We use cookies to ensure our website works properly and to personalise your experience. Cookies policy
1Department of computer science & engineering (AI and ML), S J C Institute of Technology, Chikballapur, Karnataka, India.
2Department of pharmaceutical sciences, Sri K.V. College of Pharmacy, Chikballapur, Karnataka, India.
Pharmaceutical analysis underpins the identity, purity, potency and safety of medicines. Conventional workflows based on chromatography, spectroscopy and mass spectrometry are reliable but depend on manual interpretation and generate data volumes that are hard to exploit fully. Artificial intelligence (AI) and machine learning (ML) offer data-driven tools for peak detection, retention-time and spectral prediction, impurity identification, dissolution and stability modelling, anomaly detection and real-time process analytical technology (PAT). This narrative review, based on a structured search of peer-reviewed and regulatory sources published mainly from 2021 to 2026, compares traditional and AI-enabled workflows, summarises ML methods and their applications across analytical areas, and examines explainable AI, digital twins, generative AI and autonomous laboratories. In the sources reviewed, the strongest evidence concerns spectroscopy-based prediction for PAT and dissolution and retention-time prediction supported by large public datasets, whereas evidence for stability forecasting, rare-impurity identification and autonomous laboratories is limited and seldom validated at industrial scale. Regulatory thinking is converging on risk-based, context-of-use credibility assessment and lifecycle management. We propose a conceptual Intelligent Pharmaceutical Analysis Framework (IPAF) linking instrumentation, data integrity, AI analytics, decision support and governance, and identify priorities: standardised validated datasets, explainable and uncertainty-aware models, and prospective multi-site validation.
Pharmaceutical analysis determines active substances, impurities, excipients and dosage forms throughout the product lifecycle, from development to quality control, stability testing and regulatory approval 1,2. It relies on chromatography (HPLC, UPLC, GC, LC-MS, GC-MS), spectroscopy (UV-Vis, FTIR, Raman, NIR, NMR) and capillary electrophoresis. Despite instrumental progress, workflows face large and complex datasets, time-consuming and analyst-dependent interpretation, demanding method development, limited predictive capability, rising regulatory expectations and high costs, pressures that grow with continuous manufacturing and its dense data streams 3,4.
Artificial intelligence (AI) denotes computational systems performing tasks associated with human intelligence; machine learning (ML) learns relationships from data, and deep learning (DL) uses multilayer neural networks 5. AI is widely used in pharmaceutical research and development 6,7, and reviews now address pharmaceutical analysis 8,9. Its use aligns with Quality by Design, process analytical technology (PAT) and risk-based quality systems 10,11; ICH Q2 (R2) and Q14 introduce a lifecycle approach that includes multivariate procedures 12,13, and regulators have begun to publish AI-specific expectations 14,15.
Existing reviews cover AI across the pharmaceutical value chain 16, a single technology such as neural networks in PAT 17 or digital twins 18, or analysis at a general level 8,9. To the best of our knowledge, an integrated treatment that contrasts traditional and AI-enabled workflows, grades the maturity of evidence by analytical area, maps AI use onto regulatory expectations and proposes a deployable architecture is lacking. This review therefore (i) compares traditional and AI-enabled analysis, (ii) summarises ML methods and applications, (iii) relates them to regulatory expectations, (iv) identifies research gaps and (v) proposes the Intelligent Pharmaceutical Analysis Framework (IPAF).
