Improving AlphaFold predicted contacts inalpha-helical transmembrane proteins structuresusing structural features

Background: Residue contacts maps offer a 2-d reduced representation of 3-dprotein structures and constitute a structural constraint and scaffold in structuralmodeling. In addition, contact maps are also an effective tool in identifying interhelicalbinding sites and drawing insights about protein function. While mostworks predict contact maps using features derived from sequences, we believeinformation from known structures can be leveraged for a prediction improvementin unknown structures where decent approximate structures such as onespredicted by AlphaFold2 are available. Results: Alphafold2’s predicted structures are found to be quite accurate atinter-helical residue contact prediction task, achieving 83% average precision. Weadopt an unconventional approach, using features extracted from atomic structuresin the neighborhood of a residue pair and use them to predicting residuecontact. We trained on features derived from experimentally determined structuresand predicted on features derived from AlphaFold2’s predicted structures.Our results demonstrate a remarkable improvement over AlphaFold2 achievingover 91.9% average precision for held-out and over 89.5% average precision incross validation experiments. Conclusion: Training on features generated from experimentally determinedstructures, we were able to leverage knowledge from known structures to significantlyimprove the contacts predicted using AlphaFold2 structures. Wedemonstrated that using coordinates directly (instead of the proposed features)does not lead to an improvement in contact prediction performance.