Source-linked AI summary
Mapping the Landscape of Artificial Intelligence Applications against COVID-19
Joseph Bullock, Alexandra Luccioni, Katherine Hoffmann Pham, Cynthia Sin Nga Lam, Miguel Luengo-Oroz
TL;DR
The review examines how AI can address COVID-19 challenges across molecular, clinical, and societal domains. It maps emerging applications and resources while emphasizing operational maturity, critical evaluation, and coordinated research action.
Problem
Researchers need to understand and suppress COVID-19 by improving transmission analysis, detection, vaccines and treatments, and understanding socio-economic impacts.
Method
The review maps rapidly emerging AI literature across molecular, clinical, and societal scales, covering applications, datasets, tools, resources, and implementation considerations.
Results
The review identifies AI applications in drug discovery, diagnosis, clinical outcome prediction, epidemiology, and infodemiology, while finding that few systems have operational maturity.
Takeaways & Limitations
The review calls for research road maps, application funnels, multidisciplinary collaboration, open science, and international cooperation to direct AI toward current and future pandemics.
Takeaways & Limitations
Most reviewed medical-imaging studies lacked clinical evaluation provisions, diverse datasets, robust evaluation procedures, and plans for integration into clinical workflows.
Abstract
from arXiv · showhide
COVID-19, the disease caused by the SARS-CoV-2 virus, has been declared a pandemic by the World Health Organization, which has reported over 18 million confirmed cases as of August 5, 2020. In this review, we present an overview of recent studies using Machine Learning and, more broadly, Artificial Intelligence, to tackle many aspects of the COVID-19 crisis. We have identified applications that address challenges posed by COVID-19 at different scales, including: molecular, by identifying new or existing drugs for treatment; clinical, by supporting diagnosis and evaluating prognosis based on medical imaging and non-invasive measures; and societal, by tracking both the epidemic and the accompanying infodemic using multiple data sources. We also review datasets, tools, and resources needed to facilitate Artificial Intelligence research, and discuss strategic considerations related to the operational implementation of multidisciplinary partnerships and open science. We highlight the need for international cooperation to maximize the potential of AI in this and future pandemics.
1. Introduction
This review maps AI applications addressing COVID-19 across molecular, clinical, and societal scales, while emphasizing that few applications are mature enough for operational impact. It presents a roadmap involving regulatory safeguards, international cooperation, multidisciplinary research, and open science.
- Scope and organization: AI applications are organized across molecular, clinical, and societal scales to address COVID-19 challenges.The molecular scale includes proteins, drugs, compounds, and vaccines; the clinical scale includes diagnosis and prognosis; the societal scale includes epidemiology and infodemiology.
- Molecular applications: At the molecular scale, AI supports protein-structure estimation, drug repurposing, compound discovery, vaccine-target identification, and analysis of infectivity and severity.
- Clinical applications: At the clinical scale, AI supports imaging-based diagnosis, non-invasive disease tracking, and outcome prediction from multiple data inputs.
- Societal applications: At the societal scale, AI is used for epidemiological modeling, regional comparisons, and investigation of the accompanying infodemic.
- Maturity and implementation: Few reviewed applications are mature enough to demonstrate operational impact, despite the broad range of potential AI uses.
- Strategic considerations: The review recommends regulatory and quality-assurance frameworks for critical settings and international cooperation grounded in multidisciplinary research and open science.
2. Article Selection
The review selected early COVID-19 literature through manual searches supplemented by automated CORD-19 retrieval, screening manuscripts for quality, originality, and clarity. Its scope was shaped by rapid publication, ML/AI inclusion judgments, and the prevalence of preprints requiring further evaluation.
- Search context: Over 30,000 coronavirus-related papers appeared in CORD-19 between January 1 and August 1, 2020, including over 1,000 with AI- or ML-related terms in titles or abstracts.
- Search strategy: The review combined manual searches of preprint servers and Google Scholar with an automated search of the CORD-19 dataset.
