AutoDock Vina is one of the most popular molecular docking tools. In the latest benchmark CASF-2016 for comparative assessment of scoring functions, AutoDock Vina won the best docking power among all the docking tools. Modern drug discovery is facing a common scenario of large virtual screening of drug hits from huge compound databases. Due to the seriality characteristic of the AutoDock Vina algorithm, there is no successful report on its parallel acceleration with GPUs. Current acceleration of AutoDock Vina typically relies on the stack of computing power as well as the allocation of resource and tasks, such as the VirtualFlow platform. The vast resource expenditure and the high access threshold of users will greatly limit the popularity of AutoDock Vina and the flexibility of its usage in modern drug discovery. In this work, we proposed a new method, Vina-GPU, for accelerating AutoDock Vina with GPUs, which is greatly needed for reducing the investment for large virtual screens and also for wider application in large-scale virtual screening on personal computers, station servers or cloud computing, etc. Our proposed method is based on a modified Monte Carlo using simulating annealing AI algorithm. It greatly raises the number of initial random conformations and reduces the search depth of each thread. Moreover, a classic optimizer named BFGS is adopted to optimize the ligand conformations during the docking progress, before a heterogeneous OpenCL implementation was developed to realize its parallel acceleration leveraging thousands of GPU cores. Large benchmark tests show that Vina-GPU reaches an average of 21-fold and a maximum of 50-fold docking acceleration against the original AutoDock Vina while ensuring their comparable docking accuracy, indicating its potential for pushing the popularization of AutoDock Vina in large virtual screens.
13 October 2023 in IEEE/ACM Transactions on Computational Biology and Bioinformatics
Simulated Annealing (SA) algorithm is not effective with large optimization problems for its slow convergence. Hence, several parallel Simulated Annealing (pSA) methods have been proposed, where the increase of searching threads can boost the speed of convergence. Although satisfactory solutions can be obtained by these methods, there is no rigorous mathematical analyses on their effectiveness. Thus, this paper introduces a probabilistic model, on which a theorem about the effectiveness of multiple initial states parallel SA (MISPSA) has been proven. The theorem also demonstrates that the increasing parallelism in pSA algorithm with the reducing of search depth in each thread could obtain almost the same probability of finding the global optimal solution. We validated our theorem on AutoDock Vina, a widely used molecular docking tool with high accuracy and docking speed. AutoDock Vina uses a pSA strategy to find optimal molecular conformations. Under the premise that the total searching workload (i.e., thread number * iteration depth of each thread) remains unchanged, the docking accuracy from an aggressively parallelized SA searching method is almost the same or even better than those from the default exhaustiveness (parallelism degree) configuration of AutoDock Vina. Taking complex '1hnn' as an example,with the increase (125x) in the number of initial states (from 8 to 1000) and the decrease in the search depth for each thread (from 15540 to 124, or 1/125 of the original search depth), the mean energy is -7.80 and -7.94, while the mean RMSD is 3.4 and 3.14, respectively. The result also implies that a considerable speedup (in this case 125x in theory) can be obtained by a highly parallelized SA algorithm implementation.
Ligand molecules naturally constitute a graph structure. Recently, many excellent deep graph learning (DGL) methods have been proposed and used to model ligand bioactivities, which is critical for the virtual screening of drug hits from compound databases in interest. However, pharmacists can find that these well-trained DGL models usually are hard to achieve satisfying performance in real scenarios for virtual screening of drug candidates. The main challenges involve that the datasets for training models were small-sized and biased, and the inner active cliff cases would worsen model performance. These challenges would cause predictors to overfit the training data and have poor generalization in real virtual screening scenarios. Thus, we proposed a novel algorithm named adversarial feature subspace enhancement (AFSE). AFSE dynamically generates abundant representations in new feature subspace via bi-directional adversarial learning, and then minimizes the maximum loss of molecular divergence and bioactivity to ensure local smoothness of model outputs and significantly enhance the generalization of DGL models in predicting ligand bioactivities. Benchmark tests were implemented on seven state-of-the-art open-source DGL models with the potential of modeling ligand bioactivities, and precisely evaluated by multiple criteria. The results indicate that, on almost all 33 GPCRs datasets and seven DGL models, AFSE greatly improved their enhancement factor (top-10%, 20% and 30%), which is the most important evaluation in virtual screening of hits from compound databases, while ensuring the superior performance on RMSE and $r^2$. The web server of AFSE is freely available at http://noveldelta.com/AFSE for academic purposes.
18 July 2024 in IEEE Transactions on Vehicular Technology
To address the challenge of optimizing power distribution for transmitter-receiver pairs within ad hoc wireless networks, deep learning techniques have been employed to navigate the complexities of evolving network topologies and variable channel characteristics. This approach has necessitated the use of extensive training data sets due to challenges in interpretability. Moreover, there's an anticipation for the deep learning model to unearth more generalized and distinct features. In response to these challenges, we introduce the Unsupervised Power Allocation based on Attentive Graph Representation (UPAR) method. UPAR combines an Edge Aggregated Graph Attention Regression Network (EGRN) with Deep Unfolded Weighted Minimum Mean Square Error (DU-WMMSE) within an unsupervised learning framework. Initially, EGRN leverages an attention mechanism and Gate Recurrent Units (GRU) to amalgamate features from adjacent nodes and their connecting edges, enhancing the identifiability and generalization of the graph's features. Subsequently, DU-WMMSE applies unsupervised learning to refine these graph features through regression training, transforming WMMSE iterations into efficient, interpretable layers within the neural network. This process also introduces adjustable parameters within the network, employing both feed forward and back propagation phases during end-to-end training. This method significantly narrows the solution space for the objective function in an unsupervised manner, thereby accelerating the training process and enhancing the efficiency of the entire deep learning model. Extensive comparative and ablation studies confirm the UPAR method's superior performance and adaptability across a variety of scenarios.
7 October 2021 in Journal of Chemical Information and Modeling