Protein-protein interactions are central in many biological processes, but they are difficult to characterize, especially in complex samples. Currently, protein cross-linking combined with mass spectrometry and computational modeling is gaining increased recognition as a viable tool in protein interaction studies. Here, we present TX-MS, a targeted approach that combines the sensitivity of chemical cross-linking MS (XL-MS) with the power of computational modeling to rapidly determine the quaternary structures of large protein complexes using data collected on unfractionated and biologically relevant samples. TX-MS is using artificial intelligence to combine three different MS acquisition methods (high-resolution MS1 (hrMS1), data-dependent acquisition (DDA) and data-independent acquisition (DIA)) and computational protein structure modeling. Accordingly, Rosetta docking protocol (RosettaDock) was used to produce a compendium of quaternary protein structure models; the models were then ranked using an ensemble-based machine learning algorithm using hrMS1 data revealing potential binding interfaces. The best models were iteratively improved by analyzing DDA data to find fragments of XLs within cross-linkable distance. A high-resolution phase of modeling adjusted the final model which then was validated using Sequential Windowed Acquisition of All Theoretical Fragment Ion Mass Spectra (SWATH-MS). Here, we demonstrated the applicability of TX-MS by studying two large multi-species macromolecular assemblies by analyzing a single experimental data provided by cross-linking intact bacteria in human plasma proteins with disuccinimidyl suberate (DSS) that specifically cross-links lysine residues. First, we elucidated the interactome of M1 protein as the most important virulence factor of Group A streptococcus with 15 different human plasma proteins over the surface of bacteria. This 1.8MDa large complex is supported by more than 200 XLs (an average of 13 distance constraints per interaction), revealing many new crucial interactions for bacterial survival such as the interactions of M1 with human albumin, IgG and C4BpA. Secondly, we detected the Membrane Attack Complex (MAC) in the same sample as the MAC composing components existed in large quantities. We proposed a near to atomic resolution model supported by more than 170 XLs with an average of more than 10 distance constraints per interface. TX-MS is, to our knowledge, the only method capable of determining the quaternary protein structure of large protein complexes using data from unfractionated samples. TX-MS eliminates the limitation of traditional methods by combining flexible computational models in an intelligent way. A machine learning algorithm guide the process of selecting the potential binding interfaces according to their isotopic patterns derived from hrMS1 data. This strategy reduces the computational complexity and the conformational space of the problem by order of magnitudes. Combining two more MS acquisition approaches (DDA and DIA) help to adjust and validate the final model in an iterative fashion gaining more speed and accuracy in complex samples.