- Screening: Articles released between January 1 and April 10, 2020 were screened using quality, originality, and clarity criteria.
- Scope criteria: The review included explicitly described neural-network and decision-tree applications while excluding simple linear-regression applications.
- Scope boundary: Because publication progressed rapidly, the survey was not comprehensive at publication and concentrated on the initial pandemic response.
- Evidence quality: Many cited manuscripts were preprints, so their remaining scientific rigor required assessment through peer review and other quality-control mechanisms.
3. Molecular Scale: From Proteins to Drug Development
At the molecular scale, AI research targets SARS-CoV-2 proteins to understand infection and support treatment discovery. Protein-structure methods use sequence-derived features and multiple architectures, but predicted structures can vary substantially, with some consensus for papain-like protease structures.
- Molecular targets: AI applications target SARS-CoV-2 structural and non-structural proteins, including spike, 3C-like protease, papain-like protease, and ACE2.
- Structure prediction: Protein-structure prediction methods use amino acid sequences and features from similar sequences to estimate residue-distance and angle distributions.
- AlphaFold: AlphaFold converts predicted residue distances and angles into a potential of mean force used for structure construction.
- trRosetta: trRosetta predicts residue distances and orientations with multiple output heads, while jointly learning features relevant to both predictions.
- Comparative findings: Predicted protein structures showed substantial variability across approaches, although some consensus emerged for papain-like protease structures.
3.2 Drug Repurposing
The review identifies four AI-enabled approaches to drug repurposing: knowledge graphs, binding-affinity prediction, molecular docking, and gene-expression analysis. These approaches prioritize existing or known compounds by predicting targets, interactions, or similarity to effective treatments.
- Four AI approaches support drug repurposing: biomedical knowledge graphs, binding-affinity prediction, molecular docking, and gene-expression signature analysis.
- Biomedical knowledge graphs: Knowledge graphs connect proteins, drugs, and other entities to identify candidate therapies, including Baricitinib and Poly (ADP-Ribose) Polym...
- Protein-ligand binding affinities: 10 promising drugs were identified from 4,895 drugs against 8 SARS-CoV-2-related proteins using multitask binding-affinity prediction.
- Molecular docking: Deep Docking narrowed over 1 billion ZINC compounds to 3 million predicted 3C-like protease inhibitors before docking.
- Gene-expression signatures: Gene-expression embeddings can identify therapies with effects resembling those of known treatments by classifying perturbagen-associated signatures.
3.3 Drug Discovery
AI-based drug discovery generates new SARS-CoV-2-targeting compounds and antibodies rather than only repurposing existing drugs. The reviewed approaches combine generative modeling or reinforcement learning with property prediction, docking, or molecular dynamics filtering.
- Generative and reinforcement-learning systems search for novel inhibitors of the SARS-CoV-2 3C-like protease.
- Novel compounds: Zhavoronkov et al. use protein structures, co-crystallized ligands, and homology models with 28 models per input type to generate candidates.
- Novel compounds: Bung et al. train on 1.6 million ChEMBL molecules, adapt to protease inhibitors, and propose 31 candidate inhibitors after filtering and docking.
- Novel compounds: Nguyen et al. identify 15 novel candidate drugs while also analyzing two proposed HIV drugs for estimated SARS-CoV-2 efficacy.
- Antibody discovery: A VirusNet dataset of 1,933 antigen-antibody sequences supports classifiers predicting whether antibodies neutralize antigens.
- Antibody discovery: 2,589 mutated SARS coronavirus antibody sequences were filtered using predicted effectiveness, stability, validity, and molecular dynamics to propose 8 antibodies.
3.4 Vaccine Discovery
AI-supported vaccine discovery focuses on identifying viral epitopes and determining whether they can be presented by diverse HLA-encoded MHC proteins. The review also notes that candidate vaccines using ML had reached clinical evaluation, although their methods were scarcely disclosed.
- Vaccine design must identify suitable epitopes and ensure presentation by MHC proteins associated with diverse HLA alleles.
