ICH Q2(R2) took effect in 2024 as the first major revision of the analytical procedure validation guideline in nearly two decades, and it changes what regulated laboratories must document when a validated method relies on a multivariate or machine learning (ML) model. For analytical scientists building or reviewing artificial intelligence (AI)-assisted analytical methods, the guideline's provisions on purpose-based classification, specificity, and data-driven procedures now set the practical baseline for defensible method validation.
Key Takeaways
- ICH Q2(R2) extends validation principles to multivariate analytical procedures, the entry point for machine learning methods, even though the guideline does not use the term artificial intelligence (AI) directly.
- Data-driven analytical methods must still demonstrate specificity, accuracy, and precision requirements tied to their intended purpose, the same way classical procedures are.
- Retraining or updating an AI-assisted model's parameters can trigger the same transfer and revalidation considerations as a classical method change.
- The FDA and the EMA have issued separate AI-specific guidance that complements, but does not replace, ICH Q2(R2) validation requirements.
- Regulated laboratories need documented risk assessments and model credibility evidence before deploying AI-assisted methods for release or stability testing.
ICH Q2(R2) Changes for AI-Assisted Method Validation
ICH endorsed the current Q2(R2) text in November 2023, and both the U.S. Food and Drug Administration (FDA) and the European Medicines Agency (EMA) had adopted it as official guidance by mid-2024, replacing the Q2(R1) version that had governed method validation since 2005. The revision's most consequential change for analytical scientists working with data-driven methods is the explicit inclusion of validation principles for multivariate analytical procedures, including those built on spectroscopic or chromatographic data and calibrated using chemometric or ML algorithms.
Working in analytical science?
Register for a FREE Separation Science account to subscribe to the Separation Science Newsletter.
That inclusion matters because multivariate procedures had previously been validated by analogy to univariate methods, without dedicated guidance addressing calibration models, training data, or independent sample sets. Q2(R2) does not use the term artificial intelligence anywhere in its text. Its relevance to AI-assisted methods comes through the quality guideline framework ICH maintains for multivariate procedures, and through separate guidance the FDA and the EMA have issued specifically to address AI and ML models in the medicinal product lifecycle. Method validation sits downstream of the broader AI tools for separation science workflows that are reshaping method development, spectral interpretation, and process chromatography, and validation is where a laboratory must formally account for any of those tools once they touch a release or stability result.
For analytical scientists, the practical significance is that a chemometric or ML model no longer sits in a regulatory gray area simply because it processes multivariate spectral or chromatographic data rather than a single detector response. The guideline gives such a model an explicit place in the same purpose-based validation framework that governs assay, impurity, and identification methods, rather than leaving reviewers and sponsors to negotiate an ad hoc validation approach for each new data-driven technique that reaches the laboratory.
ICH Q2(R2) Provisions for Data-Driven and Machine Learning Methods
The multivariate procedure provisions in Q2(R2) require a validated method to distinguish clearly between the calibration set used to build a model and the independent samples used to evaluate it. Independent samples are defined as samples not included in the calibration set itself, though the guideline explicitly permits them to come from the same batch as calibration samples, provided they were not part of the data used to build the model. That distinction, sample-set membership rather than batch origin, parallels the training and test set separation used to avoid overfitting in ML workflows.
This framing gives analytical scientists a regulatory vocabulary that maps onto ML practice without requiring a separate AI-specific validation standard. A chemometric model built from near-infrared spectra or an ML-based peak classification algorithm is evaluated against the same calibration and independent sample requirements that the finalized US guidance sets for any multivariate analytical procedure.
Q2(R2) also explicitly permits justified use of prior knowledge and development data as part of a validation package, provided the reliance is scientifically supported. For an AI-assisted method, that provision matters directly: development-stage data used to build or pretrain a model can, with appropriate justification, contribute to the validation record rather than requiring a fully independent validation dataset generated from scratch.
AI Method Validation Categories: Specificity, Accuracy, and Precision Requirements
ICH Q2(R2) classifies analytical procedures by their intended purpose, such as assay and potency, impurity testing, and identification, and ties the required performance characteristics to that purpose through the analytical target profile concept it shares with ICH Q14. An AI-assisted method is classified the same way any other analytical procedure would be, based on what it is intended to measure rather than on the computational approach behind it. An ML model used for assay quantitation is held to the same specificity, accuracy, and precision expectations as any other assay procedure, regardless of whether its underlying algorithm is a partial least squares model or a neural network.
