The kinetic properties that regulate drug-target interactions play a key role in determining pharmacological efficacy, as they influence how long a compound remains bound to its receptor and, consequently, the duration and intensity of the therapeutic effect. Among these, the dissociation rate constant koff is particularly relevant, because it directly relates to the average residence time of the bound drug, thereby reflecting the stability of the ligand-receptor complex. In this context, the experimental determination of kinetic parameters such as the dissociation rate constant remains a complex, slow, and often poorly reproducible process, making the use of alternative computational tools increasingly necessary to accelerate the drug discovery and development pipeline. To this end, in silico prediction methods represent a promising strategy to accelerate the study of binding kinetics and to support the early stages of drug discovery. In this thesis, a computational strategy for predicting the koff value for protein-ligand complexes has been developed, by combining in-silico methodologies including molecular modelling, machine learning and deep learning approaches. The developed prediction pipeline is divided into two layers, where starting from experimentally determined protein-ligand complexes, two complementary approaches were adopted: a ligand-based approach, focused exclusively on the molecular features of the ligand, and a structure-based approach, describing the structural properties of the protein. The predictive strength of the model results from the integration of these two complementary representations, allowing both components of the interaction to be considered synergistically when estimating the koff parameter. In the final phase, a multimodal model was developed based on transformer architectures, a class of deep learning models capable of learning contextual representations from sequential data, such as protein and molecular structures. The developed model combines a data-driven numerical representation of the receptor structure (using ProtBERT embeddings) and a corresponding numerical embedding descriptive of the small bound molecule (using ChemBERTa). The integrated representation exploits the power of pre-trained deep learning models within an optimized ML pipeline, achieving state-of-the-art predictive performance. Overall, the results demonstrate the efficiency of the developed multimodal model in estimating complex kinetic parameters. By integrating molecular, protein, and structural information the model achieves robust and generalizable performance across a wide range of protein-ligand complexes, showing that the proposed approach is not restricted to specific protein families. These insights confirm the potential of data-driven multimodal models as reliable tools for predicting binding kinetics and as valuable methodological support for rational drug design optimization.