抱歉,您的浏览器无法访问本站
本页面需要浏览器支持(启用)JavaScript
了解详情 >

今天看了一篇综述,受益良多,对一些内容做一下总结

论文链接:

Github地址:

1. 摘要

人工智能(AI)在过去十年中一直在改变着药物发现(drug discovery)的实践。各种人工智能技术已被用于广泛的应用中,如虚拟筛选和药物设计。在这项调查中,我们首先对药物发现进行了概述,并讨论了相关的应用,这些应用可以简化为两个主要任务,即分子性质预测和分子生成。此外,为了总结人工智能在药物发现方面的进展,我们介绍了相关的人工智能技术,包括所调查论文中的模型架构和学习范式。我们希望这项调查能够为那些有兴趣在AI+药物发现领域的研究人员提供指导。我们还提供了一个GitHub资源库https://github.com/dengjianyuan/Survey_AI_Drug_Discovery,其中收集了论文和代码,作为学习资源,定期更新。

2. 关键概念定义

Table 1 关键概念定义

Term Description
Drug Discovery (药物发现) 药物发现是在没有药物治疗某疾病或现有药物疗效有限(和/或)毒性严重的情况下进行的项目。
leads drug candidates
High-Throughput Screening (HTS) a hit-finding approach underpinned by development in automation and the availability of large chemical libraries
SAR structure-activity relationship
Virtual Screening (VS) Various computational techniques have been developed to search the large chemical libraries for potentially active molecules to be tested in subsequent in vitro and in vivo assays.
structure-based VS based on knowledge about the target -> increase the odds of identifying active molecules.
ligand-based VS based on knowledge about known active ligands -> increase the odds of identifying active molecules.
agonist 一种作为内源性配体激活靶点并发挥生物反应的分子
antagonist 结合靶标以抑制反应的分子。
physicochemical property water solubility, acid-base dissociation constant, lipophilicity, permeability
pharmacokinetic property absorption, distribution, metabolism, excretion
pharmacodynamic property activity, selectivity
Synthetic Accessibility Score (SAS) SAS is a heuristic score (10-1) of how hard or easy it is to synthesize a given molecule based on a combination of the molecular fragments’ contributions.
Quantitative Estimation of Drug-likeness (QED) QED is an estimate (0-1) on how likely a molecule is a viable drug candidate.
quantitative structure-activity relationship (QSAR) 对于每个感兴趣的属性,建立一个预测模型,通过分类或回归将分子结构映射到属性值。
design-make-test-analysis (DMTA) cycle 设计、制造、测试和分析周期

