Artificial Intelligence in Drug Discovery: Applications and Techniques

Motivation

• Developing a new drug costs 2.6 billion US dollars and can take more than 10 years, while fewer than 10% of drugs entering Phase I reach the market.Source: Introduction · p. 2

AI applications

• AI applications in drug discovery include virtual screening, de novo drug design, retrosynthesis, reaction prediction, and de novo protein design, grouped into predictive and generative tasks.Source: Introduction · p. 2

Scope

• The survey focuses on AI-driven discovery of small-molecule drugs and excludes biologics such as antibodies and vaccines.Source: Introduction · p. 2

Survey coverage

• It surveys drug-discovery applications, prior perspectives and reviews, data resources, molecular representations, and benchmark platforms for property prediction and molecule generation.Source: Introduction · p. 2

Techniques and outlook

• The survey explains relevant model architectures and learning paradigms, then discusses existing challenges and future directions.Source: Introduction · p. 2

Resources

• A GitHub repository collects the surveyed papers and applicable codes to provide a comprehensive overview and learning resource for interdisciplinary researchers.Source: Introduction · p. 2

Drug Discovery Overview

• Drug discovery progresses from hypothesizing therapeutic target effects through target validation, hit and lead discovery, lead optimization, preclinical studies, and clinical trials.Successful candidates can ultimately be launched as medical products.Source: Drug Discovery Overview · p. 3
• Virtual screening computationally searches chemical libraries to identify active molecules using knowledge about the target or known active ligands, increasing the odds of finding active molecules.The chemical space is estimated at around 10^30 − 10^60 molecules.Source: Drug Discovery Overview · pp. 3, 5
• Drug candidates require more than sufficient activity: they should exhibit high selectivity and satisfy a combination of physicochemical and pharmacokinetic criteria, because binding promiscuity can cause unexpected side effects.Source: Drug Discovery Overview · p. 4

Summary of Existing Reviews

• Existing reviews are grouped into general drug discovery, drug discovery in the AI era, and rethinking AI-driven drug discovery, with activity cliffs highlighting a major predictive challenge.Activity cliffs occur when structurally similar molecules have significantly different activities.Source: Summary of Existing Reviews · p. 5, General Drug Discovery Reviews · p. 6

Drug Discovery in the AI Era

• In the AI era, deep neural networks outperformed the standard random-forest model in a Kaggle competition; DDR1 inhibitors were discovered in 21 days, and halicin was identified in 2020.AI can be applied across drug discovery, from target identification and validation to drug response determination.Source: Drug Discovery in the AI Era · p. 6
• The survey focuses on lead identification through molecular property prediction and molecule generation, including realistic generation under chemical constraints and goal-directed generation with desired properties.Property prediction supports applications such as drug-target interaction, toxicity, and drug-induced liver injury prediction.Source: Drug Discovery in the AI Era · pp. 6–7

Rethinking AI-driven Drug Discovery

• AI methods cannot compensate for unsuitable representations, irrelevant data, or poorly conceived questions, making holistic evaluation necessary to avoid hype and unrealistic expectations. [8,9,31,52–55]The survey emphasizes considering hypotheses, data, representations, models, and learning paradigms together.Source: Rethinking AI-driven Drug Discovery · p. 7

Structure of the Survey

• The survey covers drug-discovery applications, data resources, molecular representations, model architectures, and learning paradigms to connect general drug discovery with AI techniques.Its applications include virtual screening, quantitative structure-activity relationships, and drug design, reduced to molecular property prediction and molecule generation.Source: Structure of the Survey · p. 7, Figure 1 · p. 8
• Because AI-technique literature is fragmented and methods are often chosen because they were not previously investigated, the survey compares architectures and learning paradigms chronologically to expose strengths and weaknesses. [51]The survey also highlights existing challenges and future directions.Source: Structure of the Survey · p. 7
• The survey identifies low-data molecular property prediction and goal-directed molecule generation as challenges addressed by paradigms including self-supervised and reinforcement learning.It also surveys few-shot, metric, meta, and active learning.Source: Figure 1 · p. 8

Data, Representation and Benchmark Platforms

• The section covers public data resources, machine-readable small-molecule representations, and benchmark platforms for molecular property prediction and molecule generation.Source: Data, Representation and Benchmark… · p. 8

