Article
Chemistry, Multidisciplinary
Ava P. Soleimany, Alexander Amini, Samuel Goldman, Daniela Rus, Sangeeta N. Bhatia, Connor W. Coley
Summary: This paper introduces a new approach to uncertainty quantification for neural network-based molecular structure-property prediction using evidential deep learning, which enables calibrated predictions, sample-efficient training, and improved experimental validation rates in the chemical and physical sciences.
ACS CENTRAL SCIENCE
(2021)
Article
Biochemical Research Methods
Kyung Pyo Ham, Lee Sael
Summary: This article proposes a method called Evidential Meta-model for Molecular Property Prediction (EM3P2) that returns uncertainty estimates and improves prediction performance. By training an evidential graph isomorphism network classifier using multi-task molecular property datasets under the model-agnostic meta-learning framework, the problem of data imbalance is addressed. The uncertainty estimates can be used to reject uncertain predictions in applications requiring higher confidence.
Article
Biochemistry & Molecular Biology
Fuhao Zhang, Min Li, Jian Zhang, Lukasz Kurgan
Summary: This study investigates sequence-based predictors of RNA-binding residues (RBRs) and finds that structure-trained predictors perform well for structure-annotated proteins, while disorder-trained predictors provide accurate results for disorder-annotated proteins. However, these methods do not perform as well on opposite annotations, leading to the development of a new integrated model for improved prediction accuracy.
NUCLEIC ACIDS RESEARCH
(2023)
Article
Biochemical Research Methods
Hehuan Ma, Yatao Bian, Yu Rong, Wenbing Huang, Tingyang Xu, Weiyang Xie, Geyan Ye, Junzhou Huang
Summary: This study explores a multi-view modeling approach (MVGNN) for molecular property prediction using graph neural networks (GNN). The approach considers both atom and bond information to provide more accurate predictions. The model's expressive power is enhanced through a cross-dependent message-passing scheme. Experimental results demonstrate that this approach outperforms state-of-the-art models on various benchmark tests.
Article
Biochemical Research Methods
Ailin Xie, Ziqiao Zhang, Jihong Guan, Shuigeng Zhou
Summary: This paper introduces a chemistry-aware fragmentation model for molecular property prediction, which utilizes contrastive learning to extract representations of molecular structures and achieves good predictive performance.
BRIEFINGS IN BIOINFORMATICS
(2023)
Article
Biochemical Research Methods
Ziqiao Zhang, Jihong Guan, Shuigeng Zhou
Summary: In this article, a fragment-oriented multi-scale graph attention network (FraGAT) is proposed for molecular property prediction, and experimental results show that it achieves state-of-the-art predictive performance in most cases, demonstrating the interpretability of the model.
Review
Pharmacology & Pharmacy
Zhen Li, Mingjian Jiang, Shuang Wang, Shugang Zhang
Summary: This review summarizes the recent applications of deep learning methods in molecular representation and property prediction. The DL methods are categorized according to the format of molecular data and common models like ensemble learning and transfer learning are discussed. The interpretability methods for these models are also analyzed, and the challenges and opportunities for DL methods in molecular representation and property prediction are highlighted.
DRUG DISCOVERY TODAY
(2022)
Article
Biochemistry & Molecular Biology
Jiahui Zhang, Wenjie Du, Xiaoting Yang, Di Wu, Jiahe Li, Kun Wang, Yang Wang
Summary: This paper proposes a new model called SMG-BERT for accurate and fast prediction of molecular properties by integrating 3D geometric parameters, 2D topological information, and 1D SMILES string into the self-attention-based BERT model. Experimental results demonstrate that SMG-BERT outperforms existing methods on 12 benchmark molecular datasets, showing good generalization and reliability.
FRONTIERS IN MOLECULAR BIOSCIENCES
(2023)
Article
Biochemical Research Methods
Xiao-Chen Zhang, Cheng-Kun Wu, Zhi-Jiang Yang, Zhen-Xing Wu, Jia-Cai Yi, Chang-Yu Hsieh, Ting-Jun Hou, Dong-Sheng Cao
Summary: This study introduces a molecular graph BERT (MG-BERT) model that integrates graph neural network mechanisms and utilizes a self-supervised learning strategy for pretraining, enhancing the model's contextual sensitivity and achieving outstanding performance in molecular property prediction.
