Predicted protein structures from AI tools like AlphaFold2 are used as alternatives to experimental data, but accuracy suffers when predicted structures are used.
This study investigates this issue, attributing performance decrease to structure embedding bias in structure representation learning. The study introduces a Protein 3D Graph Structure Learning Problem for Robust Protein Property Prediction (PGSL-RP3) and proposes a Structure embedding Alignment Optimization framework (SAO) to address the bias.
The framework seems to be effective in improving property prediction for both predicted and experimental structures.