Source-linked AI summary
Artificial Intelligence in Drug Discovery: Applications and Techniques
Jianyuan Deng, Zhibo Yang, Iwao Ojima, Dimitris Samaras, Fusheng Wang
TL;DR
Drug discovery is expensive, time-consuming, and marked by low success, while AI has expanded its use across drug-discovery applications. This survey covers AI-driven drug discovery across model architectures, learning paradigms, data and evaluation considerations, and related applications, while providing a paper-and-code repository as a learning resource.
Problem
Drug discovery is expensive and time-consuming, with clinical-trial success below 10%.
Method
The survey covers AI-driven drug discovery across model architectures, learning paradigms, data resources, evaluation considerations, and multiple related aspects.
Results
The survey synthesizes multiple aspects of AI-driven drug discovery and discusses patterns from real-world data that allow hypothesis generation and calibrate drug-design insights.
Takeaways & Limitations
The survey is intended as a guide for researchers and provides a regularly updated repository of surveyed papers and applicable codes as a learning resource.
Takeaways & Limitations
Useful AI applications require sufficient amounts of high-quality data, while incomplete data can create uncertainty and trade-offs.
Abstract
from arXiv · showhide
Artificial intelligence (AI) has been transforming the practice of drug discovery in the past decade. Various AI techniques have been used in a wide range of applications, such as virtual screening and drug design. In this survey, we first give an overview on drug discovery and discuss related applications, which can be reduced to two major tasks, i.e., molecular property prediction and molecule generation. We then discuss common data resources, molecule representations and benchmark platforms. Furthermore, to summarize the progress of AI in drug discovery, we present the relevant AI techniques including model architectures and learning paradigms in the papers surveyed. We expect that this survey will serve as a guide for researchers who are interested in working at the interface of artificial intelligence and drug discovery. We also provide a GitHub repository (https://github.com/dengjianyuan/Survey_AI_Drug_Discovery) with the collection of papers and codes, if applicable, as a learning resource, which is regularly updated.
Introduction
Drug discovery is costly, slow, and has a low clinical success rate, motivating AI-driven approaches. This survey organizes small-molecule applications around predictive and generative tasks while reviewing their data, representations, architectures, paradigms, challenges, and resources.
- Developing a new drug costs 2.6 billion US dollars on average and can take more than 10 years.
- Less than 10% of drugs entering Phase I clinical trials successfully reach the market.
- AI applications in small-molecule discovery center on predictive and generative tasks, including virtual screening, drug design, and molecule generation.
- The survey covers small-molecule AI applications, data resources, molecular representations, benchmark platforms, model architectures, and learning paradigms.
- It highlights low-data molecular property prediction and goal-directed molecule generation as continuing challenges, alongside future directions.
- A GitHub repository collects surveyed papers and applicable codes as a learning resource for researchers entering the interdisciplinary field.
Drug Discovery Overview
Drug discovery combines target and candidate development with multi-objective molecular optimization. The survey frames AI-driven discovery around molecular property prediction and molecule generation, while reviewing established applications, representations, models, and challenges.
- Drug discovery process: Drug discovery progresses from target identification and validation through hit discovery, lead optimization, preclinical studies, and clinical trials.
- Core applications: Virtual screening uses computational approaches based on target structures or known active ligands to identify molecules for subsequent assays.
- Candidate objectives: An ideal drug candidate must balance activity, selectivity, physicochemical, pharmacokinetic, pharmacodynamic, and synthesis-related properties.
- Core applications: Drug discovery is a multi-objective optimization problem in which QSAR models map molecular structures to property values through classification or regression.
- Core applications: Molecule generation complements property prediction by exploring chemical space through de novo design and iterative design-make-test-analysis cycles.
- Challenges: Activity cliffs make prediction difficult because structurally similar molecules can exhibit significantly different activities after minor structural changes.
Data, Representation and Benchmark Platforms
The survey reviews public chemical data resources, machine-readable molecular representations, and benchmark platforms for molecular property prediction and molecule generation. It contrasts fixed descriptors with learnable graph and SMILES representations, while noting practical storage, computation, and evaluation challenges.
- The survey organizes this section around public data resources, molecule representations, and benchmark platforms for molecular property prediction and molecule generation.
- Molecular Representations: Deep learning enables end-to-end molecular embeddings without hand-crafted rules, with molecular graphs and SMILES strings as two major representation formats.Graphs map atoms to nodes and bonds to edges, whereas SMILES strings are commonly converted into one-hot vectors before entering machine-learning models.
- Molecular Representations: Graph representations preserve structural information and interpretability but require substantial storage and computation, whereas SMILES strings are less computationally expensive.Canonicalization assigns one unique SMILES string to the same molecule, although graphs directly map to chemical substructures and their matrices can slow computation.
Model Architectures
The survey describes a progression from traditional machine-learning models to deep architectures for molecular property prediction and molecule generation. CNNs process molecular images or fingerprints, RNNs process SMILES sequences, GNNs model molecular graphs, and generative architectures produce candidate molecules.