3. 数据, 表征, 基准

Figure 1 AI驱动药物发现中的应用和技术概述
Figure 1 AI驱动药物发现中的应用和技术概述

1. 公开数据资源

Data Resources Description Link
PubChem PubChem是全球最大的化学数据库,收集了750个数据源的化学信息。截至2020年8月,PubChem包含1.11亿个独特的化学结构,来自120万个生物分析实验的2.71亿个活性数据点。 view
ChEMBL 一个包含7700万SMILES字符串的精选数据集(from PubChem)。在ChEMBL22 (version 22)中,有超过160万种不同的化学结构,活性值超过1400万。 view
ZINC 用于目标预测的大规模基准数据集。由UCSF的Irwin和Shoichet实验室开发,包含分子,注释配体和靶标,以及超过1.2亿类药物化合物。 view
PDBbind PDBbind数据库的目的是提供储存在蛋白质数据库(PDB)中所有类型生物分子复合物的实验测量的结合亲和力数据的全面收集。它为这些复合物的能量信息和结构信息之间提供了必要的联系,有助于生物系统分子识别的各种计算和统计研究。 view
BindingDB BindingDB是一个公共的、可在网上访问的测量结合亲和力的数据库,主要关注被认为是药物靶点的蛋白质与小的、类药物分子的相互作用。截至2021年7月4日,BindingDB包含41,300个条目,每个条目都有一个DOI,包含8,547个蛋白质靶标和992,030个小分子的2,295,072个结合数据。 view
DUD 用于基准测试虚拟筛选。 view
DUD-E 增强和重建版本的DUD,DUD-E旨在通过提供具有挑战性的decoys来帮助基准分子对接计划 view
MUV 这些数据集提供了一个用于最大无偏验证 (MUV)的虚拟筛选方法的工具。 view
STITCH 化学-蛋白质相互作用网络 view
GLL&GDD The GDD, a GPCR Decoy Database, with its accompanying GPCR Ligand Library (GLL) have been compiled to help in GPCR docking. view
NRLiSt BDB 一个非商业性的人工管理基准数据库,专门用于核受体(Nuclear Receptor, NR)配体和结构药理学档案。 view
KEGG KEGG是一个从分子水平信息,特别是基因组测序等高通量实验技术生成的大规模分子数据集,了解细胞、生物、生态系统等生物系统的高级功能和用途的数据库 view
DrugBank 一个全面的免费在线数据库,包含有关药物和药物目标的信息。 view
SIDER SIDER包含已上市药品及其记录的药物不良反应的信息。信息是从公共文档和第三方库提取的。现有信息包括副作用频率、药物和副作用分类以及进一步信息的链接,例如药物靶标关系。 view
OFFSIDES 一个关于药物副作用的数据库,但并没有在FDA的官方标签上列出。是唯一全面的双方药效库。超过3300种药物和63000种组合与数百万种潜在的不良反应有关。 view
TWO-SIDES TWO-SIDES数据库是关于药物的多药副作用的资源。该数据库包含59,220对药物和1301个不良事件之间的868,221个显著关联。这些关联仅限于那些不能明确归因于任何一种药物单独的关联(即OFFSIDES所涵盖的关联)。根据比例报告比率(PRR),该数据库还包含另外3782910个显著关联,其中药物对的副作用关联评分高于单独的药物。 view
DILIrank DILIrank DataSet是LTKB基准数据集的更新版本。DILIrank由1,036种FDA批准的药物组成,根据其导致药物诱导的肝损伤 (drug-induced liver injury, DILI)的可能性分为四类。DILI分类衍生自分析FDA批准的药物标签文件中提出的肝毒性描述,并评估文档中的因果关系。具体而言,这一最大的公开注释的DILI数据集包含三组 (Most-,Less- 和 No-DILI),并将药物与肝损伤联系起来的证据,以及一个额外的分组 (Ambiguous-DILI-concern),因果关系未确定。 view

2. 小分子表征

Figure 2 小分子表征插图
Figure 2 小分子表征插图

3. 基准平台

Table 2 基准平台总结

Benchmark platforms Description Link
MoleculeNet MoleculeNet: a benchmark for molecular machine learning (Chem Sci 2018) paper code download
MolMapNet Out-of-the-box deep learning prediction of pharmaceutical properties by broadly learned knowledge-based molecular representations (Nat Mach Intell 2021) paper code
ChemProp Analyzing Learned Molecular Representations for Property Prediction (J Chem Inf Model 2019) paper code website
REINVENT Molecular De Novo design using Recurrent Neural Networks and Reinforcement Learning (J Cheminf 2017) paper code
REINVENT 2.0 an AI Tool for De Novo Drug Design (J Chem Inf Model 2020) paper code
GraphINVENT Graph Networks for Molecular Design (aka: GraphINVENT; ChemRxiv 2020) paper code
GraphINVENT Practical Notes on Building Molecular Graph Generative Models (Applied AI Letters 2020) paper code
Guacamol GuacaMol: Benchmarking Models for de Novo Molecular Design (J Chem Inf Model 2019) paper code
MOSES Molecular Sets (MOSES): A Benchmarking Platform for Molecular Generation Models (Front Pharmacol 2020) paper code