Public Data Resources

• Public databases provide molecular structures, molecular properties, and target information, while PubChem is non-curated and its bioactivity datasets can be highly imbalanced.Source: Public Data Resources · pp. 8–9

Small Molecule Representations

• Molecular descriptors include 0D properties such as molecular weight and atom number, 2D representations of atom connectivity, and 3D descriptors; these fixed descriptors are not learnable.Source: Small Molecule Representations · p. 10

Molecular Graphs

• Molecular graphs represent atoms as nodes and bonds as edges, preserving interpretable chemical substructures but requiring substantial disk storage and computational memory.Graph representations use node and edge features together with adjacency information to encode molecular connectivity.Source: Molecular Graphs · pp. 11–12

SMILES Strings

• SMILES strings are computationally cheaper than graphs and can be canonicalized for unique molecular encoding, but their linear syntax can lose connectivity information and generate invalid molecules.SMILES are converted into one-hot vectors before machine-learning models, while ring syntax and atom valency create validity challenges during generation.Source: SMILES Strings · pp. 12–13

Other Representations

• Beyond graphs and SMILES, molecules can also be represented using sophisticated three-dimensional atomic coordinates.Source: Other Representations · p. 13
• 3D representations incorporate bond lengths, angles, torsions, coordinates, and conformation-related properties, while molecular structure images provide another modality.3D representations have been used for both prediction and molecule generation.Source: Other Representations · p. 13

Benchmark Platforms

• MoleculeNet provides property-prediction datasets spanning quantum mechanics, physical chemistry, biophysics, and physiology, with single-task and multitask settings.Its datasets include QM7–QM9, ESOL, FreeSolv, Lipophilicity, PCBA, MUV, HIV, PDDBind, BACE, BBBP, Tox21, ToxCast, SIDER, and ClinTox.Source: Benchmark Platforms · p. 13
• Because datasets can be highly imbalanced, AUPRC is preferred over AUROC when positive rates are very low, such as below 2%.Recommended dataset splits vary by task: scaffold splitting suits single-target BACE, whereas time splitting better reflects the longitudinal PDBind collection.Source: Benchmark Platforms · p. 14
• MoleculeNet lacks explicit training, validation, and test folds; the ChemBench package from MolMapNet was released to improve reproducibility, while MolMapNet adds pharmacokinetics-related datasets.Source: Benchmark Platforms · p. 14
• Generation benchmarks include REINVENT, GraphINVENT, GuacaMol, and MOSES, evaluating tasks such as distribution learning, goal-directed design, validity, diversity, novelty, FCD, and KL divergence.Goal-directed evaluation scores molecules against predefined structural, physicochemical, similarity, or dissimilarity criteria, including rediscovery and similarity.Source: Benchmark Platforms · p. 14

Model Architectures

• Deep neural networks have outperformed traditional models across various drug-discovery tasks, including classification approaches such as SVM, KNN, RF, NB, and LR.Traditional models support predictions of drug likeness, physicochemical properties, pharmacokinetic parameters, and pharmacodynamic properties.Source: Model Architectures · p. 16

Convolutional Neural Networks

• CNNs process images or fingerprints through convolution and pooling layers, sharing filter parameters to reduce memory consumption and increase computation speed.In drug discovery, CNNs support bioactivity profiling, molecular property prediction, and image-to-SMILES translation; KekuleScope achieved performance comparable to RF and DNNs on ECFPs, while compound-image predictions outperformed transcriptomic-data predictions for MeSH therapeutic-use classes.Source: Convolutional Neural Networks · pp. 16–17

Recurrent Neural Networks

• RNNs process sequential SMILES inputs for property prediction and autoregressive molecule generation, but ordinary RNNs often produce chemically invalid SMILES because they cannot capture syntax-dependent counting patterns.LSTM and GRU add memory modules, while Stack-RNN and bidirectional LSTM address validity; GraphRNN generates molecular graphs autoregressively, whereas GCPN formulates graph generation as an MDP using only the intermediate state.Source: Recurrent Neural Networks · pp. 18–19

