Regulatory Guidance for the Integration of New Approach Methodologies (NAMs) and Artificial Intelligence (AI) in Preclinical Submissions

The Saudi Food and Drug Authority (SFDA) is committed to its role as a leading, science-based international regulator. In alignment with Saudi Vision 2030, and the National Biotechnology Strategy, the SFDA recognizes that transforming healthcare and pharmaceutical regulation necessitates the integration of advanced data science and innovative biomedical technology. Traditionally, drug development has relied extensively on animal models to evaluate safety and efficacy. However, contemporary scientific evidence highlights the inherent limitations of these traditional preclinical models in predicting human physiological responses. The discrepancies, often resulting in high attrition rates during clinical trials, underscore the critical need for human-relevant methodologies. By adopting these advanced approaches, the SFDA aims to enhance the accuracy of clinical predictions, optimize the development lifecycle, and uphold the highest standards of safety and efficacy for the Kingdom.

New Approach Methodologies (NAMs)—encompassing human cell-based systems (e.g., organoids, Microphysiological Systems [MPS]), in silico computational modeling, and artificial intelligence (AI)— represent a paradigm shift toward faster, more precise, and ethically responsible risk assessments. This document provides a comprehensive guide for pharmaceutical sponsors to incorporate these technologies into their regulatory dossiers. Through early dialogue and establishment of clear technical expectations, the SFDA facilitates the availability of innovative therapies while ensuring robust public health protection.

1.1. Objective

The primary purpose of this guidance is to establish a standardized approach for the integration of NAMs and AI within preclinical drug development continuum. Specifically, this manual aims to:

• Advance the 3Rs Principle: Systematically facilitate the Replacement, Reduction, and Refinement of animal testing in drug development by encouraging the adoption of validated human-relevant technologies.

• Establish Technical Benchmarks: Provide clear criteria for NAM risk assessment report, ensuring they maintain or exceed the level of human health protection provided by traditional animal-based assessments. 

• Standardize Regulatory Review: Streamline the evaluation of complex dossiers by defining transparency and reporting standards for AI-driven and computational models, ensuring data integrity, auditability and reproducibility.

• Enhance Predictive Accuracy: Improve the reliability of preclinical safety and efficacy assessments to mitigate the high attrition rates of drug candidates during subsequent clinical phases.

• Empower Innovative Development: Provide sponsors with a clear pathway to utilize cutting-edge technologies in their submissions aligning with the Kingdom’s strategic goals for healthcare transformation and local clinical research excellence.

1.2. Scope

This document provides guidance for the utilization of NAMs and AI during the preclinical stage of drug development, which is applicable to submissions for Marketing Authorization (MA), designation programs, clinical trials, and post-authorization variations of:

1. Innovative Products or New Chemical Entities (NCEs)

2. Biological and Biotechnological Products

For the purposes of this guidance, NAMs are defined as any non-animal technology, methodology, or computational tool designed to provide human-relevant safety and efficacy outcomes, specifically addressing the translational gap between traditional non-clinical models and clinical reality. This framework assists sponsors in the classification of these methodologies to clearly define their Context of Use (CoU).

1.3. Legal Basis

This guidance is issued Pursuant to Article 5 of the Law of the Saudi Food and Drug Authority issued by Royal Decree No. (M/6) on 25/1/1428 AH, the SFDA shall undertake all regulatory, executive, and control duties necessary to achieve its purposes and objectives, including setting rules and procedures regulating the testing of drugs.

1.4. Related Guidelines and Regulatory Alignment 

Preclinical dossiers incorporating NAMs and AI derived data should be organized to clearly establish the safety basis for the proposed clinical investigation. All data submissions must be fit-for-purpose and scientifically justified within a defined CoU, which should be explicitly stated and consistently applied throughout the dossier.

This document should be read in conjunction with the following SFDA regulatory documents:

• SFDA Research and Investigational Drugs (RAID) Guideline: Overarching principles governing experimental medicinal products in the Kingdom of Saudi Arabia.

• SFDA Guideline on strategies to identify and mitigate risks for first-in-human and early clinical trials with investigational medicinal products: Applicable when NAMs-derived data is used to characterize high-risk molecules prior to clinical entry.