- Epitope discovery: Neural networks identified 405 potential T-cell epitopes presented by MHC I or MHC II proteins and two B-cell epitopes on the S-protein.
- Clinical evaluation: Three candidate vaccines reporting ML use had been approved for clinical evaluation, but corporations disclosed very limited methodological information.
3.5 Improving Viral Nucleic Acid Testing
Machine learning is being used to improve viral nucleic-acid testing by designing CRISPR assays and classifying sequence data. These approaches target faster processing, greater specificity, and improved discrimination among coronaviruses.
- ML combined with CRISPR designs assays for 67 respiratory viruses, including SARS-CoV-2.
- CRISPR assay design: The assay-design models are predicted to be sensitive, specific, and diverse in genome coverage while potentially reducing processing time and false positives.
- Sequence classification: A CNN classifies nucleic-acid sequences associated with SARS-CoV-2 against other human coronaviruses and genome sequences containing ORF1ab.
- Related sequence analysis: ML-based sequence analyses also investigate infection severity and infectivity using protein sequences from different coronaviruses.
4. Clinical Scale: From Diagnosis to Outcome Predictions
At the clinical scale, AI supports COVID-19 diagnosis through medical imaging and non-invasive measures, while also predicting patient outcomes and healthcare resource needs. These approaches remain subject to interpretability, deployment, clinical validation, and regulatory requirements.
- Diagnosis: RT-PCR limitations in resources, specimen collection, analysis time, and performance have driven interest in medical imaging for COVID-19 screening and diagnosis.
- Diagnosis: Machine-learning systems use CT scans for binary or multi-class COVID-19 classification with architectures including Inception, UNet++, and ResNet.
- Diagnosis: X-ray imaging offers a potentially more accessible and portable alternative to CT scanners or while awaiting RT-PCR results.
- Clinical Translation: Clinical use requires interpretable predictions, deployment in mobile and low-resource settings, diverse dataset validation, workflow effectiveness, and regulatory quality assurance.
- Diagnosis: Human-in-the-loop CT analysis reduced radiologists’ analysis time from over 30 minutes initially to under 5 minutes after 200 annotated examples.The system also performed segmentation, infection-region counting, and severity assessment.
- Outcome Prediction: AI studies predict hospitalization, ARDS, mortality risk, disease severity, and long-term hospitalization using clinical data, blood tests, CT scans, X-rays, or combinations.One hybrid study identified LDH and CRP among relevant clinical features.
- Resource Planning: Forecasting ICU occupancy and matching ICU beds, ventilators, and protective equipment to fluctuating demand support pandemic planning in overstretched health systems.
5. Societal Scale: Epidemiology and Infodemiology
At the societal scale, AI applications address both epidemiology and infodemiology. They augment epidemiological prediction while analyzing information spread and considering interventions to slow or halt the infodemic.
- Societal-scale AI applications span epidemiology and infodemiology, covering disease trends as well as information spread and interaction.
5.1 Epidemiology
AI epidemiology research applies machine learning to forecasting, regional comparison, intervention modeling, risk assessment, and estimating asymptomatic infection. The literature also shows that model choice, data limitations, and local heterogeneity constrain interpretation and transfer.
- Modeling and Forecasting Statistics: COVID-19 epidemiological modeling seeks to forecast transmission and national or local statistics relevant to public-health interventions.
- Modeling and Forecasting Statistics: Forecasting studies use architectures including LSTM-GRU networks and CNNs reshaping numerical data into images, but architecture, hyperparameters, and datasets interact non-trivially.
- Modeling and Forecasting Statistics: Internet searches, news activity, and mechanistic-model forecasts can be combined with health data to produce short-term forecasts of COVID-19 statistics.
- Regional Comparison: Transferred forecasting models require tailoring to local contexts because demographic characteristics and cultural norms may differ across countries.
- Regional Comparison: Country clustering can support cross-region prediction, but heterogeneous data collection and reporting limit comparability.
