BartSmiles: Generative Masked Language Models for Molecular Representations

被引:5
作者
Chilingaryan, Gayane [1 ]
Tamoyan, Hovhannes [1 ]
Tevosyan, Ani [1 ,3 ]
Babayan, Nelly [2 ,3 ]
Hambardzumyan, Karen [1 ]
Navoyan, Zaven [3 ]
Aghajanyan, Armen [4 ]
Khachatrian, Hrant [1 ,5 ]
Khondkaryan, Lusine [2 ,3 ]
机构
[1] YerevaNN, Yerevan 0025, Armenia
[2] NAS RA, Inst Mol Biol, Yerevan 0014, Armenia
[3] Toxometris Ai, Yerevan 0019, Armenia
[4] Meta AI Res, Menlo Pk, CA 94025 USA
[5] Yerevan State Univ, Yerevan 0025, Armenia
关键词
STRUCTURAL ALERTS; PREDICTION; CHEMISTRY;
D O I
10.1021/acs.jcim.4c00512
中图分类号
R914 [药物化学];
学科分类号
100701 ;
摘要
We discover a robust self-supervised strategy tailored toward molecular representations for generative masked language models through a series of tailored, in-depth ablations. Using this pretraining strategy, we train BARTSmiles, a BART-like model with an order of magnitude more compute than previous self-supervised molecular representations. In-depth evaluations show that BARTSmiles consistently outperforms other self-supervised representations across classification, regression, and generation tasks, setting a new state-of-the-art on eight tasks. We then show that when applied to the molecular domain, the BART objective learns representations that implicitly encode our downstream tasks of interest. For example, by selecting seven neurons from a frozen BARTSmiles, we can obtain a model having performance within two percentage points of the full fine-tuned model on task Clintox. Lastly, we show that standard attribution interpretability methods, when applied to BARTSmiles, highlight certain substructures that chemists use to explain specific properties of molecules. The code and pretrained model are publicly available.
引用
收藏
页码:5832 / 5843
页数:12
相关论文
共 69 条
[11]   A graph-convolutional neural network model for the prediction of chemical reactivity [J].
Coley, Connor W. ;
Jin, Wengong ;
Rogers, Luke ;
Jamison, Timothy F. ;
Jaakkola, Tommi S. ;
Green, William H. ;
Barzilay, Regina ;
Jensen, Klavs F. .
CHEMICAL SCIENCE, 2019, 10 (02) :370-377
[12]  
Conneau A., 2020, arXiv
[13]   Computer says yes [J].
Davey, Stephen G. .
NATURE REVIEWS CHEMISTRY, 2018, 2 (04)
[14]  
Devlin Jacob, 2018, 181004805 ARXIV
[15]   Graph Transformation Policy Network for Chemical Reaction Prediction [J].
Do, Kien ;
Truyen Tran ;
Venkatesh, Svetha .
KDD'19: PROCEEDINGS OF THE 25TH ACM SIGKDD INTERNATIONAL CONFERENCCE ON KNOWLEDGE DISCOVERY AND DATA MINING, 2019, :750-760
[16]  
Duvenaud D., 2015, CONVOLUTIONAL NETWOR, P2224
[17]   In silico prediction of chemical genotoxicity using machine learning methods and structural alerts [J].
Fan, Defang ;
Yang, Hongbin ;
Li, Fuxing ;
Sun, Lixia ;
Di, Peiwen ;
Li, Weihua ;
Tang, Yun ;
Liu, Guixia .
TOXICOLOGY RESEARCH, 2018, 7 (02) :211-220
[18]  
Glen RC, 2006, IDRUGS, V9, P199
[19]  
Greg L., 2013, RDKIT ASOFTWARE SUIT
[20]   Benchmark Data Set for in Silico Prediction of Ames Mutagenicity [J].
Hansen, Katja ;
Mika, Sebastian ;
Schroeter, Timon ;
Sutter, Andreas ;
ter Laak, Antonius ;
Steger-Hartmann, Thomas ;
Heinrich, Nikolaus ;
Mueller, Klaus-Robert .
JOURNAL OF CHEMICAL INFORMATION AND MODELING, 2009, 49 (09) :2077-2081