• Regulatory Guidance for Submissions and Applications: For guidance on submission pathways and procedural timelines.

• General Considerations for Preclinical Studies Submission: For providing general guidance on the preclinical data requirements for drug submissions in accordance with the Module 4 of the eCTD 

• Data Requirements for Human Drugs Submission. For guidance on data requirement in accordance with the eCTD structure.

• SFDA Policy of Appeal to Drug Sector Decisions.

1.5. Application and Methodologies

The SFDA recognizes that no single methodology is universally appropriate for all investigational products. Accordingly, this guidance adopts a Weight of Evidence (WoE) framework, under which sponsors are expected to select and integrate the most scientifically appropriate tools based on the product's modality, mechanism of action, and defined CoU. Accepted methodologies include, but are not limited to: 

• Advanced In Vitro & Ex Vivo Systems: Human-derived cellular models, three-dimensional (3D) organoids, and MPS designed to recapitulate complex tissue architecture and multi-organ interactions relevant to the investigational product.

• In Silico & Computational Modeling: Predictive structural models (e.g., Quantitative Structure-Activity Relationship [QSAR] analysis) and kinetic simulations (e.g., Physiologically Based Pharmacokinetic [PBPK] modeling) to characterize pharmacological and toxicological behavior across relevant exposure conditions. 

• AI & Machine Learning (ML) Platforms: Data-driven algorithms applied to pharmacology, absorption-distribution-metabolism-excretion (ADME) characterization, and predictive safety assessment, including the identification of structural alerts and off-target liabilities. 

• Advanced Statistical Frameworks: Quantitative methodologies, including Bayesian inference and hierarchical modeling, to integrate data from heterogeneous sources such as historical controls, rare disease populations, or pediatric extrapolation datasets.

• Next-Generation Risk Assessment (NGRA): A systematic, hypothesis-driven framework that integrates multiple non-animal data streams — including high-throughput bioactivity data, adverse outcome pathways (AOPs), and computational dose-response modeling — to characterize safety and reach risk conclusions without primary reliance on traditional animal models. NGRA submissions must include explicit documentation of the biological plausibility of each data stream and a defined strategy for uncertainty quantification.

1.6. Core Principles for Submission

Sponsors must adhere to the following principles to facilitate an efficient and thorough regulatory review of preclinical dossiers incorporating NAMs and AI -derived data.

• Data Quality and Integrity All in vitro, ex vivo, and computational data must be generated under quality standards commensurate with the intended regulatory use. Where formal Good Laboratory Practice (GLP) certification is not applicable to a given NAM, sponsors must demonstrate adherence to GLP-equivalent principles, ensuring that all data are Attributable, Legible, Contemporaneous, Original, and Accurate (ALCOA+). Deviations from these standards must be documented, justified, and assessed for their potential impact on data interpretability.

• Explicit Cross-Referencing within the Non-Clinical Overview: Sponsors must explicitly identify, within Module 2.4 (Non-Clinical Overview), each instance in which NAMs data replaces, supplements, or contextualizes traditional in vivo data. A formal bridging analysis, providing a direct, structured comparison between NAM outputs and available in vivo findings in the pharmacologically relevant species, is required for all submissions in which NAMs serve a primary evidentiary role.

• Technical Transparency and Auditability: For submissions incorporating AI or ML models, sponsors must provide sufficient technical detail to permit independent SFDA assessment, including: model architecture, software version, training and validation dataset composition, and the end-to-end data preprocessing pipeline. The SFDA reserves the right to request raw datasets and source code to verify model reproducibility and explainability.

• Case-by-Case Regulatory Review: The SFDA evaluates NAMs and AI-derived submissions on a case-by-case basis. The evidentiary weight assigned to any specific methodology is contingent upon the robustness of its validation and its fitness for the defined CoU. The SFDA reserves the authority to require additional in vivo data or targeted experimental verification when a NAM is determined to be insufficiently validated, or when residual uncertainties in the safety characterization cannot otherwise be resolved.

