A Response-Adaptive Randomization (RAR) design in clinical trials using the prediction of treatment outcomes based on Machine Learning models relating the biomarker of a patient to the treatment outcome is proposed in this article. This is studied for the case of rare diseases to efficiently allot the patients among various treatments so as to ensure the rights and maximum possible benefits to the patients in clinical trials. Based on the historical data, the best few treatments are considered for allotting patients in the subsequent steps. The method uses initial burn-in period of conventional equal randomization and helps in understanding how well the chosen treatments work in patients. However, the significant biomarker values are also observed for those patients in the burn-in period, though they are not used at this stage. Then a machine learning model is fitted relating the observed biomarker values and the treatment effective score computed. At the second stage, the biomarker values are considered to allot the patients based on the ML model fitted. The performance of the proposed design is measured by the proportion of patients cured by suitable statistical tools. The study reveals that the proposed design can be implemented practically in clinical trials of rare diseases.