- Traditional models include SVM, KNN, RF, naive Bayes, and logistic regression, while DNNs have outperformed them across varied tasks.
- Convolutional Neural Networks: CNNs extract molecular features from images or fingerprints for property prediction, with parameter sharing reducing memory consumption and increasing computation speed.CNN applications include solvation, toxicity, therapeutic-use, and other molecular-property predictions.
- Recurrent Neural Networks: RNNs sequentially process one-hot encoded SMILES, supporting property prediction and autoregressive molecule generation.LSTM and GRU variants add memory modules to address long-term dependency and gradient problems.
- Recurrent Neural Networks: RNN-based generation can produce chemically invalid SMILES because long sequences are difficult to model reliably.
- Graph Neural Networks: GNNs aggregate molecular graph features for property prediction and molecule generation, while graph-based methods target validity, diversity, and optimization.Examples include JT-VAE, CORE, compressed graph representations, and HierVAE, which can generate polymers.
- Generative Adversarial Networks: GAN-based models combine generators and discriminators to generate molecules while optimizing drug-related attributes and evaluation metrics.ORGAN improves domain-specific metrics, ORGANIC biases generation toward attributes, and RANC improves several structural and medicinal-chemistry measures.
- Generative Adversarial Networks: MolGAN generates nearly 100% valid molecules on QM9 but is susceptible to mode collapse, whereas other graph generators report strong validity and optimization performance.
Learning Paradigms
The survey presents learning paradigms for molecular property prediction and molecule generation under limited, sparse, biased, and noisy data. It emphasizes self-supervised learning and reinforcement learning, while also discussing meta-learning, few-shot learning, and metric learning.
- Challenges: Molecular property prediction is constrained by expensive labels and datasets that are limited, sparse, biased, and noisy.This is described as the low-data drug discovery problem.
- Challenges: Existing generative models can target desired properties, but latent-space mapping may restrict chemical-space exploration and reduce novelty and diversity.The survey identifies this as a challenge for molecule generation.
- Self-supervised learning: Self-supervised learning is surveyed mainly for property prediction, using generative and contrastive paradigms to reduce reliance on labeled data.Masked language modeling predicts masked tokens, while contrastive learning learns representations from augmented positive and negative pairs.
- Self-supervised learning: MolCLR uses atom masking, bond deletion, and subgraph removal on molecular graphs, then transfers learned representations to downstream tasks with state-of-the-art performance.Positive pairs come from augmentations of the same molecule, whereas other molecules form negative pairs.
- Reinforcement learning: Reinforcement learning navigates chemical space for goal-directed generation by connecting a generative agent with a predictive reward model.The survey discusses value-based, policy-gradient, and hybrid RL algorithms, including DQN, REINFORCE, and PPO.
Discussions
The discussion identifies data, interpretability, representation, and evaluation challenges that constrain AI-driven drug discovery, while highlighting real-world data as a source of design hypotheses. It concludes that coordinated attention to tasks, data, representations, architectures, and learning paradigms is needed.
- Data challenges: Existing assay data can be scarce, imbalanced, and of questionable representativeness for real-world drug discovery.The discussion notes that large chemical libraries do not guarantee sufficient data for specific assays, while benchmark quality may not reflect the vast chemical space.
- Evaluation challenges: Evaluation requires suitable datasets, balancing methods, and metrics such as AUPRC rather than relying automatically on AUROC.The authors specifically connect high class imbalance with metric selection and evaluation design.
- Hypothesis generation: Real-world data from EHRs and drug databases can reveal therapeutic properties and pharmacological components that generate and calibrate drug-design hypotheses.Examples include opioid analgesic properties and drug-drug interaction components involving carriers, transporters, enzymes, and targets.
- Interpretability: Deep learning remains difficult to interpret, motivating models that provide transparency, justification, informativeness, and uncertainty estimation.These four aspects are presented as requirements for explainable AI in drug discovery.
- Technical and benchmarking challenges: Fixed fingerprints can outperform GNN-derived molecular representations, while inconsistent datasets, splits, metrics, and procedures hinder unified evaluation.The discussion also notes that ECFPs are components of some GNN models and calls for protocols covering both property prediction and molecule generation.
- Conclusion: The survey organizes AI-driven drug discovery around tasks, data, molecular representations, model architectures, and learning paradigms.The authors present this integrated view as a basis for more meaningful contributions.
Author contributions statement
The manuscript and its accompanying GitHub repository were conceived by J.D. and Z.Y.; J.D. drafted the manuscript and built the repository, while all authors revised it.
- J.D. and Z.Y. conceived the manuscript and the GitHub repository.
- J.D. drafted the manuscript and built the GitHub repository.
- All authors made critical revisions and reviewed the manuscript.
Graphical TOC Entry
The graphical Table of Contents entry must be print-ready within the Journal of the American Chemical Society’s specified Helvetica, frame, and sizing constraints.
- The graphical entry uses Helvetica at 8 pt inside a 9 cm by 3.5 cm frame.
- The frame is the maximum permitted size for Journal of the American Chemical Society graphical Table of Contents entries.
- Content that exceeds the available space overflows rather than resizing the surrounding box.