*. 评价指标

Table 3 分子性质预测和分子生成的评价指标

Metric Definition Task Type
Accuracy Correctly predictive rate Classification
Recall True positive rate Classification
Precision Positive predictive value Classification
AUROC Area under receiver-operating curve Classification
AUPRC Area under precision-recall curve Classification
Recall@k Recall among top-k retrieved molecules Retrieval
Precision@k Precision among top-k retrieved molecules Retrieval
AP Average Precision Retrieval
MAE Mean absolute error Regression
RMSE Rooted mean squared error Regression
Validity Fraction of valid molecules Distribution learning
Uniqueness@k Fraction of non-duplicates in k valid molecules Distribution learning
Novelty Fraction of molecules not shown in training set Distribution learning
Diversity Chemical diversity within generated molecules Distribution learning
FCD Fŕechet ChemNet Distance Distribution learning
KL Divergence Kullback-Leibler Divergence Distribution learning
Scaffold Similarity Similarity based on Bemis–Murcko scaffold Goal-directed design
Rediscovery Ability to re-discover target molecule Goal-directed design

4. 模型架构

1. 卷积神经网络 (CNN)

Task: Molecular Property Prediction; Representation*: images
NameDefinitionLink
-Convolutional Embedding of Attributed Molecular Graphs for Physical Property Prediction (J Chem Inf Model 2017)paper
ChemceptionChemception: a deep neural network with minimal chemistry knowledge matches the performance of expert-developed QSAR/QSPR models (aka: Chemception; arXiv 2017)paper
Toxic ColorsToxic Colors: The Use of Deep Learning for Predicting Toxicity of Compounds Merely from Their Graphic Images (aka: Toxic Colors; J Chem Inf Model 2018)paper
KekuleScopeKekuleScope: prediction of cancer cell line sensitivity and compound potency using convolutional neural networks trained on compound images (J Cheminf 2019)paper code
-Learning Drug Functions from Chemical Structures with Convolutional Neural Networks and Random Forests (J Chem Inf Model 2019)paper
DEEPScreenDEEPScreen: high performance drug–target interaction prediction with convolutional neural networks using 2-D structural compound representation (Chem Sci 2020)paper code
Task: Molecular Property Prediction; Representation*: fingerprint
NameDefinitionLink
-Massively Multitask Networks for Drug Discovery (arXiv 2015)paper
-Convolutional Networks on Graphs for Learning Molecular Fingerprints (NeurIPS 2015)paper code
Task: Molecular Structure Extraction and Recognition
NameDefinitionLink
MSE-DUDLMolecular Structure Extraction from Documents Using Deep Learning (J Chem Inf Model 2019)paper
DECIMER-SegmentationDECIMER-Segmentation: Automated extraction of chemical structure depictions from scientific literature (J Cheminf 2021)paper
DECIMERDECIMER: towards deep learning for chemical image recognition (J Cheminf 2020)paper code
DECIMER 1.0DECIMER 1.0: Deep Learning for Chemical Image Recognition using Transformers (chemRxiv 2021)paper

2. 递归神经网络 (RNN)

Task: Molecular Property Prediction; Representation*: SMILES Strings
NameDefinitionLink
SMILES2VecSMILES2Vec: An Interpretable General-Purpose Deep Neural Network for Predicting Chemical Properties (aka: SMILES2Vec; arXiv 2017)paper
-Large-scale comparison of machine learning methods for drug target prediction on ChEMBL (aka:SmilesLSTM; Chem Sci 2018)paper code
Task: Molecule Generation; Representation*: SMILES Strings
NameDefinitionLink
-Molecular de‑novo design through deep reinforcement learning (aka: REINVENT; J Cheminf 2017)paper
-Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks (aka: CharRNN; ACS Cent Sci 2018)paper
-Exploring Deep Recurrent Models with Reinforcement Learning for Molecule Design (ICLR 2018 Workshop)paper
ReLeaSEDeep Reinforcement Learning for de novo Drug Design(aka: ReLeaSE; Sci Adv 2018)paper code
DeepFMPODeep Reinforcement Learning for Multiparameter Optimization in de novo Drug Design (J Chem Inf Model 2019)paper code
Task: Molecule Generation; Representation*: Molecular Graphs
NameDefinitionLink
GraphRNNGraphRNN: Generating Realistic Graphs with Deep Auto-regressive Models (aka: GraphRNN; ICML 2018)paper code
-Learning Deep Generative Models of Graphs (ICML 2018)paper code
MolecularRNNMolecularRNN: Generating realistic molecular graphs with optimized properties (arXiv 2019)paper

