Colorectal cancer (CRC) is the third most detected cancer and the second foremost cause of cancer deaths in the world. Intervention targeting p53 provides potential therapeutic strategies, but thus far no p53-based therapy has been successfully translated into clinical cancer treatment. Here we developed a Quantitative Structure–Activity Relationships (QSAR) classification models using empirical molecular descriptors and fingerprints to predict the activity against the p53 protein, using the potency value with the active or inactive label, were developed. These models were built using in total 10,505 molecules that were extracted from the ChEMBL, ZINC and Reaxys® databases, and recent literature. Three machine learning (ML) techniques e.g., Random Forest, Support Vector Machine, Convolutional Neural Network were explored to build models for p53 inhibitor prediction. The performances of the models were successfully evaluated by internal and external validation. Moreover, based on the best in silico p53 model, a virtual screening campaign was carried out using 1443 FDA-approved drugs that were extracted from the ZINC database. A list of virtual screening hits was assented on base of some limits established in this approach, such as: (1) probability of being active against p53; (2) applicability domain; (3) prediction of the affinity between the p53, and ligands, through molecular docking. The most promising according to the limits established above was dihydroergocristine. This compound revealed cytotoxic activity against a p53-expressing CRC cell line with an IC50 of 56.8 µM. This study demonstrated that the computer-aided drug design approach can be used to identify previously unknown molecules for targeting p53 protein with anti-cancer activity and thus pave the way for the study of a therapeutic solution for CRC.