Cancer treatments often lose effectiveness as tumors develop resistance to single-agent therapies. This challenge can be addressed through combinatorial treatments, where multiple drugs are administered simultaneously. However, the combinatorial space of N different drugs at d doses results in d^N possible treatments, making it impractical to test each combination experimentally. Moreover, mechanisms of drug action are not known in many cases. In this work, we explore the efficiency of a data-driven approach (deep learning) with categorical embeddings to identify synergic drug combinations using minimal experimental and molecular information. Comparing this method to classical machine learning approaches, we show that the best performance is achieved by a combination of the different methods, and that inclusion of drug molecular fingerprints greatly improves synergy prediction. This study leverages neural networks to extract meaningful patterns from categorical variables, and provides a guide to optimize the prediction of drug synergy, potentially enhancing treatment efficacy.