3. 图神经网络 (GNN)

Task: Molecular Property Prediction; Representation*: Molecular Graphs
WorkLink
Molecular Graph Convolutions: Moving Beyond Fingerprints (aka: Weave; J Comput Aided Mol Des 2016)paper
Convolutional Embedding of Attributed Molecular Graphs for Physical Property Prediction (J Chem Inf Model 2017)paper
Semi-supervised classification with graph convolutional networks (aka: GraphConv; ICLR 2017)paper code
Neural Message Passing for Quantum Chemistry (aka: MPNN; ICML 2017)paper code
SchNet: A continuous-filter convolutional neural network for modeling quantum interactions (aka: SchNet; NeurIPS 2017)paper code
Low Data Drug Discovery with One-Shot Learning (ACS Cent Sci 2017)paper
Large-scale comparison of machine learning methods for drug target prediction on ChEMBL (aka:SmilesLSTM; Chem Sci 2018)paper code
PotentialNet for Molecular Property Prediction (aka: PotentialNet; ACS Cent Sci 2018)paper
Molecular Property Prediction: A Multilevel Quantum Interactions Modeling Perspective (aka: MGCN; AAAI 2019)paper
Deep Learning-Based Prediction of Drug-Induced Cardiotoxicity (J Chem Inf Model 2019)paper code
DeepChemStable: Chemical Stability Prediction with an Attention-Based Graph Convolution Network (J Chem Inf Model 2019)paper
Analyzing Learned Molecular Representations for Property Prediction (aka: Chemrop, D-MPNN; J Chem Inf Model 2019)paper code
Molecule Property Prediction Based on Spatial Graph Embedding (aka: C-SGEN; J Chem Inf Model 2019)paper code
Pushing the Boundaries of Molecular Representation for Drug Discovery with the Graph Attention Mechanism (aka: Attentive FP; J Med Chem 2019)paper code
Graph convolutional neural networks as” general-purpose” property predictors: the universality and limits of applicability (J Chem Inf Model 2020)paper
N-Gram Graph: Simple Unsupervised Representation for Graphs, with Applications to Molecules (aka: N-Gram Graph; NeurIPS 2019)paper
Building attention and edge message passing neural networks for bioactivity and physical–chemical property prediction (J Cheminf 2020)paper code
A self‑attention based message passing neural network for predicting molecular lipophilicity and aqueous solubility (J Cheminf 2020)paper code
Chemically Interpretable Graph Interaction Network for Prediction of Pharmacokinetic Properties of Drug-Like Molecules (aka: CIGIN; AAAI 2020)paper code
Strategies for Pre-training Graph Neural Networks (ICLR 2020)paper code
Directional Message Passing for Molecular Graphs (aka: DimeNet; ICLR 2020)paper code
Drug–target affinity prediction using graph neural network and contact maps (RSC Advances 2020)paper
ASGN: An Active Semi-supervised Graph Neural Network for Molecular Property Prediction (aka: ASGN; KDD 2020)paper code
Meta-Learning GNN Initializations for Low-Resource Molecular Property Prediction (ICML 2020 Workshop)paper code
Task: Molecule Generation; Representation*: Molecular Graphs
WorkLink
Multi‑objective de novo drug design with conditional graph generative model (J Cheminf 2018)paper code
Graph Convolutional Policy Network for Goal-Directed Molecular Graph Generation (aka: GCPN; NeurIPS 2018)paper code
Optimization of Molecules via Deep Reinforcement Learning (aka: MolDQN; Sci Rep 2019)paper
Improving Molecular Design by Stochastic Iterative Target Augmentation (ICML 2020)paper code
DeepGraphMolGen, a multi‑objective, computational strategy for generating molecules with desirable properties: a graph convolution and reinforcement learning approach (J Cheminf 2020)paper
Reinforced Molecular Optimization with Neighborhood-Controlled Grammars (aka: MNCE-RL; NeurIPS 2020)paper code
Graph Networks for Molecular Design (aka: GraphINVENT; ChemRxiv 2020)paper code
Common GNN Models
WorkLink
Recurrent GNNs Gated graph sequence neural networks (aka: GGNN; ICLR 2016)paper code
Convolutional GNNs (Spectral-based) Convolutional Neural Networks on Graphs with Fast Localized Spectral Filtering (aka: ChebNet; NeurIPS 2016)paper code
Convolutional GNNs (Spectral-based) Semi-supervised classification with graph convolutional networks (aka: GraphConv; ICLR 2017)paper code
Convolutional GNNs (Spatial-based) Neural message passing for quantum chemistry (aka: MPNN; ICML 2017)paper code
Convolutional GNNs (Spatial-based) Inductive Representation Learning on Large Graphs (aka: GraphSAGE; NeurIPS 2017)paper code
Convolutional GNNs (Spatial-based) Graph Attention Networks (aka: GAT; ICLR 2018)paper code
Convolutional GNNs (Spatial-based) How powerful are graph neural networks? (aka: GIN; ICLR 2019)paper code

