Uncovering slow collective variables (CVs) of self-assembly dynamics is important to elucidate its numerous kinetic assembly pathways and drive the design of novel structures for advanced materials through the bottom-up approach. However, identifying the CVs for self-assembly presents several challenges. First, self-assembly systems often consist of identical monomers, and the feature representations should be invariant to permutations and rotational symmetries. Physical coordinates, such as aggregate size, lack high-resolution detail, while common geometric coordinates like pairwise distances are hindered by the permutation and rotational symmetry challenges. Second, self-assembly is usually a downhill process, and the trajectories often suffer from insufficient sampling of backward transitions that correspond to the dissociation of self-assembled structures. Popular dimensionality reduction methods, such as time-structure independent component analysis, impose detailed balance constraints, potentially obscuring the true dynamics of self-assembly. In this work, we employ GraphVAMPnets, which combines graph neural networks with a variational approach for Markovian process (VAMP) theory to identify the slow CVs of the self-assembly processes. First, GraphVAMPnets bears the advantages of graph neural networks, in which the graph embeddings can represent self-assembly structures in high-resolution while being invariant to permutations and rotational symmetries. Second, it is built upon VAMP theory, which studies Markov processes without forcing detailed balance constraints, which addresses the out-of-equilibrium challenge in the self-assembly process. We demonstrate GraphVAMPnets for identifying slow CVs of self-assembly kinetics in two systems: the aggregation of two hydrophobic molecules and the self-assembly of patchy particles. We expect that our GraphVAMPnets can be widely applied to molecular self-assembly.
29 June 2023 in Journal of Chemical Theory and Computation
Molecular docking, the task of predicting the binding structures between a protein and a small molecule ligand, plays a significant role in structural-based drug discovery. In recent years, numerous deep learning-based methods for molecular docking have emerged. State-of-the-art approaches such as DiffDock formulate the docking problem using diffusion generative models, exhibiting superior performance than traditional docking algorithms. However, despite the strong performance of these deep learning-based docking methods in predicting binding poses, they often lack a well-defined scoring function. This limitation poses challenges in effectively distinguishing between the strong and weak inhibitors during virtual screening. To address this limitation, we introduce FeatureDock, a transformer-based deep learning framework, which can accurately predict the protein-ligand binding poses as well as achieve a strong scoring power for virtual screening. FeatureDock extracts chemical features from local environments within protein structures and utilizes a Transformer encoder to predict probability density envelopes indicating where ligands are most likely to bind in the protein pocket. We also designed a scoring function, which encodes the predicted probability density envelope, to optimize and score the ligand poses. In addition, the attention mechanism in FeatureDock’s Transformer further enhances the model’s interpretability by providing the attention weights of each chemical feature from the protein structures in predicting the binding probabilities. When applied to virtual screening, we demonstrated that FeatureDock outperforms DiffDock, Smina and AutoDock Vina in distinguishing strong inhibitors from weak ones for both Cyclin-Dependent Kinase 2 (CDK2, an inactivated form) and Angiotensin-converting enzyme (ACE). The performance was assessed using Kullback–Leibler (KL) divergence and area under receiver operating characteristic (AUC) evaluation metrics. We also demonstrate that FeatureDock can accurately predict the binding poses, achieving an average RMSD of 2.4 Å when compared to CDK2-ligand co-crystal structures. We anticipate that our FeatureDock holds promise to be widely applied in virtual screening to assist in drug design. FeatureDock is available at https://github.com/xuhuihuang/featuredock.
Identifying collective variables (CVs) that accurately capture the slowest timescales of protein conformational changes is crucial to comprehend numerous biological processes. In this work, we develop a novel algorithm, the Memory kErnel Minimization based Neural Networks (MEMnets), that accurately identifies the slow CVs of biomolecular dynamics. MEMnets is distinct from popular deep-learning approaches (such as VAMPnets or SRVs) that assume Markovian dynamics. Instead, MEMnets is built on the integrative generalized master equation (IGME) theory, which incorporates non-Markovian dynamics by encoding them in a memory kernel for continuous CVs. The key innovation of MEMnets is to identify optimal CVs by minimizing time-integrated memory kernels. To accomplish this, MEMnets process time sequences of molecular dynamics (MD) conformations by using parallel encoder neural networks that project high-dimensional MD data into a low-dimensional latent space. The time-integrated memory kernels, derived from IGME theory, are then computed in the latent space as the objective function. We demonstrate that our MEMnets algorithm can effectively identify the slow CVs involved in the folding of FIP35 WW-domain with high accuracy, and successfully reveal two parallel folding pathways. Furthermore, we test MEMnets’ on the clamp opening of a bacterial RNA polymerase (RNAP), a much more complex conformational change (a system containing over 540K atoms), where sampling from all-atom MD simulations is limited. Our results demonstrate that MEMnets greatly outperforms SRVs, which is based on Markovian dynamics and may result in disconnected dynamics along the identified CVs. We anticipate that MEMnets holds promise to be widely to study biomolecular conformational changes.
Uncovering slow collective variables (CVs) of self-assembly dynamics is important to elucidate its numerous kinetic assembly pathways and drive the design of novel structures for advanced materials through the bottom-up approach. However, identifying the CVs for self-assembly presents several challenges. First, self-assembly systems often consist of identical monomers and the feature representations should be invariant to permutations and rotational symmetries. Physical coordinates, such as aggregate size, lack the high-resolution detail, while common geometric coordinates like pairwise distances are hindered by the permutation and rotational symmetry challenge. Second, self-assembly is usually a downhill process, and the trajectories often suffer from insufficient sampling of backward transitions that correspond to the dissociation of self-assembled structures. Popular dimensionality reduction methods, like tICA, impose detailed balance constraints, potentially obscuring the true dynamics of self-assembly. In this work, we employ GraphVAMPnets which combines graph neural networks with variational approach for Markovian process (VAMP) theory to identify the slow CVs of the self-assembly processes. First, GraphVAMPnets bears the advantages of graph neural networks, in which the graph embeddings can represent self-assembly structures in a high-resolution while being invariant to permutations and rotational symmetries. Second, it is built upon VAMP theory that studies Markov processes without forcing detailed balance constraint, which addresses the out-of-equilibrium challenge in self-assembly process. We demonstrate GraphVAMPnets for identifying slow CVs of self-assembly kinetics in two systems: aggregation of two hydrophobic molecules and self-assembly of patchy particles. We expect that our GraphVAMPnets can be widely to applied to molecular self-assembly.
Machine learning (ML) is rapidly reshaping the chemical sciences, with applications spanning molecular property prediction, chemical reaction design, molecular structure generation, and other data-driven discovery. With the growing integration of ML into chemical research, undergraduate chemistry students increasingly need training that bridges traditional chemical education with ML methods. Here we present Machine Learning in Chemistry (MLChem), an undergraduate-level course designed with a chemistry-first perspective to lower barriers to entry into ML while maintaining disciplinary relevance. This course introduces fundamental ML algorithms using chemical datasets, such as the small molecule solubility dataset, and the peptide activity dataset. It progresses from traditional ML algorithms to neural networks. Each chapter is accompanied by tutorial notebooks and homework assignments focused on chemistry-relevant tasks. These course materials are open-source and available at https://xuhuihuang.github.io/mlchem. These fundamental chapters are also complemented by advanced modules on emerging topics such as reinforcement learning for retrosynthesis, ML-based force fields, deep learning for the predictions of protein structure and dynamics. By combining chemical context with hands-on coding and exposure to frontier applications, MLChem equips undergraduate chemistry students with both conceptual foundations and practical skills, preparing them to participate in ML-driven chemical research.