MATERIALS AND METHODS
This is a narrative review supported by a structured search. PubMed, Scopus, Web of Science, ScienceDirect, SpringerLink, IEEE Xplore and Google Scholar were searched for 2021–2026 using Boolean combinations of terms such as “artificial intelligence in pharmaceutical analysis”, “machine learning in analytical chemistry”, “deep learning in pharmaceutical sciences”, “AI in chromatography/spectroscopy”, “process analytical technology”, “digital twins in pharmaceuticals”, “explainable artificial intelligence”, “AI agents in pharmaceutical manufacturing” and “AI in dissolution and stability studies”; reference lists and regulatory websites (FDA, EMA, ICH, PIC/S) were also screened. English-language peer-reviewed articles, reviews, conference papers and guidance documents addressing AI, ML or advanced analytics in chromatographic, spectroscopic, quality-control, PAT, dissolution, stability or impurity applications were included; duplicates, unrelated or purely clinical studies, and papers lacking methodological detail were excluded. Earlier foundational works (algorithm papers, textbooks, ICH/FDA guidelines) are cited as background, not primary evidence. Extracted items (technique, AI method, data, metrics, limitations) were synthesised narratively by theme; no pooling or formal risk-of-bias scoring was attempted.
RESULTS AND DISCUSSION
3.1 Traditional versus AI-based analysis
Traditional methods combine validated techniques with expert interpretation and remain the reference for compliance 1,2. AI adds automated processing, pattern recognition and prediction, with reported benefits in speed, reduced analyst variability and preventive quality management 17; these depend on representative data, validation, model maintenance and human oversight (Table 1). Complexity does not guarantee superiority: ANN models predicted dissolution more accurately than PLS in one study 19, yet no single retention-time model is best across datasets 20.
Table 1: Traditional Versus Ai-Enabled Pharmaceutical Analysis (Conceptual)
| Parameter | Traditional | AI-enabled (potential and caveats) |
|---|---|---|
| Data processing | Manual or semi-automated review by analysts | Automated feature extraction and classification; needs validated pipelines and oversight |
| Method development | Experimentation and design of experiments | Predictive models can narrow experiments; transfer between systems is limited 20 |
| Process monitoring | Off-line or at-line sampling | In-line PAT with multivariate or ML models 4,17 |
| Spectral interpretation | Expert evaluation, PLS calibration 21,22 | ML/DL classification and quantification 23; needs calibration sets and maintenance |
| Dissolution testing | Destructive, time-consuming tests | Surrogate models from spectra or process data, at research stage 19,24 |
| Impurity profiling | Manual spectral interpretation | Candidate structures from MS/MS; orthogonal confirmation needed 25,26 |
| Cost and documentation | High labour; manual reports | Possible long-term savings; validation, security and skills investment 16,27 |
3.2 Machine learning techniques
Supervised methods (random forests 28, support vector machines 29, gradient boosting 30) support quality prediction, spectral classification and dissolution modelling; unsupervised methods (PCA, clustering) support exploration and outlier detection; deep networks (ANN, CNN, LSTM 31, transformers 32) model nonlinear and sequential data (Table 2). Chemometric PLS 21 remains the essential baseline.
Table 2: Machine Learning Algorithms and Pharmaceutical Applications
| Algorithm | Category | Illustrative applications | Strengths / limitations |
|---|---|---|---|
| Random forest | Supervised | Quality prediction, dissolution modelling | Robust to noise / limited transparency |
| Support vector machine | Supervised | Spectral classification, impurity detection | Good for high dimensions / parameter sensitive |
| Gradient boosting | Supervised | Stability prediction, process optimisation | High accuracy on tabular data / overfitting risk |
| PCA, clustering | Unsupervised | Chemometrics, batch grouping, outliers | Simplify data / linear or cluster assumptions |
| ANN | Deep learning | Dissolution prediction, process modelling | Captures nonlinearity / needs adequate data |
| CNN | Deep learning | Image and spectral classification | Automatic features / large labelled sets |
| LSTM | Deep learning | Process monitoring, time-series forecasting | Handles sequences / training complexity |
| Transformer | Deep learning | Spectral interpretation, documentation | Contextual learning / costly, hard to validate |
3.3 Chromatographic analysis
AI can automate peak integration, deconvolution, retention-time (RT) prediction, method optimisation and outlier detection (Figure 1). RT prediction is best documented: large datasets such as METLIN SMRT (about 80,000 molecules) and RepoRT (88,325 records from 49 LC systems) enabled deep-learning models, which generally outperform classical regressors but degrade for dissimilar molecules 20,33–35. This evidence comes mainly from metabolomics; transfer to validated pharmaceutical assays is unproven. For LC-MS, CSI:FingerID, SIRIUS 4, CFM-ID 4.0 and MSNovelist propose structures for unknowns 25,26,36,37; such candidates are hypotheses requiring confirmation.
