How reliable is your AI-ML model for predicting time to response, complete vs. partial response, degree of response, and durability of response, for monotherapy and for synergistic combination therapy? Is AUC the major predictor?

The CertisAI-ML model uses Area Under the Curve (AUC), especially for in vitro data, because it explains sensitivity and incorporates Cmax (versus IC50). However, we do not need AUC because we combine learned in vitro and in vivo (PDX) datasets, as well as large screening studies from Novartis and other big institutions. For PDX, you want to use TGI, for in vitro you want to use AUC, and for clinical studies, you want to use classes (complete, partial, non-response). We normalize all the data into a metric that is relative to that dataset. Instead of AUC, you normalize all the data to distribution and convert back to what you want to predict (TGI, AUC).

PDX Response Prediction Scale:

  • 100 TGI (complete response)
  • 51 – < 99 (partial response)
  • < 50 TGI (no response)

The scale is about learning and figuring out what that scale is for patients.