Our research focuses on molecular recognition, with the goal of providing methods and software for solving biomedical problems. The primary areas of interest are protein-protein interactions and the ligand binding properties of proteins. We believe that predictive methods will be substantially improved during the next five years due to the increasing amount of information on sequences, structures, and interactions of molecules in the cell, and the development of powerful machine learning tools. To take advantage of these opportunities we will integrate the use of structural templates, co-evolutionary information, and machine learning into classical biophysical methods. This part of the research is facilitated by the availability of much improved neural network algorithms. We already combined our popular ClusPro protein-protein docking program with Alphafold2 and Alphafold-Multimer. For the prediction of antibody-antigen complexes the combined method substantially improves the rates of generating acceptable models when considering the five top ranked predictions, but the top ranked model is not necessarily the best one. The quality and the selection of the models can be improved by forced sampling, i.e., re-running the AFM program many times with perturbed initial weights of the neural network, which require substantial GPU resources. Examples of applications include the design of PROteolysis TArgeting Chimeras (PROTACs) for modulating a target protein by degradation, the prediction of antibody epitopes, and searching for epitope-specific antibodies. Several tools will be combined for the structural analysis of protein interaction networks, including a novel method of constructing multi-protein complexes based on pre- calculated tables of interaction energies between pairs of proteins. To study the ligand binding properties of proteins we focus on binding hot spots, regions of proteins that are major contributors to the binding free energy. Our FTMap server globally samples the surface of target proteins using fragment sized molecular probes and provides reliable hot spot and pharmacophore information. We are in the process of improving the scoring function using neural networks, and substantially expanded the set of probes to obtain generalized pharmacophores that identify regions in the protein binding site with preferences for specific functional groups and a number of bound fragments. This can be achieved by iterative mapping, first using the standard probe set, and then adding specifically selected probes from a large library for functional characterization of the site, again increasing the need for computer power. Since the bound fragment positions can be used to find larger ligands, the goal is to convert FTMap into a fragment-based ligand discovery platform. We will also improve our template- based server LigTBM, which docks small molecules to proteins, and will integrate models based on machine learning with FTMap.