Figure 1: AI-assisted Chromatographic Workflow (Conceptual Schematic; No Experimental Data).
3.4 Spectroscopic analysis
Multivariate calibration is long established for NIR and Raman 22, and ML/DL extends it to mixture identification, blend monitoring and process control 23,38,39. FTIR (material identification, counterfeit screening) and NMR (structure and impurity elucidation) applications are described in the original draft and need supporting pharmaceutical references.
3.5 Quality control and PAT
PAT encourages timely measurement of quality attributes 11, and ANNs can integrate data from several PAT tools for real-time monitoring and feedback 17. NIR-based ML surrogates have supported real-time release testing of dissolution 40, an in-line surrogate was shown in continuous blending 24, and interpretable ANNs gave retrospective process understanding 41 (Figure 2). Model maintenance and calibration transfer remain open issues.
Figure 2: AI-integrated PAT Framework with Closed-Loop Feedback (Conceptual Schematic; No Experimental Data).
3.6 Dissolution, stability and impurity profiling
ANNs trained on NIR or Raman spectra predicted dissolution profiles of extended-release tablets (37 manufacturing settings) more accurately than PLS 19, and of immediate-release tablets with f2 above 59 42; studies are formulation-specific with little cross-formulation validation. Stability follows ICH protocols and stability-indicating methods 43,44; AI could support shelf-life prediction and QbD formulation optimisation 45, but few prospectively validated examples were found. For impurities, MS-based tools can generate candidate structures 26,37, but rare pharmaceutical impurities are underrepresented in training data and confirmation remains necessary.
3.7 Emerging technologies
Explainable AI (for example Shapley-value attributions 46; reviews 47) supports scrutiny but not validation. A digital twin mirrors a physical process 48; a 2020 review found no full digital twin in pharmaceutical manufacturing 18, with later reviews broadening the scope 49. Language models are explored for documentation 27 and LLM agents have planned and executed chemistry experiments 50, as have robotic platforms 51, but these are not quality-control deployments; hallucination, confidentiality and validation are obstacles, and draft EU GMP Annex 22 was reported to exclude generative AI from critical GMP use 52. Federated learning is a candidate answer to data scarcity and confidentiality.
3.8 Regulatory and validation considerations
ICH Q2 (R2) and Q14 provide a lifecycle view of analytical procedures including multivariate methods 12,13; the EMA reflection paper covers AI across the medicinal product lifecycle 14; the FDA draft guidance (January 2025) proposes a seven-step, risk-based credibility assessment tied to a defined context of use 15; and draft Annex 22, with revised Annex 11, would set GMP requirements for AI in manufacturing (consultation July–October 2025) 52. Practically, laboratories should define the context of use first; separate training, validation and test data; apply data-integrity principles; monitor drift; place model changes under change control; and keep human oversight and documentation.
