Background: Anticancer drugs often have strong toxicity against tumours and normal cells. Some natural products demonstrate high tumour specificity. We have previously reported the cytotoxic activity and tumour specificity of various chemical compounds. In this study, we constructed a database of previously reported compound data and predictive models to screen a new anticancer drug. Methods: We collected compound data from our previous studies and built a database for analysis. Using this database, we constructed models that could predict cytotoxicity and tumour specificity using random forest method. The prediction performance was evaluated using an external validation set. Results: A total of 494 compounds were collected, and these activities and chemical structure data were merged as database for analysis. The structure-toxicity relationship prediction model showed higher prediction accuracy than the tumour selectivity prediction model. Descriptors with high contribution differed for tumour and normal cells. Conclusions: Further study is required to construct a tumour selective toxicity prediction model with higher predictive accuracy. Such a model is expected to contribute to the screening of candidate compounds for new anticancer drugs.