13129 Background: Therapy in inoperable non-small cell lung cancer has limited effectiveness. New therapy directed at tumour markers is attractive but limited by marker prevalence and intrinsic effectiveness. Prior work (Abstract 9639, 2005 PASCO) on 191 trials of first line therapy shows a “typical drug” needs at least a 35% difference in response rate, treatment versus control, to predict a survival gain. The current work describes the characteristics of targeted drugs likely to predict this gain at smaller differences, i.e., more discriminating drugs. We determine how much greater effects on response rate and survival, compared to the typical drug, need to be to predict this survival gain. Methods: By using 1.0, 1.1, 1.2, 2.0-fold increments in response rates over traditional therapy and the same increments for median survival up to that used for response, we defined 55 differently constructed targeted drugs. The drugs were all analysed at each of four marker prevalence rates, 5%, 20%, 50% and 100%, each analysis determining the response rate difference needed to predict a survival gain. For each analysis individual trial response rates were modified in one arm before applying linear regression. The response rate difference to predict a survival gain is the portion of the prediction band fully exceeding zero. Results: Improvements in discrimination were rare and highly sensitive to marker prevalence. Little improvement occurred for any drug with 5% or 20% prevalence, and if no survival improvement occurred, discrimination worsened (up to 90%). If both tumour response and survival were doubled, discrimination improved to 33%, 27%, 17%, and 7% for marker prevalence of 5%, 20%, 50%, or 100%, respectively. Modest changes (×1.4, ×1.6) in response and survival only improved discrimination with 100% prevalence. Clearly improved discrimination (to 20% or less) only occurred with at least ×1.8 increase in response and survival and 50% or more prevalence. Conclusions: The results show that targeting a marker only substantially improves discrimination if it doubles both tumour response rate and median overall survival and if marker prevalence is at least 50%. This approach may help rationalize aspects of drug development in oncology. No significant financial relationships to disclose.