Machine learning in drug discovery promises faster drug screening and fewer failures, but weak data limits gains. See what the latest 2026 review means for pharma.
Machine learning in drug discovery promises faster screening, fewer late-stage failures, and sharper targets. Those promises are being tested by their evidence.
The answer is vital for those funding AI programmes and it looks like it’s nothing new: The models work, but only as well as the data behind them.
Image
What Machine Learning in Drug Discovery Delivers Today
The latest and most wide-ranging review searched databases including PubMed, Scopus, and Web of Science for work from 2014 to 2024 for evidence on machine learning (ML) and deep learning (DL) across the discovery pipeline, with a nanomedicine focus.
The review notes that discovery programmes screen millions of compounds, yet only about 1,000 to 1,500 candidates reach Phase I each year. Just 30 to 50 new drugs win approval annually.
Small gains at early stages therefore compound.
Research output on the topic has surged. Publications on ML and DL rose from about 100 in 2014 to 1,610 in 2024, a 16-fold increase.
The review also gathers headline performance figures from the literature:
Up to 85% accuracy when predicting drug-target interactions.
Up to 40% better molecular property prediction from deep learning.
Up to 30% lower candidate attrition in AI-driven workflows.
20% to 40% faster discovery timelines.
Remember these aren’t ceiling numbers. As a literature review pools results from different studies, no single controlled comparison sits behind the numbers.
Therefore, to truly capitalise off this data, ask which programme, dataset, and stage each figure reflects, and whether your own data resemble it.
Predictive modelling and screening
ML models predict activity, toxicity, and pharmacokinetics from chemical structure.
The review covers quantitative structure-activity relationship (QSAR) models, which link molecular features to biological effect, and ADME predictions for absorption, distribution, metabolism, and excretion. Teams use both to rank compounds before wet-lab work begins.
Virtual screening follows two routes. Ligand-based screening searches for compounds that resemble known actives. Structure-based screening docks candidates into the 3D shape of a target protein.
Both types of screening can reduce the amount of testing needed to reach a credible hit.
One team cited by the review merged physics-based modelling with deep learning. The hybrid cut computing cost while keeping accuracy and improved explainability.
Image
Target identification and validation
Here the review points to protein-ligand modelling, molecular dynamics simulations, and biological pathway analysis.
Pathway tools merge genomic, proteomic, and metabolomic data into cell-level models. Teams then prioritise targets at influential nodes and spot likely side effects early.
Pathway models also predict how blocking one protein ripples through a disease network, which supports target validation.
Deep learning for structures and genomes
Deep learning adds protein structure prediction, medical imaging, and genomic analysis.
It also warns that black-box models and correlation-led results still need experimental validation. Image models, meanwhile, have classified melanoma subtypes from dermoscopic images, for example, and microscopy classifiers can successfully group cell phenotypes in high-throughput screens.
Toxicity, dosing, and manufacturing
The review extends beyond early discovery.
Predictive toxicology models flag risky compounds before clinical trials, and adverse reaction tools mine health records and pharmacovigilance databases.
Additionally, dosing models built on pharmacokinetic data can support personalised regimens for anticoagulants such as warfarin and apixaban, for example.
Manufacturingfeatures too. The authors cite real-time analytics that can cut production costs by 20%, plus formulation models lifting stability by 30% and bioavailability by 40% in oral dosage forms. These figures come from cited studies, so verify them before building a business case.
Find out how Quality-by-Design improves reliability and compliance in pharma and how you can maximise innovation through quality-based pharma development systems.
Case Studies That Test the Claims
The review’s case studies span binding prediction, repurposing, target identification, and new datasets. Each shows where machine learning earns its place and where it needs help.
Predicting how tightly drugs bind
Binding affinity measures how strongly a molecule grips its target, but docking and dynamics simulations are slow and costly.
Convolutional neural networks (CNN) and graph neural network models (GNN) trained on resources such as PDBbind learn the same relationship from known complexes, and the review says they often beat classical docking.
The review also notes that AtomNet was used to flag potential Ebola inhibitors in under a day.
However, the authors admit the weaknesses of neural networks. Many models ignore cellular context and protein movement, and both change how a ligand binds.
Finding new uses for existing drugs
Repurposing offers a quicker route to patients because safety profiles already exist. Models combine drug-target data, gene expression, and clinical outcomes to predict new indications.
One ML framework called Drug Repurposing in Alzheimer’s Disease (DRAID) linked gene lists to Alzheimer’s disease stages using brain expression data.
The review lists ML mining tools such as the Drug Repurposing in Alzheimer’s Disease (DRIAD) and PharmMapper, and cites a trial of fluoxetine against COVID-19 that followed computational predictions of antiviral potential.
It does concede that very few predictions generated so far have actually become a therapy.
Mapping targets through networks
Network analysis treats biology as a graph. Researchers build protein-interaction networks with tools such as STRING and Cytoscape, then rank the best-connected nodes as candidate targets.
The review credits this approach with identifying BRD4 as a regulator of cancer gene programmes, which led to BRD4 inhibitors.
In a separate study cited, researchers used deep learning to find a binding site on the coronavirus spike protein. That site showed lower mutation tolerance and higher conservation than the receptor-binding domain, which could make it a more durable target. This research was used during the pandemic for vaccine research.
Screening for tuberculosis
One cited study screened about 11,576 DrugBank compounds against Mycobacterium tuberculosis.
Aldoxorubicin and quarfloxin inhibited the H37Rv strain and showed activity against multidrug-resistant isolates, which matters because drug resistance remains a global treatment problem.
Docking, simulations, and binding experiments then confirmed that both compounds bind DNA gyrase.
