Predictive Biosignature and Risk Scoring for Bone Disease in Multiple Myeloma

Introduction

Multiple myeloma (MM) is a hematological malignancy of bone marrow plasma cells. Clinically, it is manifested with a spectrum of symptoms including focal/bone lesions, hypercalcemia, anemia, and renal dysfunction. Bone disease (BD) in MM is characterized by osteopenia, osteoporosis, and osteolytic lesions, significantly impacting quality of life and survival (Kanellias et al., 2022). Diagnosis and clinical monitoring of BD involves advanced and expensive imaging techniques like whole body computer tomography (CT) scans, magnetic resonance imaging (MRI) and positron emission tomography (PET) scans, which along with new treatments, contribute to MM’s high economic burden. There is currently no clear guideline for BD monitoring frequency or biomarkers for stratifying newly diagnosed patients by BD risk. At the same time, the molecular mechanisms of BD in MM are not fully understood, and it is unclear why some patients develop BD and others do not.

Identifying BD risk at diagnosis is crucial for personalized treatment, healthcare recourse utilization and improving quality of life. Concurrently, further research into MM’s molecular dysregulations, both with and without BD, can help develop tailored therapies.

Methods

In the first step, we discovered a biosignature predictive of BD using a discovery cohort of newly diagnosed MM patients (n=86, BD=49, non-BD=37). BD was determined by X-ray before 2014 (n=24) and by CT after 2014 (n=62). The discovery cohort included 25084 transcriptomic (RNA-Seq) and 9 clinical and biochemical features. Using the JASPAR-v2022 database, we built a global transcription-factor-gene network. Biosignature mining was performed by integrating web datamining and machine learning (ML) algorithms with 5-fold cross validation, similar to AlgoOS (Kashif et al., 2023). Model performance was evaluated by accuracy and harmonic mean of precision and sensitivity (F1-score).

For additional evaluation, we benchmarked the transcriptomic feature selection using the transcription-factor-gene network against both random selection of 15 features and the entire transcriptomic dataset.

In the second step, we developed a BD prediction score using the identified biosignature by calculating logarithmic transformation of feature ratios. The scoring system was tested in the discovery cohort and validated in an independent external validation cohort (n=25, BD=19, non-BD=6). Performance was assessed by receiver operating characteristic (ROC) analysis using area under curve (AUC). All the analyses were performed in R and Python.

Results

In the first step, using the discovery cohort, we identified a 15-feature biosignature predictive of BD risk with 91% accuracy and 92% F1-score. This included 9 transcriptomic (TDO2, PLPP1, ST8SIA4, FER, HOMER1, RHOBTB3, PNKD, GNPTAB, SENP7) and 6 biochemical features (calcium, β2M, hemoglobin, creatinine, FLC ratio and type of M component heavy chains). Additionally, benchmarking transcriptomic feature selection using the transcription-factor-gene network showed a 32% improvement in prediction accuracy over random feature selection and a 27% improvement over the entire dataset.

In the second step, the final biosignature was used to develop the BD risk score. This scoring system included only 4 genes (TDO2 HOMER1, GNPTAB, SENP7) and calcium. It achieved an AUC of 89% (95% CI[77, 97]) in the discovery cohort and an AUC of 83% (95% CI[63, 96]) in the external validation cohort.

A bibliographic review showed that all 4 genes in the risk score are relevant to MM or bone metabolism. Tdo2 is linked to inflammation in osteoarthritis (Rong et al., 2022), Homer1 to osteoblast activation and osteoporosis (Rybchyn et al., 2021), Gnptab to lysosome function in osteoclasts (VanMeel et al., 2011) and SENP family members to bortezomib resistance (Xie et al., 2020). Calcium levels are correlated with BD conditions (Tian et al., 2018), indicating these features are biologically relevant and potential targets for understanding BD.

Conclusions

We discovered a biosignature and developed a risk score with only 5 features (4 genes and calcium) for prediction of BD in MM patients and validated externally. These tools hold potential for personalized healthcare, better resource utilization, and may provide insights into BD’s molecular mechanisms. Further validation in larger cohorts and clinical trials is needed.