Drug Molecular Conformation Prediction Based on Two-way Constraint Model

In this paper, a drug molecule conformation prediction method based on a two-way constraint model is proposed, aiming to cope with the complexity and challenges in the drug design process. The model first maps the drug molecular map data into a matrix network and pre-trains it with random masks to adapt to the prediction problem with a small sample size, and employs a visual transformer (ViT) as the backbone network to take advantage of its powerful global modeling capability. Next, the raw data are fed into the decoupled network, which is mapped into structure vectors and action vectors by two multilayer perceptrons (MLPs) for different constraint processing tasks, respectively. The structure features focus on the different permutations of atoms on the surface, while the action features are mainly concerned with the interaction patterns between the drug and the target. In addition, we design a spatially enhanced predictor to further improve the prediction performance by secondary extraction of the reorganized structure vectors and action vectors. The experimental results show that on two real datasets, DrugBank and TWOSIDES, our model demonstrates significant advantages in evaluation metrics such as accuracy (ACC), area under the curve (AUC), and F1 score compared to multiple mainstream prediction models including CNN, RNN, LSTM, GRU, GNN, and ViT, with 92.5% vs. 88.5%-90.2% (Dataset1) and 93.9% vs. 82.7%-92.6% (Dataset2), demonstrating the effectiveness and superiority of the proposed two-way constraint model in drug molecular conformation prediction, and providing a new strategy and technological means for future drug discovery.