O08_03
Conditional structure prediction of protein-compound complex
Atsuhiro TOMITA *
Drug Discovery, Preferred Networks, inc.
( * E-mail: atomita@preferred.jp)
Deep learning based protein structure prediction, such as Alpha Fold2 [1], has made it possible to predict protein structures with high accuracy. Most of these methods use multiple sequence alignment (MSA) as input and predict the protein structure. By manipulating this input MSA, researchers have solved various problems. For example, it is known that proteins can adopt various conformational states in vivo, but conventional prediction methods had the problem of predicting conformations that were biased toward a particular state. Then, it was reported that by directly modifying the input MSA, the biased conformation can be relaxed and the conformation of other states can be stochastically obtained [2]. In addition, another application of MSA engineering was reported: a method for predicting the structure of protein-compound complexes by extending MSA features to compounds other than amino acids [3]. With the advent of such methods, structure prediction can now be utilized in the drug discovery area, including low-molecular-weight compounds. However, the prediction accuracy for protein-compound complexes is not sufficient, and improved methods are still needed.
To improve this problem, we propose a method to optimize MSA features indirectly by incorporating external knowledge into the network of protein-compound complex prediction.
Recently, in order to efficiently predict a protein structure in a specific state, methods have been investigated that indirectly modify MSA features toward the desired state using user-defined constraints [4]. These methods sample the target state more efficiently than conventional methods that directly modify the MSA to sample the structure stochastically. We applied these indirect MSA modification methods to the prediction of 3D structures of protein-compound complexes.
We incorporated external knowledge such as pharmacophores and similarity to other protein-compound complexes into the structure prediction. As a result, we optimized the MSA feature to generate a complex structure that is consistent with both the structural validity and the introduced external knowledge. The results showed improvement in the prediction of the complex structure that were incorrect in the prediction without external knowledge. In this presentation, we report that our approach is effective in predicting the structure of protein-compound complexes, a field where few effective improvement methods have been reported so far.
[1] Jumper, J. et al. Highly accurate protein structure prediction with AlphaFold. Nature 596, 583–589 (2021)
[2] Diego, A. et al. Sampling alternative conformational states of transporters and receptors with AlphaFold2. eLife 11:e75751. (2022)
[3] Bryant P. et al. Structure prediction of protein-ligand complexes from sequence information with Umol. Nature Communications 15, Article number: 4536 (2024)
[4] Xie, T. et al. Conditioned Protein Structure Prediction. bioRxiv 2023.09.24.559171 (2023)