The pattern is very practical. ML narrowed thousands of candidates to a short list, and wet-lab work confirmed it.
Better training material is arriving too. The MISATO dataset pairs molecular dynamics and quantum data for about 20,000 protein-ligand complexes.
Image
Why Does Data Quality Still Hold Progress Back?
The review is candid about weak points.
In nanomedicine, the data problem is more obvious than in more general small-molecule work, and ML in drug discovery, therefore, inherits every flaw in its training sets.
Nanoparticle formulations vary in size, shape, surface charge, composition, and batch reproducibility.
Labs report these traits inconsistently and test them with different protocols. Two particles with near-identical chemistry can behave very differently in the body, so that error propagates straight into models.
The authors flag four recurring weaknesses:
Heterogeneous formulations that make datasets hard to compare.
Protein corona effects, where blood proteins coat a particle and change its biological identity.
Narrow coverage, with datasets skewing towards some particle types and missing many polymeric, lipid-based, and hybrid carriers.
Siloed data types, with no harmonised way to fuse omics, imaging, and pharmacokinetic data.
The protein corona deserves special attention. It forms within moments of exposure to biological fluids, varies with disease context, and shapes cell recognition, immune clearance, and efficacy.
Experimental datasets on it are large but condition-dependent and unstandardised.
Interpretability compounds the issue. Most DL models behave as black boxes, which makes nano-bio predictions difficult to explain mechanistically.
Multimodal learning and knowledge graphs could help, yet the authors describe their use in nanomedicine as in the early stage.
And ultimately, a model that scientists cannot explain will struggle to earn regulator trust.
Cost adds pressure. Nanomedicine needs extra characterisation, long-term toxicity studies, and complex manufacturing, which lengthen development timelines.
Regional differences in nanomedicine rules add documentation burden. ML can trim spend by optimising formulations and flagging toxic candidates early, but only once data standards and validation frameworks mature.
Regulators are trying to help improve the system.
The U.S. Food and Drug Administration’s January 2025 draft guidance sets a credibility framework for AI models that support regulatory decisions, yet the Federal Register notice states that it excludes AI used in drug discovery. Discovery teams should therefore set their own validation standards.
The Looming Threat of Geopolitical Risk in Clinical Supply Chains
Geopolitical risk is reshaping how clinical supply chains are managed, demanding resilience and foresight. Learn why building supply chain resilience is the key for clinical trial success.
Image
The Algorithms Are Ready. Is Your Data?
Machine learning in drug discovery already shortens screening, sharpens target choice, and supports repurposing.
The reported gains include up to 85% accuracy and timelines shortened by 20% to 40%, but each depends on the study and the dataset. Nanomedicine shows this limit most clearly, because inconsistent formulation data weakens every model trained on it.
For drug development teams, four moves are now necessary. Audit your data, demand external validation, and set your own standards while regulators catch up. Fund shared, standardised datasets too, because they gain value across every programme.
At Pharmatica, we separate evidence from excitement in pharmaceutical AI. Our Insights connect data, regulation, and market dynamics, so leaders can invest with confidence.
Pharmatica: Insight. Connection. Impact.
Frequently Asked Questions
What is machine learning in drug discovery?
Machine learning in drug discovery uses algorithms that learn from chemical, biological, and clinical data to predict drug behaviour. Teams apply it to screening, target selection, toxicity prediction, and repurposing, which reduces reliance on laboratory testing.
How accurate is machine learning at predicting drug-target interactions?
A 2026 review reported 85% accuracy as the upper figure for ML models predicting drug-target interactions. That figure is a ceiling drawn from varied studies, so results depend on the dataset, target, and validation method.
What is the difference between machine learning and deep learning?
Deep learning is a subset of machine learning that uses multi-layered neural networks. It handles high-dimensional inputs such as images, genomic data, and protein sequences, while simpler ML models often suit smaller structured datasets.
Why is nanomedicine harder for AI models than small molecules?
Nanoparticles vary in size, surface charge, composition, and batch quality, and proteins coat them in blood. These factors create inconsistent, poorly standardised data, which weakens model accuracy and generalisability.
Does FDA guidance cover AI used in drug discovery?
No. The FDA’s January 2025 draft guidance addresses AI models that support regulatory decisions on safety, effectiveness, or quality. It states that it does not address AI used in drug discovery.
Nicole (BSc Molecular Medicine, Honours Medical Biochemistry) has many years of pharmaceutical experience, having worked for top CROs and biopharma companies for more than a decade.
Did you enjoy the content?
2
Why not support
Nicole Dale
by giving this content a like
Close the Drug Discovery Loop by Leveraging the DMTA Cycle
The DMTA cycle is the engine of modern pharma R&D. Learn how integrated DMTA workflows reduce early-stage attrition and compress drug discovery timelines.
Preclinical vs Clinical Research Strategy for Drug Development
Explore the critical differences between preclinical vs clinical research, and how closing the translation gap can reduce drug development attrition and improve R&D decision-making.
What 13 Years of Data Shows About Drug Innovation in Psychiatry
From 2012 through 2024, the FDA only approved 22 drugs for the treatment of psychiatric conditions. Pharma R&D teams must rethink development strategies to unstall drug innovation in psychiatry
Pharmacoinformatics is redefining drug discovery strategy. This analysis covers the AI platforms, molecular design tools, and regulatory frameworks shaping R&D pipelines in 2026.
Why Precision Alone Isn't Enough in Oncology Drug Discovery
Precision alone isn’t enough in oncology drug discovery. See why execution, biomarkers, and trial design determine scalable success in precision medicine.
Comments (0)