A Machine Learning Platform to Optimize the Translation of Personalized Network Models to the Clinic

Area of Science:

• Computational biology
• Systems biology
• Cancer research

Background:

• Dynamic network models offer insights into disease mechanisms and clinical prognosis.
• Personalizing these models is hindered by the need for extensive input data, limiting clinical application.

Purpose of the Study:

• To establish computational platforms that reduce input requirements for prognostic network models.
• To demonstrate the feasibility of optimizing model inputs for clinical translation.

Main Methods:

• Applied APOPTO-CELL, a prognostic model for apoptosis signaling, using two pipelines: Ensemble (probabilistic, consistent inputs) and Tree (machine learning, personalized inputs).
• Developed and validated models on a virtual cohort (3.2 million patients) and an in-house stage III colorectal cancer cohort (120-117 patients).
• Utilized reverse phase protein array and immunohistochemistry for protein profiling.

Main Results:

• Ensemble and Tree achieved 92% and 99% accuracy, respectively, reducing inputs by 40% and 46%.
• Both methods retained prognostic utility in the colorectal cancer cohort.
• Immunohistochemistry proved suitable for quantifying model inputs in routine clinical settings.

Conclusions:

• A generalizable framework was developed to optimize network-based prognostic assays.
• This framework facilitates the integration of prognostic models into routine clinical workflows.