• Protocol and Model Lifecycle Management: To preserve the regulatory validity of a qualified methodology, the protocol, model architecture, and assay conditions established at the time of submission must remain consistent throughout the development program. Changes are classified as follows:

o Minor Modifications: Changes that do not materially affect model performance or assay output must be documented in periodic submission updates, with a brief scientific justification.

o Substantial Modifications: Any significant change to the model architecture, training dataset, or assay protocol constitutes a substantial modification requiring a new technical assessment submission. Sponsors must demonstrate that the integrity and performance of the updated framework are equivalent to or exceed those of the prior version.

Preclinical evidence derived from non-animal sources is categorized according to its intended role in the regulatory decision-making process. Sponsors must clearly identify the applicable classification when defining the CoU for each submitted methodology.

Table 1: Context of Use (CoU) classification

2.1. NAM Technical Assessment Report Criteria

The Technical Assessment Report submitted in support of a NAM must comprehensively characterize the methodology's biological relevance, analytical performance, and fitness for the defined CoU. The report must be organized according to the following sections:

A. Rationale and Context of Use (COU)

• Intended Regulatory Purpose: Clearly define the specific drug development decision the NAM is intended to address (e.g., supporting dose selection, justifying species omission, or characterizing off-target effects).

• Biological and Mechanistic Scientific Basis: Provide a detailed description of the biological pathways or Adverse Outcome Pathways (AOPs) investigates.

• Human Biological Relevance: Justify the selection of cellular systems (e.g., induced pluripotent stem cells [iPSCs], primary human cells, immortalized cell lines) and three-dimensional architecture (e.g., organoids, microvascular scaffolds) with respect to their capacity to recapitulate relevant human anatomical and functional characteristics. The physiological basis for the model's translational relevance must be explicitly stated.

• Predictive Goal: Classify the NAM's intended role as one of the following:

o Replacement: serving as a stand-alone substitute for a traditional study

o Supplemental: addressing a defined data gap

o Integrated: contributing to a WoE determination.

B. Applicability Domain

The Domain of Applicability defines the scope within which a submitted methodology produces reliable and interpretable results. Sponsors must explicitly characterize the domain for each NAM or AI/ML model, addressing both its chemical and biological dimensions:

• Chemical Applicability Domain: Defines the structural and physicochemical space of compounds for which the methodology has been validated. Characterization must reference training dataset composition, molecular descriptor coverage, and structural diversity. Test articles falling outside this space must be identified, and accompanied by an explicit uncertainty statement and a scientific justification regulatory reliance.

• Biological Applicability Domain: Defines the species, cell types, tissue systems, and biological pathways that the methodology is designed and validated to evaluate. Sponsors must identify the specific mechanistic or pathway-based scope of the model, including relevant AOPs, and provide a scientific basis for the human relevance of the biological system.

• Boundary Characterization: Requires equal rigor in defining both what falls within the validated scope and what lies beyond it. Sponsors must provide:

o A structured description of compounds, biological conditions, and exposure scenarios for which the methodology is considered fit-for-purpose.

o An explicit delineation of out-of-scope conditions (e.g. structural classes, species, pathways, or exposure durations) for which predictions are unreliable, absent, or require additional experimental corroboration.

• AI/ML Applicability Domain: For computational models, define the chemical space coverage and demographic representativeness of training and testing datasets. Where the test article approaches or exceeds the boundaries, sponsors must assess the impact on the overall WoE determination and propose mitigation strategies. The SFDA reserves the right to request additional validation data when the Applicability Domain is insufficiently characterized.

C. Technical Characterization and Protocol Standard Operating Procedures (SOPs)

• Assay Protocol: Provide a complete, reproducible protocol specifying test substance preparation, dosing frequency, exposure concentrations, detection instrumentation, primary/secondary endpoints, predefined statistical acceptance criteria (classifying results as positive, negative, or inconclusive), and predefined specifications for test articles.

• System Stability and Working Duration: Define the physiological window over which the model maintains homeostatic and metabolic competence. Demonstrate that critical materials, including cell batches and reagents, remain stable across the study duration.

• Cellular and Platform Metadata: Document cell source (e.g., patient-derived, commercial vendor catalogue number), donor phenotype (age, sex, relevant genetic characteristics), passage number, culture media, and substrate coating conditions.