Graph Neural Networks

• GNNs naturally model molecular graphs for node-, edge-, and graph-level tasks, using neighborhood aggregation, pooling, and readout through recurrent or convolutional architectures.ConvGNNs generalize convolution from grids to graphs, while MPNNs use message passing and readout; early Weave graph convolutions did not outperform fingerprint-based models.Source: Graph Neural Networks · p. 20
• GNNs propagate neighbor features into atom or edge messages, aggregate them into graph representations, and support property prediction through downstream fully connected layers.Directed MPNN (D-MPNN) uses bond-associated messages to prevent repeated node updates, while RDKit global features can be concatenated with learned features; attention variants capture adjacent-atom interactions.Source: Graph Neural Networks · pp. 20–21
• GNN-based molecule generation models graphs through state-dependent actions, outperform SMILES-based generation especially in validity, and support scaffold-focused and goal-directed design.MolMP models graph growth as a Markov decision process with append, connect, and terminate actions; reinforcement learning is often integrated for desired-property optimization, as in GCPN, MolDQN, DeepGraphMolGen, and MNCE-RL.Source: Graph Neural Networks · p. 22

Variational Autoencoders

• VAEs encode molecules into organized, low-dimensional continuous latent spaces regularized by KL divergence, enabling sampling and property-directed generation.For SMILES, property regression can organize the latent space toward desired values, but syntax-related sparsity creates invalid-molecule “dead areas,” motivating GrammarVAE and syntax-directed VAE.Source: Variational Autoencoders · pp. 22–23
• Graph VAEs represent molecules through probabilistic graphs, junction trees, or hypergraphs, enabling valid, diverse, constraint-aware generation and scaffold retention.JT-VAE searches latent junction-tree substructures and guarantees validity of generated molecules; related models add graph matching, reinforcement learning, and auxiliary property prediction.Source: Variational Autoencoders · p. 24
• Graph VAE advances address inaccurate rare-substructure prediction and scalability by combining scaffolding-tree generation with adversarial training, compressed representations, and hierarchical motifs.CORE combines scaffolding tree generation and adversarial training; compressed graphs reduce computational complexity while maintaining validity and diversity, and HierVAE uses multi-resolution atom-to-motif encoding with coarse-to-fine decoding.Source: Variational Autoencoders · pp. 24–25
• HierVAE can generate polymers, extending variational autoencoder applications beyond molecular generation.Source: Variational Autoencoders · p. 25

Generative Adversarial Networks

• Objective-reinforced GANs generated molecules exhibiting drug-like structures and improved evaluation metrics; ORGANIC generated molecules with biased distributions, while RANC outperformed ORGANIC.Source: Generative Adversarial Networks · pp. 25–26

Normalizing Flow Models

• Flow models learn invertible mappings with exact and tractable likelihood estimation, enabling efficient one-shot inference and 100% reconstruction of training data; molecular flow models generate valid molecules with minimal duplicates.GRF has comparable performance with GraphNVP while using much less parameters; GraphAF generates molecules with 100% validity and significantly faster training.Source: Normalizing Flow Models · pp. 27–28

Transformers

• Transformers replace recurrence with self-attention and positional embeddings, handling long sequences and enabling self-supervised molecular pretraining.SMILES-BERT 215 uses transformer encoder layers and is pretrained on large-scale SMILES corpora through a SMILES recovery task for molecular property prediction.Source: Transformers · pp. 29–30
• MolBERT uses BERT-based molecular representation learning with masked language modeling and other tasks, improving molecular property prediction over fixed fingerprints. [95]ChemBERTa and the SMILES Transformer are additional sequence-based approaches for molecular representation learning. [59, 216]Source: Transformers · p. 30
• Transformer encoders can be applied to molecular graphs because they can be viewed as a variant of graph attention networks. [150]GROVER exemplifies this graph-based use through a message passing transformer. [63]Source: Transformers · p. 30
• GROVER pretrains graph representations on 10 million unlabeled molecules using contextual property and graph-level motif prediction, achieving state-of-the-art performance on 11 benchmark datasets. [63]Its message passing transformer framework is illustrated in Fig 9B.Source: Transformers · p. 30
• Transformers can generate molecules by mapping products to reactants, as in MoleculeChef, or by translating target-protein amino acid sequences into ligand SMILES. [217, 218]The protein-specific formulation uses the amino acid sequence as input and ligand SMILES as output.Source: Transformers · p. 30

