Missing data on which proteins do not interact can be recovered from existing experiments.

Inside our cells, proteins bump into each other and work together. Scientists call these connections protein-protein interactions. Millions of these links have been tracked, but the overall map has major issues. It is incomplete, heavily biased, and packed with false alarms. To fix this, scientists need to know which proteins do not interact, but experiments almost never report these non-interactions. The researchers built a method to uncover this missing negative data from existing studies. Their method reconstructs the original experimental space that past tests actually checked. Next, they made a model that calculates how likely a link is real based on how often it was tested and observed. These reconstructed non-interacting pairs help calculate error rates and calibrate screens. Using this negative data to train a simple machine learning model instead of random protein pairs helps build a far more accurate map.

Why it matters

Researchers mapping protein interactions currently rely on reported positive interactions, but many are false and the network is incomplete. This work shows that non-interacting pairs, typically ignored, can be reconstructed from existing data. Using these negatives improves prediction accuracy and reduces bias, leading to a more reliable map of human protein interactions crucial for understanding disease and developing therapies.

Questions answered