Heterogeneity-optimized method for predicting immune checkpoint blockade response.

Area of Science:

• Computational biology
• Cancer research
• Immunotherapy

Background:

• Interpatient tumor heterogeneity presents multimodal distributions in genomic, transcriptomic, and microenvironmental profiles.
• This heterogeneity violates unimodal assumptions in conventional machine learning, hindering accurate prediction of immune checkpoint blockade (ICB) response.
• Existing predictive models struggle to account for complex tumor variations.

Purpose of the Study:

• To develop a heterogeneity-optimized framework for improved ICB response prediction.
• To address the limitations of conventional machine learning models in handling multimodal tumor data.
• To enable biologically interpretable precision immunotherapy by modeling multimodal heterogeneity.

Main Methods:

• Applied K-means clustering to stratify patients into hot-tumor and cold-tumor subgroups, outperforming hierarchical and DBSCAN clustering.
• Developed subtype-specific predictive models: a support vector machine for hot-tumor and a random forest for cold-tumor subtypes.
• Utilized seven heterogeneity-associated biomarkers to build models that circumvent unimodal constraints.

Main Results:

• The proposed framework significantly enhanced ICB response prediction across melanoma, NSCLC, other cancer types, and pan-cancer datasets.
• Achieved a mean accuracy gain of at least 1.24% compared to 11 baseline methods.
• Performance improvements were consistently validated in an independent external cohort.

Conclusions:

• The heterogeneity-optimized framework effectively models multimodal tumor heterogeneity for superior ICB response prediction.
• This approach offers a pathway to biologically interpretable precision immunotherapy.
• The findings demonstrate a significant advancement in predicting patient response to cancer immunotherapy.