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
[1]  
Aghajanyan Armen, 2022, arXiv
[2]  
Aghajanyan Armen, 2020, arXiv
[3]  
[Anonymous], 2012, EXTRACTION CHEM STRU
[4]  
[Anonymous], 2014, arXiv
[5]  
Baines M., 2021, FAIRSCALE GEN PURPOS
[6]  
Bao H., 2021, ARXIV
[7]   The properties of known drugs .1. Molecular frameworks [J].
Bemis, GW ;
Murcko, MA .
JOURNAL OF MEDICINAL CHEMISTRY, 1996, 39 (15) :2887-2893
[8]   Machine learning with physicochemical relationships: solubility prediction in organic solvents and water [J].
Boobier, Samuel ;
Hose, David R. J. ;
Blacker, A. John ;
Nguyen, Bao N. .
NATURE COMMUNICATIONS, 2020, 11 (01)
[9]  
Bradshaw J., 2018, ARXIV
[10]  
Chithrananda S., 2020, CHEMBERTA LARGE SCAL