BRIEFINGS IN BIOINFORMATICS
(2021)
Article
Chemistry, Multidisciplinary
Xuan Zang, Xianbing Zhao, Buzhou Tang
Summary: The authors develop a hierarchical molecular graph self-supervised learning framework to learn molecule representation for property prediction. This framework improves the encoding of structural information and chemical functions in molecular graphs.
COMMUNICATIONS CHEMISTRY
(2023)
Article
Chemistry, Multidisciplinary
Yuanbing Song, Jinghua Chen, Wenju Wang, Gang Chen, Zhichong Ma
Summary: This paper proposes an end-to-end double-head transformer neural network (DHTNN) for high-precision molecular property prediction, addressing the limitations of existing deep learning methods in the generalization ability of the nonlinear representation of molecular features and the reasonable assignment of feature weights. It significantly improves the accuracy of molecular property prediction.
JOURNAL OF CHEMINFORMATICS
(2023)
Article
Biology
Patrice Monkam, Songbai Jin, Wenkai Lu
Summary: This study proposes a two-phase framework for fast generation of annotated echo data, showing significant performance improvement when using generated realistic annotated images for pretraining deep learning models.
COMPUTERS IN BIOLOGY AND MEDICINE
(2022)
Article
Chemistry, Multidisciplinary
Yinghui Jiang, Shuting Jin, Xurui Jin, Xianglu Xiao, Wenfan Wu, Xiangrong Liu, Qiang Zhang, Xiangxiang Zeng, Guang Yang, Zhangming Niu
Summary: The authors propose a multi-level Pharmacophoric-constrained Heterogeneous Graph Transformer (PharmHGT) to capture the pharmacophore structure and chemical information. Their model achieves superior performance on molecular properties prediction and better representation capacity.
COMMUNICATIONS CHEMISTRY
(2023)
Article
Chemistry, Multidisciplinary
Yang Zhou, Nan He, Zheshuai Lin, Xiaoying Shang, Xueyuan Chen, Yanqiang Li, Weiqi Huang, Maochun Hong, Sangen Zhao, Junhua Luo
Summary: This study reports a non-pi-conjugated molecular crystal, NH3BH3, which achieves a subtle balance between second-harmonic generation, bandgap, and birefringence. NH3BH3 exhibits a large second-harmonic generation response, deep-UV transparency, and moderate birefringence. It is considered a promising candidate for deep-UV nonlinear optical applications.
Article
Biochemical Research Methods
Shen Han, Haitao Fu, Yuyang Wu, Ganglan Zhao, Zhenyu Song, Feng Huang, Zhongfei Zhang, Shichao Liu, Wen Zhang
Summary: Accurate prediction of molecular properties is important in drug discovery. Previous works have developed various representation schemes for capturing chemical information in molecules. This study proposes a novel framework, HimGNN, which combines atom- and motif-based graphs to learn hierarchical molecular topology representations. HimGNN achieves promising performances on classification and regression tasks in molecular property prediction.
BRIEFINGS IN BIOINFORMATICS
(2023)
Article
Chemistry, Multidisciplinary
Hung V. -T. Nguyen, Yivan Jiang, Somesh Mohapatra, Wencong Wang, Jonathan C. Barnes, Nathan J. Oldenhuis, Kathleen K. Chen, Simon Axelrod, Zhihao Huang, Qixian Chen, Matthew R. Golder, Katherine Young, Dylan Suvlu, Yizhi Shen, Adam P. Willard, Michael J. A. Hore, Rafael Gomez-Bombarelli, Jeremiah A. Johnson
Summary: This study synthesized water-soluble chiral bottlebrush polymers using macromonomers of different rigidity, and discovered that polymers with conformationally flexible mirror image side chains exhibited significant differences in properties compared to those with comparably rigid mirror image side chains. The observations were rationalized by correlating greater conformational freedom with enhanced chiral recognition, providing insights for the design of future biomaterials.
