T21. DEVELOPMENT OF PROTEOMIC PREDICTION MODELS FOR OUTCOMES IN THE CLINICAL HIGH RISK STATE AND PSYCHOTIC EXPERIENCES IN ADOLESCENCE: MACHINE LEARNING ANALYSES IN TWO NESTED CASE-CONTROL STUDIES

BACKGROUND: Individuals at clinical high risk (CHR) of psychosis have an approximately 20% probability of developing psychosis within 2 years, as well as an associated risk of non-psychotic disorders and functional impairment. People with subclinical psychotic experiences (PEs) are also at risk of future psychotic and non-psychotic disorders and decreased functioning. It is difficult to accurately predict outcomes in individuals at risk of psychosis on the basis of symptoms alone. Biomarkers for accurate prediction of outcomes could inform the clinical management of this group. METHODS: We conducted two nested case-control studies. 1. Study 1, a clinical high-risk (CHR) sample, was nested within the European Network of National Schizophrenia Networks Studying Gene-Environment Interactions (EU-GEI) cohort study. The sample comprised 133 CHR participants who were followed clinically for up to 6 years, of whom 49 transitioned to psychosis and 84 did not. 2. Study 2, a general population sample, was nested within the Avon Longitudinal Study of Parents and Children (ALSPAC). The sample comprised 121 participants who did not report psychotic experiences (PEs) at age 12, of whom 55 later reported PEs at age 18 and 66 did not. We employed discovery-based proteomic methods to analyse protein expression in baseline plasma samples in EU-GEI and age 12 plasma samples in ALSPAC using liquid chromatography mass spectrometry. Differential expression of quantified proteomic markers was determined by analyses of covariance (with false discovery rate of 5%) comparing expression levels for each marker between those who did not and did not develop psychosis in Study 1 (adjusting for age, gender, body mass index and years in education), and between those who did and did not develop PEs in Study 2 (adjusting for gender, body mass index and maternal social class). Support vector machine algorithms were used to develop models for prediction of transition vs. non-transition (as determined by the Comprehensive Assessment of At Risk Mental States) and poor vs. good functional outcome at 2 years in Study 1 (General Assessment of Functioning: Disability subscale score </=60 vs. >60). Similar algorithms were used to develop a model for prediction of PEs vs. no PEs at age 18 in Study 2 (as determined by the Psychosis Like Symptoms Interview). RESULTS: In Study 1, 35 of 166 quantified proteins were significantly differentially expressed between CHR participants who did and did not develop psychosis. Functional enrichment analysis provided evidence for particular implication of the complement and coagulation cascade (false discovery rate-adjusted Fisher’s exact test p=2.23E-21). Using 65 clinical and 166 proteomic features a model demonstrated excellent performance for prediction of transition status (area under the receiver-operating curve [AUC] 0.96, positive predictive value [PPV] 83.0%, negative predictive value [NPV] 93.8%). A model based on the ten most predictive proteins accurately predicted transition status in training (AUC 0.96, PPV 87.5%, NPV 95.8%) and withheld data (AUC 0.92, PPV 88.9%, NPV 91.4%). A model using the same 65 clinical and 166 proteomic features predicted 2-year functional outcome with AUC 0.72 (PPV 67.6%, NPV 47.6%). In Study 2, 5 of 265 quantified proteins were significantly differentially expressed between participants who did and did not report PEs at age 18. A model using 265 proteomic features predicted PEs at age 18 with AUC 0.76 (PPV 69.1%, NPV 74.2%). DISCUSSION: With external validation, models incorporating proteomic data may contribute to improved prediction of clinical outcomes in individuals at risk of psychosis.