4. 变分编码器 (VAE)

Task: Molecule Generation; Representation*: SMILES Strings
WorkLink
Automatic chemical design using a data-driven continuous representation of molecules (arXiv 2016; ACS Cent Sci 2018)paper code
Grammar Variational Autoencoder (aka: GrammarVAE; ICML 2017)paper
Application of Generative Autoencoder in De Novo Molecular Design (Mol Inform 2017)paper
Syntax-Directed Variational Autoencoder for Structured Data (aka: SD-VAE; ICLR 2018)paper code
Conditional Molecular Design with Deep Generative Models (aka: Continuous SSVAE; J Chem Inf Model 2018)paper code
Molecular generative model based on conditional variational autoencoder for de novo molecular design (aka: CVAE; J Cheminf 2018)paper code
Constrained Graph Variational Autoencoders for Molecule Design (aka: CGVAE; NeurIPS 2018)paper code
NEVAE: A Deep Generative Model for Molecular Graphs (aka: NeVAE; AAAI 2019)paper code
De Novo Molecular Design by Combining Deep Autoencoder Recurrent Neural Networks with Generative Topographic Mapping (aka: GTMVAE; J Chem Inf Model 2019)paper
Re-balancing Variational Autoencoder Loss for Molecule Sequence Generation (aka: re-balanced VAE; ACM BCB 2020)paper code
CogMol: Target-Specific and Selective Drug Design for COVID-19 Using Deep Generative Models (aka: CogMol; NeurIPS 2020)paper code
VAE变体: AAE
WorkLink
Application of Generative Autoencoder in De Novo Molecular Design (Mol Inform 2017)paper
druGAN: An Advanced Generative Adversarial Autoencoder Model for de Novo Generation of New Molecules with Desired Molecular Properties in Silico (aka: druGAN; Mol Pharm 2017)paper
Entangled Conditional Adversarial Autoencoder for de Novo Drug Discovery (aka: SAAE; Mol Pharm 2018)paper
Task*: Molecule Generation; Representation*: Molecular Graphs
WorkLink
GraphVAE: Towards Generation of Small Graphs Using Variational Autoencoders (aka: GraphVAE; arXiv 2018)paper
Junction Tree Variational Autoencoder for Molecular Graph Generation (aka: JT-VAE; ICML 2018)paper code
Constrained Generation of Semantically Valid Graphs via Regularizing Variational Autoencoders (aka:Regularized VAE; NeurIPS 2018)paper
Molecular Hypergraph Grammar with Its Application to Molecular Optimization (aka: MHG-VAE; ICML 2019)paper code
Efficient learning of non‑autoregressive graph variational autoencoders for molecular graph generation (J Cheminf 2019)paper code
Deep learning enables rapid identification of potent DDR1 kinase inhibitors (aka: GENTRL; Nat Biotechnol 2019)paper code
Scaffold-based molecular design using graph generative model (aka: ScaffoldVAE; arXiv 2019)paper
Learning Multimodal Graph-to-Graph Translation for Molecule Optimization (aka: VJTNN; ICLR 2019)paper code
CORE: Automatic Molecule Optimization Using Copy & Refine Strategy (AAAI 2020)paper code
Hierarchical Generation of Molecular Graphs using Structural Motifs (aka: HierVAE; ICML 2020)paper code
Compressed graph representation for scalable molecular graph generation (J Cheminf 2020)paper code
Task*: Reaction & Retrosynthesis Prediction; Representation*: Molecular Graphs
WorkLink
Generating Molecules via Chemical Reactions (ICLR 2019 Workshop)paper
Barking up the right tree: an approach to search over molecule synthesis DAG (NeurIPS 2020)paper code
paper

