A novel method combining covariance analysis with global inference via message passing is introduced to identify direct residue contacts in protein-protein interactions. The method distinguishes between directly and indirectly correlated residues, which is crucial for understanding the molecular determinants of protein-protein interactions. Applied to bacterial two-component signaling systems, the method successfully identifies residue pairs that are in close spatial proximity without requiring ad hoc tuning parameters. It is effective for both hetero-interactions between sensor kinase (SK) and response regulator (RR) proteins and homo-interactions between RR proteins. The method uses mutual information (MI) and direct information (DI) to assess the strength of correlations, with DI being a more reliable indicator of direct residue contacts than MI. The method was validated against structural data, showing that DI values closely match the actual distances between residues in the co-crystal structure of Spo0B/Spo0F. The method is applicable to proteins present in a single copy per genome, which is increasingly common as the number of sequenced genomes grows. The approach has potential applications in drug design and understanding protein interaction mechanisms. The method is robust to sampling biases and can be used to predict protein-protein interaction surfaces, aiding in the design of synthetic proteins with specific interactions. The study highlights the importance of distinguishing direct from indirect interactions in protein-protein interactions and demonstrates the effectiveness of the message-passing approach in achieving this.A novel method combining covariance analysis with global inference via message passing is introduced to identify direct residue contacts in protein-protein interactions. The method distinguishes between directly and indirectly correlated residues, which is crucial for understanding the molecular determinants of protein-protein interactions. Applied to bacterial two-component signaling systems, the method successfully identifies residue pairs that are in close spatial proximity without requiring ad hoc tuning parameters. It is effective for both hetero-interactions between sensor kinase (SK) and response regulator (RR) proteins and homo-interactions between RR proteins. The method uses mutual information (MI) and direct information (DI) to assess the strength of correlations, with DI being a more reliable indicator of direct residue contacts than MI. The method was validated against structural data, showing that DI values closely match the actual distances between residues in the co-crystal structure of Spo0B/Spo0F. The method is applicable to proteins present in a single copy per genome, which is increasingly common as the number of sequenced genomes grows. The approach has potential applications in drug design and understanding protein interaction mechanisms. The method is robust to sampling biases and can be used to predict protein-protein interaction surfaces, aiding in the design of synthetic proteins with specific interactions. The study highlights the importance of distinguishing direct from indirect interactions in protein-protein interactions and demonstrates the effectiveness of the message-passing approach in achieving this.