Article
Chemistry, Multidisciplinary
Kevin P. Greenman, William H. Green, Rafael Gomez-Bombarelli
Summary: Optical properties play a central role in molecular design for various applications, with existing theoretical and statistical methods balancing accuracy, generality, and cost. This study utilizes neural networks to predict molecular absorption peaks in solution, achieving higher accuracy and generalizability through a multi-fidelity approach based on an auxiliary model.
Article
Chemistry, Multidisciplinary
Estefania Bello-Jurado, Daniel Schwalbe-Koda, Mathias Nero, Cecilia Paris, Toni Uusimaki, Yuriy Roman-Leshkov, Avelino Corma, Tom Willhammar, Rafael Gomez-Bombarelli, Manuel Moliner
Summary: A novel methodology based on high-throughput simulations has been developed to design unique biselective organic structure-directing agents (OSDAs) that enable the efficient synthesis of CHA/AEI zeolite intergrowth materials with controlled phase compositions. These materials exhibit outstanding catalytic performance and hydrothermal stability, surpassing even the performance of commercial CHA-type catalysts. This methodology opens up possibilities for synthesizing new zeolite intergrowth materials with more complex structures and unique catalytic properties.
ANGEWANDTE CHEMIE-INTERNATIONAL EDITION
(2022)
Article
Chemistry, Physical
James Damewood, Daniel Schwalbe-Koda, Rafael Gomez-Bombarelli
Summary: Efficient and accurate calculation of thermodynamic potentials and observables is crucial for the application of statistical mechanics simulations in materials science. Existing naive Monte Carlo methods cannot handle the calculation demands of complex materials, so we transform machine learning-based generative models into the semi-grand canonical ensemble to address this issue. The resulting models are transferable across different thermodynamic conditions and can be used with various internal energy models.
NPJ COMPUTATIONAL MATERIALS
(2022)
Article
Chemistry, Physical
Wujie Wang, Zhenghao Wu, Johannes C. B. Dietschreit, Rafael Gomez-Bombarelli
Summary: In this study, a general stochastic method called DiffSim is proposed to learn pair interactions from data using differentiable simulations. The method uses molecular dynamics simulations and stochastic gradient descent to directly learn interaction potentials based on structural observables. DiffSim is flexible and can simultaneously simulate and optimize multiple systems, such as different temperatures or compositions. The results show that DiffSim can explore a wider functional space of pair potentials compared to traditional methods like iterative Boltzmann inversion. The methods can also be used to simultaneously fit potentials for simulations at different compositions and temperatures to improve transferability.
JOURNAL OF CHEMICAL PHYSICS
(2023)
Article
Chemistry, Multidisciplinary
Simon Axelrod, Eugene Shakhnovich, Rafael Gomez-Bombarelli
Summary: This article introduces a computational tool for predicting the thermal half-lives of azobenzene derivatives, which are key photoswitches in light-activated drugs. Through machine learning and quantum chemistry data, the authors automated the prediction of thermal half-lives for 19,000 azobenzene derivatives and explored trends and trade-offs between barriers and absorption wavelengths.
Article
Chemistry, Multidisciplinary
Gabriel Bradford, Jeffrey Lopez, Jurgis Ruza, Michael A. Stolberg, Richard Osterude, Jeremiah A. Johnson, Rafael Gomez-Bombarelli, Yang Shao-Horn
Summary: Solid polymer electrolytes (SPEs) have the potential to improve lithium-ion batteries by enhancing safety and enabling higher energy densities. A chemistry-informed machine learning model was developed to predict ionic conductivity of SPEs, using data from hundreds of experimental publications. The model encodes the Arrhenius equation into the readout layer of a neural network and has improved accuracy in predicting ionic conductivity.
ACS CENTRAL SCIENCE
(2023)
Article
Materials Science, Multidisciplinary
Aik Rui Tan, Shingo Urata, Masatsugu Yamada, Rafael Gomez-Bombarelli
Summary: Analyzing the atomic structure of glassy materials is challenging, but using a graph-theoretical approach can help understand the topological differences between disordered structural arrangements. By comparing different thermodynamic states of silica glass, it was found that silica glasses exhibit distinct topological features at temperatures higher than the fictive temperature. Graph-based analysis suggests that the anomalous density behavior of silica glass may be attributed to the increased formation of oxygen triclusters and reduced number of larger sized cycles at the density minimum temperature.
