The vastness of the chemical space of compound scaffolds is humongous and it represents a large playground for potential lead drug discovery or repurposing. With the accumulation of experimental data over the years, together with the development of more complex statistical frameworks, screening of such elaborate chemical spaces is finally possible. There are several well-defined problem areas for drug screening efforts, the most popular being the inhibition activity against a multitude of protein targets in human cells related to often occurring diseases. Some examples of highly targeted protein spaces include protein kinases, g-protein coupled receptors, and/or (non)selective serotonin re-uptake inhibitors. Mutation and dysregulation in any of the three of the mentioned protein groups can result in hereditary disorders, tumors, and mental disorders. Contrary to the available machine learning frameworks for prediction of direct physical interactions between compounds and protein targets, certain chemical activity predictions are not well-represented or defined in the literature, e.g. phytotoxic activity. In this work, publicly available data is collected with regard to the experimentally measured binding affinities of diverse compounds against one of the most popular target protein families, protein kinases. This protein super-family is one of the most important enzyme groups responsible for the regulation of most of the important cellular processes, including cell metabolism, cell growth, and division. Protein kinases regulate biochemical cycles by transferring high energy phosphoryl group from adenosine-3-phosphate (ATP) to specific amino acid residues of the target protein substrates. All members of this enzyme family are characterised by the highly conserved protein kinase (PK) domain, but depending on the phosphorylation site and the activation mechanisms of individual members of this family, this superfamily can be divided into several kinase groups. Due the specific characteristics of this protein group and kinase inhibitors, it is important to investigate how each of these chemical or biological spaces impact models performance and how to achieve more optimal predictive performance. On the other hand, we examine a different subspace of biological activity, focusing mostly on synthetic compounds with determined phytotoxic or herbicidal activity. We define this problem as a multiclass classification problem by using two predefined classification systems: main one, by the Herbicide Resistance Action Committee (HRAC), and the second one, by the Weed Science Society of America (WSSA). Considering that no defined machine learning framework for modeling and prediction of herbicidal activity was publicly available, an effort was made to collect the representative data set and define the optimal computational approach to maximize the prediction accuracy for mode of action (MoA) prediction. Considering that the classification of phytotoxic compounds was mostly performed by visual inspection of phenotypic changes in the affected weeds, there is a great need for an automated, systematic approach to this endeavor. Due to the limited size of the collected data, consisting of molecular structures of known activity and denoted by a MoA group, we further tested several “shallow” learners. The panel of tested algorithms includes naive bayes (NB), support vector machines (SVM), extreme-gradient boosting approach (XGBoost) and random forest (RF). All the approaches mentioned were trained in a ten times repeated ten fold (10x10-fold) cross validation mode. A comparison of trained models over all hundred resamples was performed using a non-frequentist approach - Bayesian analysis. For the first time for the herbicide activity modeling, we have implemented a computational framework from feature processing and selection to the training of several learners and, ultimately, a statistical comparison of their performance. However, due to the sheer size of the publicly available experimental data for protein kinase inhibitors, modeling of physical interactions between small compound spaces and the human kinome has allowed for more complex modeling techniques - but has also been more challenging in defining and engineering the feature space for over 8000 compounds and the nuance of the protein kinase family. Both of the aforementioned methods are founded on the QSAR (Quantitative structure-activity relationship) modeling principles. The definition of the applicability domain (AD) for a specified problem is one of the pillars of QSAR modeling. However, defining the boundaries of the chemical space within which the model can make accurate predictions is not simple and is dependent on the nature of the trained model. In the case of predicting general biological activity in the form of a phenotypic signal, as is the case with herbicidal activity, the applicability domain can be simply defined in two-dimensional space by considering the structural similarity of available molecules and a model output, such as the probability of belonging to a particular class. Predicting the physical interaction between any two entities, such as compounds and protein targets, adds complexity that cannot be accommodated by the conventional applicability domain. In this instance, we intend to extend the standard applicability domain to include both entities and generate a quantitative estimate of prediction confidence using the conformal prediction framework. Conformal predictors can reliably estimate a prediction region based on the computed nonconformity of test samples. The disadvantage of this method is that the nonconformity is defined in the label space of predefined calibration samples, resulting in estimates that work well in general but are not specific to any tested compound-target pair, thus failing for samples that are not already available in the training set. Combining concepts from both frameworks, we dynamically define similarity-based applicability domains or conformity regions for each new sample and then calculate nonconformity scores - we refer to this approach as the dynamic applicability domain (dAD). The dAD approach was shown to produce tighter prediction regions when compared to the original conformal predictors algorithm. More importantly, complementary to the prediction regions, when it comes to realistic use-case scenarios (S2, S3), dAD achieves lower error rates for any confidence level. More importantly, merging the concept of applicability domain with a conformal predictors corrects for existing bottlenecks in the traditional applicability domain definition and allows for the evaluation of model behavior in an abstract interaction space between any number of interacting entities. This way, it is a valuable and informative approach for validation of data quality in subregions of interaction space specific for biomolecular complexes.