Artificial intelligence tools swept through pharmaceutical research laboratories in the years after the COVID-19 pandemic, but a new systematic review suggests their impact has been concentrated almost entirely at the earliest, computational end of the drug development pipeline. Researchers from the National University of Singapore analysed 57 peer-reviewed studies published between 2023 and 2025 and found that AI-enabled Pharma 4.0 technologies and predictive modelling were predominantly applied to early-stage drug discovery and computational research, with reported benefits of more efficient and accurate processes, while evidence of clinically meaningful pharmacological effects remained scarce.
The review, published in Frontiers in Artificial Intelligence, followed the Preferred Reporting Items for Systematic Review and Meta-Analyses (PRISMA) method. The team searched PubMed and Scopus, two of the largest multidisciplinary biomedical databases, and identified 593 articles. After removing four duplicates, screening titles and abstracts, and applying eligibility criteria to full texts, 57 articles remained for detailed analysis. The search was conducted between 25 August 2025 and 5 February 2026, and the post-pandemic window was deliberately defined as 2023 to 2025, beginning after the World Health Organization ended COVID-19’s Public Health Emergency of International Concern status on 5 May 2023.
Pharma 4.0 refers to the pharmaceutical industry’s adoption of the Industry 4.0 toolkit: artificial intelligence, machine learning, the Internet of Things, big-data analytics, robotics and cloud computing, integrated across the pharmaceutical product life cycle. The pandemic exposed the vulnerabilities of the older, manual, fragmented way of working. Border restrictions disrupted supply chains, remote work strained workflows, and shortages of critical medications were amplified, prompting companies to accelerate the transition towards digitalised, interconnected systems that share data in real time.
To characterise what the included studies actually measured, the reviewers catalogued the performance metrics reported in each paper and grouped them into three categories: drug candidate evaluation metrics, computational evaluation metrics, and AI/ML model evaluation metrics. Across the 57 articles they identified 152 individual metric mentions. Because a single study could report several metrics, percentages were calculated against the 152 mentions rather than the 57 articles. AI/ML model evaluation metrics dominated, accounting for 80.9 percent of all mentions, split between classification metrics at 60.5 percent and regression metrics at 20.4 percent. Drug candidate evaluation metrics made up 9.9 percent and computational evaluation metrics 9.2 percent.
The individual metrics tell a similar story. AUC-ROC, in its various notations, was the most frequently reported classification metric at 12.5 percent of mentions, followed by sensitivity or recall at 17 mentions and accuracy at 16. RMSE was the most common regression metric at 6.6 percent. IC50 or pIC50, a laboratory measure of drug potency, led the drug candidate category with 9 mentions, and RMSD, a structural measure used in molecular simulations, topped the computational category with 6. The reviewers interpreted this distribution as evidence that many predictive models are built for categorical prediction tasks, such as identifying active compounds from large chemical libraries, rather than for downstream pharmacological evaluation.
The technology breakdown reinforced the same pattern. Deep learning was the most common technology in the reviewed studies at 35.1 percent, followed by artificial intelligence or machine learning and virtual screening at 14.0 percent each. The most common research area was drug discovery and design at 54.4 percent, with computational prediction at 28.1 percent. Almost all the included literature described applications in the pre-commercialisation stage, where a final product has yet to reach the market. In drug repurposing studies, for example, researchers worked on existing drug candidates rather than entirely new compounds.
The studies themselves spanned a striking range of computational approaches applied to coronavirus-relevant targets. One team reported an AI-accelerated protein-ligand docking method for SARS-CoV-2 that was 100-fold faster with no significant change in detection. Another used machine learning to streamline the development of liposomal drug delivery systems, achieving an R-squared of 0.74 and an AUC of 0.98. A random forest model for predicting SARS-CoV-2 main protease binding reached an accuracy of 0.514, while a support vector machine study of cathepsin L inhibitors reported an R-squared of 0.9676. Graph neural networks were used to predict drug-target binding affinity and SARS-CoV-2-human protein interactions, and generative adversarial networks were applied to low-data molecular design.
When the reviewers examined the conclusions of the included papers, they found a consistent vocabulary of optimism. The keyword “effective” appeared in 66.7 percent of the articles, often substantiated by metrics such as EC50 values. Some 36.8 percent used “quick” or “time saving”, typically supported by AUROC values, and 24.6 percent used “accurate”, backed by accuracy percentages. Other recurring terms included “cost saving” and “widely applicable”. Yet the reviewers noted that most conclusions addressed the strengths of the models themselves, with little information on how these technologies integrate the different stages of the product life cycle, which they attributed to the infancy of AI in pharmaceutical workflows.
That gap between algorithmic performance and therapeutic reality is the review’s central caution. The positive performance of a machine learning model, the authors write, may not directly guarantee positive therapeutic outcomes of the investigated compound. Because the included articles are technology-centred rather than clinically centred, they cannot be used alone to conclude that the pharmaceutical products being researched are clinically beneficial. The reviewers also assessed risk of bias using the PROBAST tool and found that the majority of articles were at high risk of bias, reflecting methodological limitations inherent to early-stage drug discovery, where studies often lack external validation and are more suitable for hypothesis generation than clinical application.
The authors suggest that future research should move towards predictive model training on compounds’ effects on the body, such as pharmacokinetics and pharmacodynamics, to help bridge the gap between computational performance and clinical outcomes. They also acknowledge limitations: only two databases were searched, the publication window was narrow, publication bias may have inflated reported benefits, and no prospective review protocol was registered. For now, their conclusion is measured. AI-enabled Pharma 4.0 applications show potential utility as complementary tools in early-stage drug discovery, potentially offering shorter lead times and greater productivity, but clinical effectiveness and patient outcomes were not established by the included evidence, and further externally validated, prospective studies are needed to determine whether these computational advantages translate into real-world therapeutic benefit.
Subject of Research: A systematic review of AI-enabled Pharma 4.0 applications and predictive modelling in post-pandemic pharmaceutical development
Article Title: Navigating the Pharma 4.0 revolution: how AI-enabled applications and predictive modelling shaped pharmaceutical development in the post-pandemic era: a systematic review
Article References: Chang, X. Y., Teo, C., & Goh, C. H. (2026). Navigating the Pharma 4.0 revolution: how AI-enabled applications and predictive modelling shaped pharmaceutical development in the post-pandemic era: a systematic review. Frontiers in Artificial Intelligence, 9, Article 1976101. https://doi.org/10.3389/frai.2026.1976101
Image Credits: AI Generated
DOI: 10.3389/frai.2026.1976101
Keywords: Pharma 4.0, artificial intelligence, drug discovery, systematic review, COVID-19, machine learning, predictive modelling, pharmaceutical industry, deep learning, virtual screening, PRISMA, risk of bias
News Source: Louis Brooks. (November 5, 2026). AI drug discovery tools dominated early-stage pharma research after COVID-19, review finds. Scienmag.