• Platform Integrity and Test Article Behaviour: For organ-on-a-chip and complex MPS, specify engineering parameters including flow rates, shear stress values, oxygen tension, and scaffold biocompatibility. Sponsors must rule out physicochemical interference with endpoint measurements, including non-specific binding/absorption into platform materials (e.g., polydimethylsiloxane [PDMS] or culture plastics), evaporation at air-liquid interfaces, and material-interface incompatibilities.

• AI/ML Computational Architecture: For AI/ML models, define the specific architecture (e.g., Graph Neural Networks [GNNs], QSARs), software versioning, and the complete end-to-end pre-processing pipeline applied to input data.

D. Reference Standards and Comparative Benchmarking

• Test Substance: Provide a comprehensive characterization including concentration, composition, purity, and dosing intervals. Document compatibility of the test substance with assay materials to confirm absence of interference with endpoint measurements.

• Reference Controls: Define and justify the selection of positive, negative, and vehicle controls that establish the assay's sensitivity, specificity, and dynamic range.

• Translational Bridging: 

o Concurrent/Integrated Classification: Provide head-to-head comparative data between NAM outputs and in vivo responses in the pharmacologically sensitive species.

o Bridge Assessment Classification (Low Species Translation): Provide a systematic literature review, supported by high-impact peer-reviewed evidence, demonstrating the historical failure of traditional animal models to predict human outcomes for the relevant drug class, and establishing that the proposed NAM offers superior human predictivity.

E. AI/ML Model Integrity and Reproducibility

• Architecture Transparency: Fully describe the model architecture, software environment, dependencies, and all preprocessing steps applied prior to training.

• Explainability (XAI): Employ interpretability methods (e.g., SHapley Additive exPlanations [SHAP], Local Interpretable Model-Agnostic Explanations [LIME]) to demonstrate that predictive logic is mechanistically consistent with established toxicological pathways.

• Reproducibility Package: Submit all hyperparameters, random seeds, and data-split methodologies (train/validation/test splits) to enable independent verification of model stability and performance by the SFDA.

• Chemical Space and Demographic Coverage: Characterize training datasets with respect to structural diversity and, where applicable, demographic representation of biological data sources.

F. Methodology control and validation 

• Intra- and Inter-Laboratory Transferability: Provide data evaluating reproducibility across independent laboratories and within the same laboratory, demonstrating robustness against biological variation. Include batch codes, statistical characterization of observed variability based on assay acceptance criteria, and comparisons to reference data.

• Longitudinal Stability: For repeated or extended exposure scenarios, demonstrate that relevant findings can be reliably reproduced across the intended study duration.

• Batch Testing: Submit data demonstrating that test article batches used for NAM validation meet predefined release specifications.

• Predictive Performance Metrics: Report sensitivity, specificity, positive/negative predictive values, and—for quantitative models—Root Mean Square Error (RMSE) across both internal cross-validation procedures and independent external test sets. Demonstrate consistent model performance across structurally diverse compounds within the defined applicability domain.

G. Integrated and Relevance Assessment

• WoE Contribution: Describe how NAM findings shift the overall risk determination for the investigational product when integrated with all available data.

• Data Transparency: Submit complete raw datasets alongside tabulated summaries evaluating methodology reliability and relevance. For AI/ML models, identify and characterize out-of-distribution (OOD) samples and discuss their implications for result interpretation.

H. Fitness-for-Purpose Assessment and System Limitations

• Known System Limitations: Provide a transparent, scientifically grounded discussion of intrinsic model limitations (e.g., absence of immune components, restricted metabolic clearance, lack of systemic circulation) and their impact on result interpretation within the defined CoU.

• Contribution to Risk Characterization: Articulate how methodology findings integrate into overall human health risk characterization, outlining residual uncertainties and proposed mitigation strategies.

• Applicability Domain Limitations: Explicitly flag predictions generated outside the defined range of compounds, structural classes, or biological conditions, accompanying them with formal uncertainty statements.​

The SFDA has established a suite of product designation programs designed to incentivize pharmaceutical innovation, accelerate regulatory review, and facilitate the integration of NAMs and AI into evidentiary packages. Sponsors are encouraged to identify the applicable designation pathway at the earliest feasible stage of development.