5. 对抗生成网络 (GAN)

Task: Molecule Generation; Representation*: SMILES Strings
WorkLink
Objective-Reinforced Generative Adversarial Networks (ORGAN) for Sequence Generation Models (aka: ORGAN; ArXiv 2017)paper code
Optimizing distributions over molecular space. An objective-reinforced generative adversarial network for inverse-design chemistry (aka: ORGANIC; ChemRxiv 2017)paper code
Reinforced Adversarial Neural Computer for de Novo Molecular Design (aka: RANC; J Chem Inf Model 2018)paper
Task*: Molecule Generation; Representation*: Molecular Graphs
WorkLink
MolGAN: An implicit generative model for small molecular graphs (aka: MolGAN; ICML 2018 Workshop)paper code-tensorflow code-pytorch

6. Normalizing Flow Models

Task*: Molecule Generation; Representation*: Molecular Graphs
WorkLink
GraphNVP: An Invertible Flow Model for Generating Molecular Graphs (aka: GraphNVP; arXiv 2019)paper code
Graph Residual Flow for Molecular Graph Generation (aka: GRF; arXiv 2019)paper
GraphAF: a Flow-based Autoregressive Model for Molecular Graph Generation (aka: GraphAF; ICLR 2020)paper code
MoFlow: An Invertible Flow Model for Generating Molecular Graphs (aka: MoFlow; KDD 2020)paper code
GraphDF: A Discrete Flow Model for Molecular Graph Generation (aka: GraphDF; ICML 2021)paper

7. Transformers

Task*: Molecular Property Prediction; Representation*: SMILES Strings
WorkLink
SMILES-BERT: Large Scale Unsupervised Pre-Training for Molecular Property Prediction (aka: SMILES-BERT; ACM BCB 2019)paper
SMILES Transformer: Pre-trained Molecular Fingerprint for Low Data Drug Discovery (aka: SMILES Transformer; arXiv 2019)paper code
ChemBERTa: Large-Scale Self-Supervised Pretraining for Molecular Property Prediction (aka: ChemBERTa; arXiv 2020)paper code
Molecular representation learning with language models and domain-relevant auxiliary tasks (aka: MolBERT; NeurIPS 2020 Workshop)paper code
Task*: Molecular Property Prediction; Representation*: Molecular Graphs
WorkLink
Self-Supervised Graph Transformer on Large-Scale Molecular Data (aka: GROVER; NeurIPS 2020)paper
Task*: Molecule Generation; Representation*: Molecular Graphs
WorkLink
A Model to Search for Synthesizable Molecules (aka: Molecule Chef; NeurIPS 2019)paper code
Transformer neural network for protein-specific de novo drug generation as a machine translation problem (Sci Rep 2021)paper

5. 学习范式

1. 分子性质预测中的自监督学习

Generative Learning
WorkLink
Strategies for Pre-training Graph Neural Networks (ICLR 2020)paper code
Molecular representation learning with language models and domain-relevant auxiliary tasks (aka: MolBERT; NeurIPS 2020 Workshop)paper code
Self-Supervised Graph Transformer on Large-Scale Molecular Data (aka: GROVER; NeurIPS 2020)paper
Contrastive Learning
WorkLink
MolCLR: Molecular Contrastive Learning of Representations via Graph Neural Networks (ArXiv 2021)paper

