The issue of in silico analysis plays a crucial role in designing new medicines in modern day pharmaceutical industry. Selecting the best candidate for a new drug among countless molecules is a challenge which can be facilitated by machine learning methods. Following article addresses the problem of computational prediction of Human Immunodeficiency Virus (HIV) inhibition level among molecules. We introduced the cross siamese network (CSN) - a novel architecture based on siamese neural network - generating an embedding aiming to enhance the prediction process. The proposed neural net is a hybrid type model which combines embeddings generated from several subnetwork trained in estimating HIV inhibition level and other chemical properties like: solubility, lipophilicity or toxicological effects. To verify the efficiency of the proposed solution we trained a set of k-nearest neighbors classifiers on starting molecules’ fingerprints and embeddings outputted by the experimental models. The results from this test showed that some versions of model enhanced the molecular embeddings, improving their utility for predicting HIV inhibition.