A Lightweight Risk Scoring System for Stratification of Newly Diagnosed Multiple Myeloma Patients

Introduction:

Multiple myeloma (MM) is a hematological malignancy of bone marrow plasma cells, with patient overall survival (OS) ranging from few weeks to more than 10 years. Stratifying at-risk newly diagnosed (NDMM) patients is crucial for tailoring treatment strategies. Historically, prognostic factors such as albumin, lactate dehydrogenase (LDH), β2-microglobulin (β2M), extramedullary involvement, and cytogenetic abnormalities have been studied extensively in MM. That lead to development of various clinically adopted prognostic models such as International Staging System (ISS) and its revisions R-ISS, R2-ISS, and International Myeloma Working Group (IMWG) criteria.

Additionally, machine learning (ML) models are being tested to discover biosignatures for risk stratification using next-generation sequencing (NGS) data. These models often consist of several features and lack integration with clinical data, making their clinical utility challenging. Therefore, most NDMM patients receive the same front-line treatment and personalized treatment options are not yet available.

Methods:

To address this issue, in the first step, we developed a biosignature of risk stratification by integrating transcriptomic (RNASeq), clinical and biochemical data from NDMM patients. We used a discovery cohort of 241 white patients from the MMRF CoMMpass study, and categorized them into short-term survivors (ST, died ≤ 3 years, n=97) and non-short-term survivors (non-ST, lived > 3 years, n=144). Using the JASPAR-v2022 database, we built a global transcription-factor-gene network. A Google PageRank-like algorithm and support vector machine (SVM) identified top predictive transcriptomic features. These selected features were then integrated with clinical and biochemical data to train a random forest model and a biosignature for stratification of at-risk patients was discovered. The performance of model was evaluated using the harmonic mean of precision and sensitivity, known as F1-score, with 5-fold Monte Carlo cross-validation. This method was similar to AlgoOS (Kashif et al., 2023).

For additional evaluation, we benchmarked the transcriptomic feature selection using the transcription-factor-gene network against both random selection, and feature reduction methods, including tSNE and UMAP.

In the second step, a risk prediction score was developed using the identified biosignature by calculating the logarithmic transformation of feature ratios. The performance of the scoring system was assessed by receiver operating characteristic (ROC) analysis, with the area under the curve (AUC). The scoring system was evaluated in the discovery cohort and validated in external cohort-1 (non-white, n=99; non-ST=50, ST=49) and external cohort-2 (Swedish patients, n=15; non-ST=10, ST=5). All the analyses were performed in R and Python.

Results:

In the first step, we discovered a biosignature for stratifying NDMM patients that achieved F1-score of 85%. Benchmarking showed that transcriptomic feature selection by the transcription-factor-gene network achieved the highest F1-score, followed by UMAP, tSNE and random selection.

In the second step, we developed a risk score using the identified biosignature, consisting of only eight features: five transcriptomic (RBPJ, PFDN2, SNRPC, TCF12, GMNN), two biochemical (albumin, creatinine) and one clinical (age). This system achieved an AUC of 81% (95% CI [73, 88]) in the discovery cohort.

In the validation cohort-1, the system achieved an AUC of 76% (95% CI [67, 85]) and in validation cohort-2, it achieved an AUC of 92% (95% CI [73, 100]),

The scoring system performed significantly better than the ISS in the discovery cohort (p value=0.0005) and validation cohort-1 (p value= 0.0056). In validation cohort-2, the ISS and R-ISS achieved AUCs of 55%, while the scoring system achieved AUC of 92%. However, reliable p values could not be computed due to the small sample size of cohort-2.

Conclusions:

We developed a lightweight scoring system for stratifying at-risk MM patients, using only eight features, which could be easily implemented in clinical settings.

It holds potential for real-time risk assessment and personalized treatment decisions for MM patients. However, these findings need validation in larger cohorts and prospective clinical trials.