3.9 Literature matrix and research gaps
Table 3: Literature Survey Matrix (Representative Sources)
| No. | Source | Area / technique | Reported finding | Gap for this review |
|---|---|---|---|---|
| 1 | Nagy et al. 17 | PAT; ANN review | ANNs combine multiple PAT data streams for real-time monitoring and feedback | Industrial-scale validation |
| 2 | Dharmamoorthy et al. 8 | Analysis; AI review | Covers quality control, process optimisation, regulatory compliance and challenges | No technique-level evidence grading |
| 3 | Kowtharapu et al. 9 | Analysis; AI models | Overviews AI models and conceptual instrument designs | Designs not validated |
| 4 | Liu et al. 20 | RT prediction; DL review | Large datasets enabled DL; no single best model | Benchmarks; pharma transfer |
| 5 | Pan et al. 38 | Raman; ML/DL review | Reviews preprocessing and AI for mixture identification | Real-time pharma use |
| 6 | Galata et al. 19 | Dissolution; ANN vs PLS | ANN beat PLS on NIR/Raman data | Single formulation |
| 7 | Mészáros et al. 24 | RTRT; NIR, PLS, ANN | In-line blend uniformity and dissolution prediction | Multi-product validation |
| 8 | Chen et al. 18 | Digital twins; review | No full digital twin yet in pharmaceutical manufacturing (2020) | Analytical-lab twins |
Table 4: Research Gaps and Suggested Directions
| Area | Gap | Suggested direction |
|---|---|---|
| Chromatography | Real-time peak models; transfer of RT models to validated pharma methods | Pharmaceutical benchmark datasets; prospective evaluation |
| Spectroscopy | Few large validated spectral datasets; calibration transfer | Shared reference data; transfer learning |
| Quality control | Explainable, regulation-compliant systems | Context-of-use validation; uncertainty-aware models |
| Dissolution / stability | Formulation-specific models; little long-term validation | Multi-site external validation; real-time confirmation |
| Impurity profiling | Scarce data for rare impurities | Curated libraries; orthogonal confirmation |
| Digital twins / agents | Early-stage adoption; outside regulated QC | Pilot twins; human-in-the-loop agents under GxP |
3.10 Proposed Intelligent Pharmaceutical Analysis Framework (IPAF)
IPAF is a conceptual six-layer architecture (Figure 3): instrumentation and sensors; data infrastructure and integrity; data processing; AI/ML analytics; decision support and control; and governance, validation and regulatory alignment, with explainability, data governance and human oversight as cross-cutting requirements. Autonomy is staged from assisted to automated to autonomous; most reviewed evidence supports the first two. Adoption should begin with low-risk use cases and assisted analytics, move to validated automation for defined contexts of use, and consider closed-loop operation only after sustained monitored performance.
Figure 3: Conceptual Layers of IPAF (Proposed in This Review; Not Experimentally Validated).
3.11 Limitations and future perspectives
This narrative review is limited to English sources and listed databases, without quantitative synthesis; much RT and MS evidence comes from metabolomics; regulatory documents are evolving; and IPAF is unvalidated. Priorities are standardised curated pharmaceutical datasets (including impurity and degradation data), explainable and uncertainty-aware models, prospective multi-site validation, practical lifecycle management with drift monitoring, privacy-preserving collaboration, and carefully governed use of generative AI beginning with non-critical tasks.
CONCLUSION
AI and ML are becoming valuable complements to established analytical methods. The strongest evidence supports spectroscopy-based prediction for PAT and dissolution and data-driven retention-time and structure prediction; evidence for stability forecasting, rare-impurity identification, analytical digital twins and autonomous laboratories is limited. Barriers include scarce standardised datasets, interpretability, data security, model maintenance and validation frameworks acceptable to regulators. Regulation is converging on risk-based, context-of-use credibility assessment and lifecycle control, and IPAF offers a conceptual route toward connected, increasingly automated analytical laboratories.
ACKNOWLEDGEMENT
The authors thank the researchers whose published work underpins this review.
CONFLICT OF INTEREST
The authors declare no financial, commercial, professional, institutional or personal conflicts of interest. No external funding was received.
REFERENCES
Nagarjuna G. R., Veerap Anusha, Artificial Intelligence and Machine Learning Applications in Pharmaceutical Analysis: From Traditional Analytical Approaches to Intelligent Pharmaceutical Systems – A Comprehensive Review, Int. J. of Pharm. Sci., 2026, Vol 4, Issue 10, 4866-4874. https://doi.org/10.5281/zenodo.23829637
10.5281/zenodo.23829637