Specificity, accuracy, and precision remain the core performance characteristics a data-driven method must demonstrate. For models trained on complex spectral or chromatographic datasets, specificity testing typically requires deliberately challenging the model with samples containing known interferents, degradation products, or matrix variability that was not represented in the calibration set. The table below summarizes how these expectations shifted from Q2(R1) to Q2(R2).
| Validation element | ICH Q2(R1) approach | ICH Q2(R2) approach for data-driven methods |
|---|---|---|
| Scope | Focused on univariate analytical procedures | Explicitly addresses multivariate and data-driven procedures |
| Calibration and test data | No dedicated distinction required | Requires calibration set and independent sample sets to be kept separate |
| Specificity demonstration | Interference challenge testing for a single response | Interference and matrix variability challenge testing across the model's input space |
| Prior knowledge | Limited guidance on reuse of development data | Explicit allowance for justified use of development and prior study data |
Accuracy and precision assessments for AI-assisted methods generally require a larger and more representative sample set than a comparable univariate procedure, since model performance can vary across the range of inputs it will encounter in routine use rather than at a single concentration or condition.
Interpretability is not a formal validation category under Q2(R2), but it shapes how easily a laboratory can defend a data-driven method to a reviewer or auditor. A partial least squares or principal component model whose contributing variables can be examined directly is generally easier to justify against the specificity requirement than a deep learning model whose internal decision process is harder to inspect, even when both achieve comparable accuracy. That difference does not disqualify more complex models, but it often means additional documentation is needed to demonstrate why a less transparent model's output can be trusted for a given validation category.
AI-Assisted Method Transfer and Revalidation Under ICH Q2(R2)
Method transfer and revalidation triggers under Q2(R2) apply to AI-assisted procedures in the same way they apply to classical ones, but the practical trigger points differ. Retraining a model with new data, adding analytes to its scope, or moving it to a different instrument platform each functions as a change to the analytical procedure and should be evaluated against the same partial or full revalidation criteria used for a classical method change.
USP's analytical procedure lifecycle chapter provides a complementary, risk-based framework for deciding how much revalidation a given change requires, and it applies equally to data-driven procedures. On the regulatory side, the EMA's AI reflection paper sets an expectation that sponsors define the risk level and intended use of an AI model before relying on its output, a principle that maps directly onto the transfer and revalidation decisions method validation scientists already make under Q2(R2) and ICH Q14.
Method transfer between sites adds a further consideration for multivariate and AI-assisted procedures specifically. A model validated on instrumentation and reagent lots at one site can encounter subtly different spectral or chromatographic baseline characteristics at a receiving site, even on nominally identical instrument platforms, so transfer verification for a data-driven method typically needs to confirm performance under the receiving site's actual operating conditions rather than relying solely on the originating site's validation data.
AI Method Validation Steps for Regulated Laboratories
Building a defensible validation package for an AI-assisted analytical method under Q2(R2) generally follows a consistent sequence:
- Define the intended use and validation category for the method before selecting or training a model.
- Separate calibration and independent sample sets, and document the rationale for sample selection.
- Challenge the model with interferents, matrix variability, and edge cases the training data did not fully represent.
- Establish acceptance criteria for accuracy, precision, and specificity appropriate to the validation category.
- Document a change control plan specifying what triggers partial versus full revalidation if the model is retrained or its scope changes.
Laboratories that already maintain AI-assisted quality control and validation programs will recognize much of this sequence from broader QC governance work, since the same risk-based documentation expectations extend to out-of-specification investigations that touch AI-generated results. Where a validated method later produces an unexpected result, AI-assisted out-of-specification investigations should reference the same validation and revalidation documentation established during method development, and the FDA's AI guidance for drug development outlines a risk-based credibility framework, still in draft form as of this writing, that many laboratories are adapting for this purpose ahead of finalization. Method development teams working from the earlier stages of the workflow can find complementary guidance in resources on AI-assisted chromatographic method development, which addresses how models are built before they reach the validation stage covered here.
The five-step sequence above is easiest to apply consistently when it is written into a laboratory's standard operating procedures for method validation rather than handled case by case for each new model. A documented policy also gives quality assurance and regulatory affairs staff a shared reference point when an inspector or reviewer asks how a specific AI-assisted method was validated, since the answer can point to a consistent internal process rather than a justification assembled after the fact for that particular method.
ICH Q2(R2) and AI Method Validation: Building Regulatory-Ready Programs
ICH Q2(R2) does not create a separate validation pathway for AI. Instead, it extends the existing purpose-based validation framework, and the specificity, accuracy, and precision requirements that come with it, to cover the multivariate and data-driven procedures that AI-assisted methods represent. Analytical scientists who already understand calibration and independent sample separation have most of the vocabulary they need to build a Q2(R2)-compliant validation package for an AI-assisted method.
The remaining work is procedural: documenting model risk and intended use, setting clear revalidation triggers, and connecting method validation records to the complementary AI-specific guidance the FDA and the EMA have each published. Laboratories that treat those three elements as part of standard method validation, rather than as a separate AI compliance exercise, are best positioned to meet what regulated laboratories now need to know under ICH Q2(R2).
This article was produced under Separation Science's AI Editorial Guidelines.