- Intervention Modeling: Autoencoder latent features and neural-network regularizers model similarities among regions and the changing strength of quarantine interventions.
- Risk Assessment: Risk scores simplify outbreak trends for rapid analysis but may be non-robust to changes in underlying data or coverage and should be interpreted cautiously.
- Bayesian Analysis: 17.9% of Diamond Princess patients were estimated to be asymptomatic, although applicability to the broader population was unclear.
5.2 Infodemiology
Infodemiology examines the spread, quality, and social effects of COVID-19 information, including misinformation, disinformation, and hate speech. AI-supported monitoring and interventions include social-media analysis, fact-checking, official-information dissemination, and chatbots.
- Infodemic Context: The infodemic involves an overabundance of accurate and inaccurate information that makes trustworthy guidance difficult to find.
- Infodemic Context: Rapid growth in COVID-19 scientific publications and reliance on preprints make quality assessment and critical synthesis increasingly important.
- Spread and Interaction: Information-propagation research analyzes Twitter themes, myths, unreliable-post exposure, cross-platform engagement, advertisements, and hate speech.
- Hate Speech: Studies report increasing hateful, malicious, and Sinophobic COVID-19 content across social-media channels, motivating intervention to protect vulnerable groups.
- Positive Action: Interventions include WHO’s EPI-WIN coordination, collaboration with platforms and search companies, content curation, automated fact-checking, and similarity matching to WHO recommendations.
- Positive Action: Multilingual chatbots can disseminate official information while relieving pressure on question-and-answer hotlines.
6. Datasets and Resources
The review surveys datasets and resources supporting AI research on COVID-19 across epidemiological, textual, biomedical, clinical, genomic, and drug-discovery domains. It emphasizes both expanding data-sharing initiatives and persistent limitations in diagnostic data availability, model sharing, and expert annotation.
- AI research on COVID-19 requires substantial data and computing power, motivating dedicated datasets and collection efforts.
- Case Data: Case datasets support epidemic tracking, growth-rate calculation, and assessment of preventive measures, with public repositories aggregating data from health organizations.
- Case Data: CHIME uses SIR modeling to estimate infections, predict outcomes under specified circumstances, and plan hospital-bed requirements.
- Case Data: Mobility datasets are being used to assess changes in movement and their impact on local epidemic evolution.
- Textual Data: NLP can mine scientific articles, news, and social media to extract passages, topics, and indicators for stakeholders.
- Textual Data: CORD-19 and other literature resources support questions about transmission, risk factors, interventions, vaccines, ethics, and medical-care practices.
- Textual Data: Twitter, institutional, news, and television datasets can support monitoring of misinformation, rumors, and official messaging.
- Biomedical Data: Biomedical data across clinical and molecular scales helps models account for patient variation, viral evolution, structure, and therapy effectiveness.
7. Discussion
The discussion finds that AI and ML applications span drug discovery, diagnosis, clinical prediction, epidemiology, and infodemiology, but few reviewed systems are operationally mature. It calls for shared data, multidisciplinary partnerships, open science, and international cooperation to translate research into adaptable global solutions.
- Discussion: AI and ML can support COVID-19 responses across drug discovery, diagnosis, clinical outcome prediction, epidemiology, and infodemiology, although few systems are operationally mature.
- Discussion: A research road map and application funnel can organize how AI assists the current pandemic, later stages, and future pandemics.
- Calls for Action: Open repositories for scalable data and model sharing could accelerate model development and enable knowledge transfer between medical institutions.
- Calls for Action: Deploying AI in the pandemic requires diverse, complementary teams and long-term multidisciplinary partnerships, including potential work in robotics and logistics.
- Calls for Action: Open science and international cooperation can help share proven solutions, adapt them locally, and build capacity in health systems with fewer resources.
- Conclusion: The review presents itself as an initial step toward identifying valuable AI domains, collaboration opportunities, and research agendas for current and future pandemics.