3.1. Designation Applications

3.1.1. Research and Investigational Drugs (RAID) Designation 

The Research and Investigational Drugs (RAID) designation, referred to as “Rā’id”, is a specialized regulatory status intended to accelerate the development of innovative investigational products, with particular emphasis on fostering local clinical research within the Kingdom of Saudi Arabia.

NAM Utilization: The SFDA specifically encourages RAID applicants to incorporate qualified NAMs where traditional animal models exhibit limitations in human translatability. A NAM-based evidence submission must:

• Define the CoU with precision, articulating the specific safety or pharmacological endpoint the NAM is intended to address;

• Provide a scientific justification for the superior or equivalent predictive value of the NAM relative to the corresponding in vivo model; and

• Demonstrate consistency with internationally recognized qualification frameworks.

For eligibility criteria, preclinical requirements, regulatory incentives, and application procedures, please refer to the SFDA Research and Investigational Drugs (RAID) Guideline. https://www.sfda.gov.sa/en/guide/19000 

3.1.2. Breakthrough Medicine Designation 

The Breakthrough Medicine Designation is designed to expedite the development and regulatory review of medicinal products intended to treat serious or life-threatening conditions where preliminary clinical evidence demonstrates a substantial improvement over available therapy. This designation reflects a commitment to facilitating patient access to transformative medicines without compromising the scientific rigor of the SFDA’s evidentiary standards.

Evidence Substitution via NAMs: Where species-specificity renders conventional animal studies non-predictive or scientifically inappropriate, e.g., in the case of certain monoclonal antibodies, cell and gene therapies, or highly human-specific biologics, the SFDA may accept qualified NAM-based evidence as a primary component of the preclinical safety and pharmacology package, subject to the following conditions:

• The proposed NAM must be accompanied by a robust biological rationale demonstrating its mechanistic relevance to human physiology;

• NAM-derived data must exhibit a demonstrable and documented correlation with available preliminary clinical observations; and

• Acceptance is subject to scientific evaluation by the SFDA, taking into consideration any relevant prior scientific advice.

For eligibility criteria, regulatory incentives, and application procedures, please refer to the applicable SFDA Breakthrough Medicine Designation Guideline.

https://www.sfda.gov.sa/en/guide/breakthrough-medicine-program 

The SFDA does not grant blanket waivers of animal study requirements under this designation. Each substitution request will be evaluated individually on the basis of scientific merit, the nature of the endpoint in question, and the maturity of the proposed NAM.

3.1.3. Rare Diseases Drugs Designation Programs and Orphan Drugs

In alignment with Orphan Drug Designation (ODD) and the New Accelerator Program for Drugs for Rare Diseases (NADR) the SFDA supports the development of medicinal products intended for the diagnosis, prevention, or treatment of rare diseases affecting a limited patient population within Saudi Arabia. Regulatory incentives are offered through the Orphan Drug Designation (ODD) and the New Accelerator Program for Drugs for Rare Diseases (NADR).

For ultra-rare conditions where conventional preclinical or clinical studies are scientifically or logistically infeasible, the SFDA may accept evidence generated through advanced methodologies, such as Patient-Derived Organoids (PDOs), for confirmatory evidence of efficacy and target engagement in disease-relevant biological systems. 

• Organoid models must be qualified and results contextualized within a WoE framework.

• Submissions incorporating these modalities must be prospectively discussed with the SFDA through the Scientific Advice mechanism prior to formal submission.

3.1.4. Well-established Use (WEU) and Reliance

The SFDA accepts Marketing Authorization Applications based on the Well-Established Use (WEU) pathway for active substances with a documented and stable safety and efficacy profile established through more than 10 years of medicinal use within Stringent Regulatory Authority (SRA).

NAMs as Bridging Evidence: Under the WEU pathway, NAMs may be submitted as bridging data in connection with new formulations or minor modifications of well-established active substances. Where NAM-based evidence is accepted in this context, sponsors may be exempted from conducting additional in vivo animal studies, provided that:

• The NAM is qualified and reproducible under defined experimental conditions.

