Abstract Introduction: Gene amplification and/or protein overexpression of EGFR have been observed in a variety of carcinomas, including lung, colorectal, urinary bladder, breast, head/neck, esophageal, and gastric carcinomas. Increase in EGFR expression has been observed to be associated with advanced stage and an unfavorable prognosis in some carcinoma, e.g., non-small cell lung carcinoma (NSCLC) and colorectal carcinoma (CRC). Cetuximab is a recombinant human/mouse chimeric monoclonal antibody against EGFR, which was approved for treating EGFR-expressing metastatic CRC (mCRC) without activating KRAS mutation, and squamous cell carcinoma of the head and neck (SCCHN)(1,2). However, response to cetuximab treatment in CRC seems to have no correlation with EGFR expression. Furthermore, preclinical studies demonstrate cetuximab treatment can inhibit tumor growth in a variety of indications, including gastric cancer, with drug efficacy positively correlated with EGFR expression level. Response to cetuximab may involve different biologic pathways in different indications. Methods: We collected genomics data and cetuximab efficacy data for PDX cohorts of 26 CRC, 27 gastric carcinoma, and 37 NSCLC. We used linear mixed model to compare the tumor growth rates of different tumors under cetuximab treatment, and to analyze the interaction between EGFR expression level and tumor growth, as compared to the traditional method using tumor growth inhibition (TGI) as efficacy endpoint. In addition, we developed a machine learning protocol to identify potential biomarkers for cetuximab treatment in these carcinomas at both gene levels and pathway levels. Results: By our new analysis method using machine learning, we have made some new discoveries using our available datasets. First, under cetuximab treatment, tumor growth rate of CRC is significantly higher than that of NSCLC and gastric carcinoma. Second, for gastric cancer we confirmed that EGFR overexpression resulted in better response to cetuximab treatment, while we found the association of overexpression of PROX1 with drug resistance. Third, for CRC, while we somehow confirmed the earlier observation of RAS-signature score correction with the cetuximab resistance based on TGI, we also attempted to identify additional gene changes significantly better correlated with efficacy according to our new analysis method using machine learning. Fourth, we are also revealing new observation by performing the same in NSCLC, where our previous attempts using TGI failed. We will report our new findings in the upcoming presentations. Conclusions: Cetuximab treatment in different indications results in significantly different results, which can be correlated with genetic makeup of different cancers. Our new analysis method using machine learning can be a powerful tool to reveal new predictive insights of drug response.