2. 分子生成中的强化学习

Reinforcement Learning in Molecule Generation
WorkLink
Objective-Reinforced Generative Adversarial Networks (ORGAN) for Sequence Generation Models (aka: ORGAN; ArXiv 2017)paper code
Molecular de‑novo design through deep reinforcement learning (aka: REINVENT; J Cheminf 2017)paper
Optimizing distributions over molecular space. An objective-reinforced generative adversarial network for inverse-design chemistry (aka: ORGANIC; ChemRxiv 2017)paper code
Reinforced Adversarial Neural Computer for de Novo Molecular Design (aka: RANC; J Chem Inf Model 2018)paper
Exploring Deep Recurrent Models with Reinforcement Learning for Molecule Design (ICLR 2018 Workshop)paper
MolGAN: An implicit generative model for small molecular graphs (aka: MolGAN; ICML 2018 Workshop)paper code-tensorflow code-pytorch
Deep Reinforcement Learning for de novo Drug Design(aka: ReLeaSE; Sci Adv 2018)paper code
Graph Convolutional Policy Network for Goal-Directed Molecular Graph Generation (aka: GCPN; NeurIPS 2018)paper code
Deep learning enables rapid identification of potent DDR1 kinase inhibitors (aka: GENTRL; Nat Biotechnol 2019)paper code
Deep Reinforcement Learning for Multiparameter Optimization in de novo Drug Design (aka: DeepFMPO; J Chem Inf Model 2019)paper code
Optimization of Molecules via Deep Reinforcement Learning (aka: MolDQN; Sci Rep 2019)paper
Efficient learning of non‑autoregressive graph variational autoencoders for molecular graph generation (J Cheminf 2019)paper code
GraphAF: a Flow-based Autoregressive Model for Molecular Graph Generation (aka: GraphAF; ICLR 2020)paper code
Reinforcement Learning for Molecular Design Guided by Quantum Mechanics (cka: MolGym; ICML 2020)paper
DeepGraphMolGen, a multi‑objective, computational strategy for generating molecules with desirable properties: a graph convolution and reinforcement learning approach (aka: DeepGraphMolGen; J Cheminf 2020)paper
Reinforced Molecular Optimization with Neighborhood-Controlled Grammars (aka: MNCE-RL; NeurIPS 2020)paper code
Common RL Algorithms
WorkLink
Value-based Playing Atari with Deep Reinforcement Learning (aka: DQN; NeurIPS Workshop 2013)paper
Value-based Human-level control through deep reinforcement learning (aka: DQN; Nature 2015)paper
Value-based Deep Reinforcement Learning with Double Q-learning (aka: Double Q-learning; AAAI 2016)paper
Value-based Prioritized Experience Replay (aka: DQN with Experience Replay; ICLR 2016)paper
Value-based Dueling Network Architectures for Deep Reinforcement Learning (aka: Dueling Network; ICML 2016)paper
Policy-gradient Simple statistical gradient-following algorithms for connectionist reinforcement learning (aka: REINFORCE; Mach Learn 1992)paper
Policy-gradient Policy Gradient Methods for Reinforcement Learning with Function Approximation (aka: Random Policy Gradient; NeurIPS 1999)paper
Policy-gradient Deterministic Policy Gradient Algorithms (aka: DPG; ICML 2014)paper
Policy-gradient Trust Region Policy Optimization (aka: TRPO; ICML 2015)paper
Policy-gradient Proximal Policy Optimization Algorithms (aka: PPO; arXiv 2017 2015)paper
Hybrid Continuous control with deep reinforcement learning (aka: DDPG; ICLR 2016)paper
Hybrid Asynchronous Methods for Deep Reinforcement Learning (aka: A3C; ICML 2016)paper
Pareto Optimality (帕累托最优)
WorkLink
De Novo Drug Design of Targeted Chemical Libraries Based on Artificial Intelligence and Pair-Based Multiobjective Optimization (J Chem Inf Model 2020)paper code
Multiobjective de novo drug design with recurrent neural networks and nondominated sorting (J Cheminf 2020)paper
DrugEx v2: De Novo Design of Drug Molecule by Pareto-based Multi-Objective Reinforcement Learning in Polypharmacology (ChemRxiv)paper
paper
Reaction & Retrosynthesis Optimization
WorkLink
Optimizing chemical reactions with deep reinforcement learning (ACS Cent Sci 2017)paper

