surviving the failed initial chemotherapy and its side effects, which include prolonged bone marrow failure (aplasia) and serious infection risk.
The fundamental clinical dilemma is that doctors currently cannot reliably identify these resistant patients before starting treatment. If they could, alternative strategies (less toxic approaches, clinical trials of novel agents, faster referral for transplant) could be offered upfront. Several scoring tools exist, but the best achieve only an AUC of 0.71-0.78 in validation - 'fair' performance by standard criteria. This study aimed to build a substantially better predictor by combining gene expression data with clinical information.
The researchers used three independent patient cohorts. Training set 1 comprised 407 patients from the multicenter AMLCG-1999 trial (a German AML Cooperative Group phase III trial), with gene expression measured on Affymetrix microarrays. Training set 2 consisted of 449 patients from the HOVON trials in the Netherlands, also measured on Affymetrix arrays. Both training sets used the same platform, enabling consistent variable selection across them.
The validation set was designed to be maximally stringent: 250 patients from the AMLCG-2008 trial, with gene expression measured by RNA sequencing - a completely different technology from the Affymetrix arrays used in training. Validating a gene expression predictor across platforms is a high bar; many such scores fail when the measurement method changes. To increase the number of treatment-resistant cases in the validation set (which was needed for adequate statistical power), an additional 40 resistant patients from the AMLCG-1999 trial not used in training were added.
The development process was comprehensive: clinical variables, cytogenetic risk classification (MRC criteria and European LeukemiaNet ELN 2017), and the mutation status of 68 recurrently mutated AML genes were all considered as candidate predictors alongside gene expression data. Penalized logistic regression (Lasso) was used to select variables and estimate their weights - a method that automatically discards weak predictors and prevents overfitting. Crucially, the final predictor and its cut-off threshold were defined before any data from the validation set was accessed.
The final predictor, named PS29MRC (Predictive Score 29 MRC), combines expression levels of 29 genes with the MRC cytogenetic risk classification (favorable vs. unfavorable cytogenetics). The score is calculated as a weighted linear sum of these components - each gene contributes positively or negatively depending on whether its high expression is associated with treatment resistance or treatment response. In the validation set, scores ranged from -2.75 to +3.72, reflecting a wide spread in predicted resistance probability.
In validation, PS29MRC as a continuous variable achieved an AUC of 0.76 - 'fair' by standard classification, but superior to all currently used clinical scoring tools tested in the same cohort (the Walter score achieved AUC 0.71; the stemness-based LSC17 score, retrained for resistance prediction, achieved AUC 0.61). When the pre-defined threshold was applied to create a binary high-risk/low-risk classification, the odds ratio for treatment resistance was 8.03 - meaning patients in the high-risk group were eight times more likely to be resistant to induction therapy than low-risk patients.
Accuracy in the validation set was 77%. The score was designed with high specificity (90%) in mind - meaning very few patients who would actually respond to treatment are wrongly flagged as resistant. The tradeoff is lower sensitivity (46%): about half of the patients who will ultimately be resistant are not captured. This design choice reflects clinical reality: excluding a patient from a potentially curative treatment (induction chemotherapy) requires very high confidence that resistance is likely.
To understand whether PS29MRC adds information beyond what is already known, the researchers constructed multivariable models that included all variables with significant associations with treatment resistance in both the training and validation sets. After controlling for everything else, only three variables remained independently significant: PS29MRC, patient age, and TP53 mutation status. All other clinical and molecular markers lost their individual significance when PS29MRC was included - a strong statement that the score captures most of the predictive information contained in the other variables.
TP53 mutations in AML are associated with the most aggressive, treatment-resistant disease and are known to be a special category requiring different management. The fact that TP53 mutation remained significant alongside PS29MRC - but that common molecular markers like NPM1 mutation and FLT3-ITD did not - suggests that TP53-mutated AML involves distinct resistance mechanisms not fully captured by the gene expression signature. Consistently, PS29MRC showed limited predictive power specifically within the TP53-mutated subgroup.
