MARM: a framework for malignancy risk prediction from host-derived CNV in bronchoalveolar lavage fluid mNGS data with microbial admixture.

Early identification and risk assessment of malignancy are essential for improving clinical decision-making and patient outcomes. Bronchoalveolar lavage fluid (BALF) metagenomic next-generation sequencing (mNGS) data contain both microbial and host-derived signals, and a key challenge in extending such data to tumor-associated applications is the robust extraction of host features with discriminative value for malignancy from this complex, admixed background. To address this problem, we developed MARM, a malignancy risk prediction method centered on host-derived copy number variation (CNV). Using host-derived reads from BALF mNGS data, MARM performs genome-wide window-based coverage quantification, normalization and bias correction, reference baseline construction, and principal component-based denoising to derive window-level CNV features for malignancy risk modeling. In addition, a pseudo-label-based extension strategy was introduced to incorporate weakly labeled samples through high-confidence screening, and the performance of XGBoost, Random Forest, and generalized linear models (GLM) was systematically evaluated using CNV features, microbial features, and combined features. Models built on host-derived CNV features consistently outperformed those based on microbial features and achieved performance comparable to combined-feature models, while joint modeling did not provide a stable additional benefit. These findings indicate that, under the current data setting and feature construction strategy, CNV represents a more stable and informative discriminative signal than microbial features. Among the evaluated classifiers, XGBoost showed the best compatibility with window-level CNV features and outperformed Random Forest and GLM overall. On the independent validation set, the pseudo-label-enhanced MARM achieved the best overall performance, with a sensitivity of 0.686, specificity of 0.975, accuracy of 0.847, and Youden index of 0.671. By contrast, microbial features did not show stable independent discriminative ability, and combined modeling did not yield clear or sustained performance gains. Together, these results indicate that, in microbially admixed BALF mNGS data, host-derived CNV is more suitable than the evaluated microbial features as the core modeling signal for malignancy risk prediction. MARM provides a new methodological framework for malignancy prediction in complex clinical samples and offers a reference for deeper exploitation of host-derived signals in mNGS data and related auxiliary diagnostic applications.