3. Other

Metric Learning
WorkLink
Machine-guided representation for accurate graph-based molecular machine learning (Phys Chem Chem Phys 2020)paper
Embedding of Molecular Structure Using Molecular Hypergraph Variational Autoencoder with Metric Learning (Mol Inform 2020)paper
Few-Shot Learning
WorkLink
Low Data Drug Discovery with One-Shot Learning (ACS Cent Sci 2017)paper
paper
Meta Learning
WorkLink
Meta-Learning GNN Initializations for Low-Resource Molecular Property Prediction (ICML 2020 Workshop)paper code
Active Learning
WorkLink
ASGN: An Active Semi-supervised Graph Neural Network for Molecular Property Prediction (aka: ASGN; KDD 2020)paper code
Evidential Deep Learning for Guided Molecular Property Prediction and Discovery (NeurIPS 2020 Workshop)paper

6. 解决现有挑战

1. Model Interpretation

Work Link
Drug Discovery Maps, a Machine Learning Model That Visualizes and Predicts Kinome−Inhibitor Interaction Landscapes (J Chem Inf Model 2018) paper
Using attribution to decode binding mechanism in neural network models for chemistry (PNAS 2019) paper
Interpretation of QSAR Models by Coloring Atoms According to Changes in Predicted Activity: How Robust Is It? (J Chem Inf Model 2019) paper
Building of Robust and Interpretable QSAR Classification Models by Means of the Rivality Index (J Chem Inf Model 2019) paper

2. Dataset Concerns

Work Link
In Need of Bias Control: Evaluating Chemical Data for Machine Learning in Structure-Based Virtual Screening (J Chem Inf Model 2019) paper
Deep Learning-Based Imbalanced Data Classification for Drug Discovery (J Chem Inf Model 2020) paper code

3. Uncertainty Estimation

Work Link
General Approach to Estimate Error Bars for Quantitative Structure−Activity Relationship Predictions of Molecular Activity (J Chem Inf Model 2018) paper
Assessment and Reproducibility of Quantitative Structure−Activity Relationship Models by the Nonexpert (J Chem Inf Model 2018) paper
Deep Confidence: A Computationally Efficient Framework for Calculating Reliable Prediction Errors for Deep Neural Networks (J Chem Inf Model 2018) paper
Reliable Prediction Errors for Deep Neural Networks Using Test-Time Dropout (J Chem Inf Model 2019) paper
Minimal-uncertainty prediction of general drug-likeness based on Bayesian neural networks (Nat Mach Intell 2020) paper
Assigning Confidence to Molecular Property Prediction (arXiv 2021) paper
Gi and Pal Scores: Deep Neural Network Generalization Statistics (ICLR 2021 Workshop) paper

4. Representation Capacity

Work Link
Ligand-Based Virtual Screening Using Graph Edit Distance as Molecular Similarity Measure (J Chem Inf Model 2019) paper
Optimal Transport Graph Neural Networks (arXiv 2020) paper

5. Out-of-Distribution Generalization

Work Link
Dissecting Machine-Learning Prediction of Molecular Activity: Is an Applicability Domain Needed for Quantitative Structure−Activity Relationship Models Based on Deep Neural Networks? (J Chem Inf Model 2018) paper
Most Ligand-Based Classification Benchmarks Reward Memorization Rather than Generalization (J Chem Inf Model 2018) paper
Molecular Similarity-Based Domain Applicability Metric Efficiently Identifies Out-of-Domain Compounds (J Chem Inf Model 2018) paper

7. 参考文献

  1. Deng, Jianyuan & Yang, Zhibo & Ojima, Iwao & Samaras, Dimitris & Wang, Fusheng. (2021). Artificial Intelligence in Drug Discovery: Applications and Techniques.

评论