Learning Paradigms

• Limited, sparse, biased, and noisy labeled datasets define the low-data drug discovery problem, while latent-space mappings can restrict molecular novelty and diversity.These challenges motivate learning paradigms for molecular property prediction and molecule generation.Source: Learning Paradigms · pp. 30–31

Self-Supervised Learning

• Self-supervised learning addresses limited labels by recovering data through generative or contrastive objectives, with MolCLR achieving state-of-the-art molecular property-prediction performance.MolCLR uses atom masking, bond deletion, and subgraph removal to construct molecular graph augmentations for contrastive learning.Source: Self-Supervised Learning · pp. 31–32
• Self-supervised pretraining can avoid negative transfer from supervised pretraining, while meta-learning initialization outperforms multi-task baselines on 16 of 20 in-distribution tasks and all out-of-distribution tasks.The section also describes few-shot and metric-learning approaches for prediction with few examples and mixed distributions.Source: Self-Supervised Learning · pp. 32–33
• Metric learning alleviates structure-property mismatch by better separating molecules in latent space and can support efficient goal-directed molecule generation.A VAE-based molecular embedding framework aligns latent representations with molecular properties to enable efficient search.Source: Self-Supervised Learning · p. 33

Reinforcement Learning

• Reinforcement learning navigates high-dimensional, non-convex molecular design spaces by connecting generative agents with predictive models that assign rewards from predicted properties.The survey covers value-based, policy-gradient, and actor-critic methods, while noting policy-gradient variance and TRPO’s second-order computational cost.Source: Reinforcement Learning · pp. 33–34
• Reinforcement learning is expected to generate chemical libraries shifted toward desired properties.Source: Reinforcement Learning · p. 34
• Drug design is a multi-objective optimization problem, so non-dominated sorting or pair-based comparisons are used to identify Pareto-optimal solutions.Source: Reinforcement Learning · pp. 34–35
• Reinforcement learning faces an exploration–exploitation dilemma arising from uncertainty and incomplete knowledge of rewards across states.Source: Reinforcement Learning · p. 35
• Active learning may help resolve the exploration–exploitation trade-off by querying an expert or other information sources during learning.Source: Reinforcement Learning · p. 35

Data and evaluation

• Although chemical libraries contain many molecules, data for specific assays can be very scarce, and benchmark datasets may have questionable representativeness for real-world drug discovery across vast chemical space.Source: Discussions · p. 35

Hypothesis-driven design

• Drug design should be driven by a need or certain hypotheses, while real-world data, including electronic health records and marketed-drug databases, can generate insights into drug design [241].Source: Discussions · pp. 35–36

Explainability

• Deep learning models remain difficult to interpret, motivating explainable systems that provide transparency, justification, informativeness, and uncertainty estimation [31].The desired outcome is for AI to help scientists refine their knowledge and beliefs about the investigated process.Source: Discussions · p. 36

Technical concerns

• Fingerprints can still outperform GNN-derived representations for molecular property prediction [244], while AI-driven drug-discovery studies lack unified protocols, using different benchmark datasets, split folds, and evaluation metrics.ECFPs are a component of some GNN models [63,71].Source: Discussions · p. 36

Overall perspective

• Successful AI applications require understanding the task, data, molecule representation, model architecture, and learning paradigm as an integrated whole.The survey covers multiple aspects centered on AI-driven drug discovery and anticipates more meaningful contributions as these aspects become better understood.Source: Discussions · p. 37

• J.D. and Z.Y. conceived the manuscript and GitHub repository, while J.D. drafted the manuscript and built the repository.Source: Author contributions statement · p. 37
• All authors made critical revisions and reviewed the manuscript.Source: Author contributions statement · p. 37

Graphical TOC Entry

• The entry uses Helvetica 8 pt within a 9 cm by 3.5 cm frame, which is the maximum permitted size for Journal of the American Chemical Society graphical TOC entries.Content that exceeds the frame overflows rather than resizing, and the box with its title prints on a separate final page.Source: Graphical TOC Entry · p. 65