Multimodel inference for biomarker development: an application to schizophrenia.

Area of Science:

• Psychiatric research
• Proteomics
• Biomarker discovery

Background:

• Predictive models for complex psychiatric disorders often lack reproducibility.
• Independent validation of predictive models is crucial for clinical utility.

Purpose of the Study:

• To enhance the predictive performance and reproducibility of models for complex psychiatric disorders.
• To investigate the utility of multimodel inference for predicting schizophrenia probability.
• To identify potential protein biomarkers and their interactions with sex in schizophrenia.

Main Methods:

• Utilized proteomic data from drug-naive schizophrenia patients and controls.
• Applied lasso regression with repeated tenfold cross-validation for model training.
• Employed feature extraction and model averaging for multimodel inference.

Main Results:

• Achieved good predictive performance on training (AUC > 0.80) and test sets (AUC > 0.80).
• Demonstrated the utility of multimodel inference for reproducible schizophrenia prediction.
• Identified four proteins with sex-modified effects, including fetal hemoglobin as a novel potential biomarker.

Conclusions:

• Multimodel inference offers a robust approach for developing reproducible predictive models in psychiatry.
• Sex-specific analyses are warranted in schizophrenia research due to identified effect modifications.
• Future biomarker studies should incorporate multimodel approaches and consider interactions beyond main effects.