COMPUTATIONAL MATERIALS SCIENCE
(2023)
Article
Multidisciplinary Sciences
Pau Ferri, Chengeng Li, Daniel Schwalbe-Koda, Mingrou Xie, Manuel Moliner, Rafael Gomez-Bombarelli, Mercedes Boronat, Avelino Corma
Summary: Approaching the level of molecular recognition of enzymes with solid catalysts is a challenging goal, achieved in this work for the competing transalkylation and disproportionation of diethylbenzene catalyzed by acid zeolites. The key diaryl intermediates for the two competing reactions only differ in the number of ethyl substituents in the aromatic rings, and therefore finding a selective zeolite able to recognize this subtle difference requires an accurate balance of the stabilization of reaction intermediates and transition states inside the zeolite microporous voids.
NATURE COMMUNICATIONS
(2023)
Article
Energy & Fuels
Gavin Winter, Rafael Gomez-Bombarelli
Summary: Li10Ge(PS6)(2) (LGPS) is a highly concentrated solid electrolyte with Coulombic repulsion between neighboring cations hypothesized as the reason for ion hopping mechanism. By using a neural network potential trained on density functional theory (DFT) simulations, MD simulations were conducted to study ion conduction mechanisms at a range of temperatures including previous simulations and experimental studies. The results showed a Li+ sublattice phase transition in LGPS near 400 K which drastically reduced the ab-plane diffusivity. The sublattice phase transition was accompanied by less cation-cation correlation and more harmonic vibrations at lower temperature, indicating slower ion conduction.
JOURNAL OF PHYSICS-ENERGY
(2023)
Article
Chemistry, Multidisciplinary
Simon Axelrod, Eugene Shakhnovich, Rafael Gomez-Bombarelli
Summary: Molecular photoswitches, such as azobenzene, play a crucial role in light-activated drugs. This study presents a computational tool for predicting the thermal half-lives of azobenzene derivatives, using a fast and accurate machine learning potential trained on quantum chemistry data. The research explores trends and trade-offs between barriers and absorption wavelengths, and provides open access to the data and software for further research in photopharmacology.
ACS CENTRAL SCIENCE
(2023)
Proceedings Paper
Computer Science, Artificial Intelligence
Wujie Wang, Minkai Xu, Chen Cai, Benjamin Kurt Miller, Tess Smidt, Yusu Wang, Jian Tang, Rafael Gomez-Bombarelli
Summary: In this paper, a novel model is proposed for reconstructing fine-grained coordinates from coarse-grained coordinates. The model encodes the uncertainties of the fine-grained representation into a latent space and decodes them back to fine-grained geometries using equivariant convolutions. Experimental results demonstrate that this approach can recover more realistic structures and outperforms existing data-driven methods.
INTERNATIONAL CONFERENCE ON MACHINE LEARNING, VOL 162
(2022)
Article
Chemistry, Multidisciplinary
Simon Axelrod, Daniel Schwalbe-Koda, Somesh Mohapatra, James Damewood, Kevin P. Greenman, Rafael Gomez-Bombarelli
Summary: Designing new materials is crucial for addressing societal challenges, and computational techniques such as atomistic simulation and machine learning (ML) offer an avenue for rapid material invention. This article reviews the recent contributions of simulation and ML in materials design, discussing numerical representation of materials, ML methods for enhancing atomistic simulation, and high-throughput virtual screening. The limitations of ML and simulation are also discussed.
ACCOUNTS OF MATERIALS RESEARCH
(2022)
Article
Computer Science, Artificial Intelligence
Somesh Mohapatra, Joyce An, Rafael Gomez-Bombarelli
Summary: This article presents the development of a chemistry-informed graph representation of macromolecules, allowing for quantification of structural similarity and interpretable supervised learning. It enables quantitative chemistry-informed decision-making and iterative design in the macromolecular chemical space.
MACHINE LEARNING-SCIENCE AND TECHNOLOGY
(2022)