The score also predicted overall survival independently in multivariable analysis (HR 2.15, P=0.00052). When integrated with the ELN 2017 genetic risk classification, four groups with dramatically different survival were defined: the PS29MRC high-risk group had a median survival of only 8 months (12% alive at 24 months), compared to the ELN favorable-risk group where median survival was not reached (76% alive at 24 months). This four-group stratification provides finer resolution than either the ELN system or PS29MRC alone.
One of the most important features of this study is that the score was validated using RNA sequencing in the validation cohort, while it was built using Affymetrix microarrays in the training cohorts. These are fundamentally different technologies - one measures fluorescent signal intensities on probe arrays, the other counts RNA molecules directly by sequencing. Maintaining predictive power across this technology gap is a stringent test that many published gene expression signatures fail.
Performance varied across AML genetic subgroups. PS29MRC showed high accuracy in common molecular subtypes including core binding factor AML (CBF), KMT2A-rearranged AML, CEBPA double-mutated AML, and NPM1-mutated AML. Importantly, it also showed significant prognostic value in the large subgroup of patients without any classifiable genetic mutation - a group that is particularly hard to risk-stratify with existing molecular tools.
The score showed moderate predictive accuracy in high-risk subgroups defined by TP53 mutations or chromosomal aneuploidy. This is a recognized limitation: these aggressive AML subtypes have resistance mechanisms that likely operate through pathways not fully represented in the 29-gene signature. For these patients, additional research is needed to identify alternative markers or completely different therapeutic strategies.
PS29MRC identifies approximately 20% of intensively treated AML patients as high-risk for primary resistance. These patients have a median survival of only 8 months and a 24-month survival rate of 12% - outcomes that are devastating given the toxicity of intensive induction chemotherapy, which subjects patients to weeks of profound bone marrow failure, infection risk, and organ stress. The paper explicitly raises the question of whether standard induction chemotherapy can still be considered appropriate for this subgroup.
Currently, the only potentially curative option for treatment-resistant AML is an allogeneic stem cell transplant. But reaching transplant requires surviving failed induction therapy - a difficult and sometimes impossible hurdle for patients in the poorest condition. If PS29MRC could identify high-risk patients before treatment begins, they could potentially be enrolled in clinical trials of alternative induction strategies: less toxic hypomethylating agents, venetoclax-based regimens, or novel targeted therapies that might achieve remission with less upfront toxicity and enable a path to transplant.
The score also has implications for clinical trial design. Currently, AML induction trials enroll all patients regardless of their probability of response. With a validated pre-treatment resistance score, trials could be designed to stratify by resistance probability - testing whether more aggressive or novel approaches specifically benefit the high-resistance-probability group, while confirming that standard therapy remains appropriate for likely responders.
PS29MRC represents a meaningful advance over existing AML resistance prediction tools. It was rigorously validated in an independent cohort using a different gene expression measurement platform, dominated all competing models in head-to-head comparisons, and remained independently significant even after controlling for age, cytogenetics, and 68 gene mutations. The combination with ELN 2017 genetic risk stratification creates four prognostically distinct groups that are more informative than either system alone.
An important biological observation from the development process is that adding mutation data from 68 AML-associated genes did not improve the model beyond what cytogenetics and gene expression already provided. This mirrors findings from the Walter et al. group and suggests that gene expression - capturing the functional output of all mutations simultaneously - is a more complete representation of resistance biology than cataloging individual mutations. Gene expression integrates the effects of all genetic and epigenetic alterations into a single phenotypic readout.
Practical implementation will require standardized RNA sequencing workflows for routine clinical use - a challenge that is becoming increasingly feasible as costs fall and laboratory capacity grows. The authors envision PS29MRC being integrated into the initial AML diagnostic workup, providing resistance probability alongside the cytogenetic and molecular reports that already guide treatment decisions. The next step is prospective clinical validation to confirm that using PS29MRC to guide treatment selection actually improves outcomes for patients in the highest-risk group.