One of the major parts of modern medicine research is detecting reactions of various molecules and microorganisms. This paper explores a solution for predicting the reactions of small molecules and microorganisms from the open-access CO-ADD database. Various techniques for improving the results were used: scaling and centering the input data, removing overrepresented data, duplicating rare cases, transforming the labeling problem into a classification problem, algorithmic search for the best model hyperparameters, removing contradictory data, and training multiple times. Even after applying all of these techniques, the obtained results are far from optimal. The data from the used database and the Mordred toolkit for converting SMILES representations of molecules into their features isn’t sufficient for a high-quality solution to this problem. This paper shows a significant advantage in the performance of the multi-label classification technique compared to the more conventional binary classification technique. This advantage is followed by its drawbacks like difficulties with data processing and slightly poorer results. Further research is needed for a more robust comparison of these two techniques. The most useful data preprocessing technique was principal component analysis (PCA), which reduced the number of data inputs and thus enabled results with a median F1 score of 0.429 for multi-label classification, and 0.438 for binary classification. This paper confirms the usefulness of multi-task learning with the goal of lower resource usage while also demonstrating the complexity of predicting the biological reactions of small molecules and microorganisms