• The modification does not alter the established pharmacokinetic or toxicological profile of the active substance in a clinically meaningful manner; and

• The bridging rationale is documented within a structured scientific justification submitted at the time of application.

Global Reliance: The SFDA may leverage prior scientific assessments conducted by Stringent Regulatory Authorities (SRAs) where such authorities have previously accepted NAM-based evidence for the same product or active substance. Sponsors wishing to rely on SRA assessments under this provision must submit the relevant assessment reports in their original language, accompanied by a certified English translation where applicable.​

The SFDA recognizes that the integration of NAMs and AI into regulatory submissions requires sustained iterative dialogue between sponsors and the SFDA. To minimize regulatory uncertainty, reduce the risk of data non-acceptability, and support evidence-based innovation, the SFDA has established a suite of structured facilitation mechanisms, as described in this section.

4.1. Scientific Advice and Pre-Designation Meetings 

Sponsors are strongly encouraged to engage with the SFDA through Scientific Advice or Pre-Designation Meetings prior to initiating studies intended to generate pivotal NAM-based evidence. Pre-Designation Meetings are specifically available to sponsors seeking RAID, NADR or Breakthrough Medicine designation. These interactions are non-binding with respect to final regulatory decision-making; however, positions agreed in writing during such meetings will be accorded substantial weight during subsequent substantive review. 

A. Context of Use (CoU) Definition

The sponsor must articulate a precise and scientifically justified CoU for each proposed NAM. The CoU must define:

• The specific regulatory purpose the NAM is intended to serve (e.g., replacement of a specific in vivo toxicology study, supplementation of a clinical pharmacology dataset).

• The scope and boundaries of the NAM’s applicability, including the disease indication, species model being replaced, and the specific endpoint addressed.

• The intrinsic limitations of the NAM and the scientific rationale for why those limitations do not compromise the regulatory adequacy of the submission.

B. Biological Relevance and Cross-Species Correlation

Applicants must provide a documented scientific rationale, referred to herein as a “Scientific Bridge”, demonstrating how NAMs outputs correlate with human physiological responses. This justification is mandatory where the sponsor contends that NAM-derived results offer superior predictive value relative to conventional in vivo data. The Scientific Bridge must be anchored in peer-reviewed mechanistic evidence and, where feasible, supported by biomarker correlation data from early clinical observations.

C. Justification for the Omission of Animal Studies

A robust scientific rationale must be presented for the proposed waiver of specific in vivo studies. This rationale should be:

• Framed within the 3Rs principle (Replacement, Reduction, Refinement), with explicit reference to the applicable guideline or regulatory precedent.

• Supported by evidence that the proposed NAM provides an equivalent or superior level of risk characterization compared to the animal study it replaces; and

• Cross-referenced against the Integrated Approaches to Testing and Assessment (IATA) framework applicable to the endpoint in question.

D. Data Integrity and Computational Governance

For submissions incorporating in silico models or large-scale datasets, sponsors are advised demonstrate that all computational outputs meet the standards for scientific integrity. Specifically, sponsors must address:

• Traceability: All data must be fully auditable from raw acquisition to final model prediction, including documentation of pre-processing steps, parameter selection, and code/model version control.

• Reproducibility: The computational model must yield consistent outputs under defined and documented experimental or parametric conditions.

E. AI Algorithm Validation and Explainability (XAI)

Where AI/ML algorithms are used to generate safety, efficacy, or pharmacological predictions, sponsors must submit a prospective validation methodology addressing:

• Training Data Quality: Training datasets must be representative of the intended population, free from systematic bias, and adequately annotated. The source, composition, and known limitations of the dataset must be disclosed.

• Explainability (XAI): Toxicological or therapeutic predictions generated by the model must be interpretable in the context of established biological pathways. The SFDA considers AOPs as an appropriate mechanistic framework for model anchoring.

F. Integrated Approaches to Testing and Assessment (IATA)

Sponsors proposing the use of multiple non-animal methods in combination must outline how these methods are integrated within a formal IATA framework. The WoE assessment must demonstrate that the aggregate evidentiary conclusion derived from the integrated dataset is of sufficient scientific robustness to support clinical progression or marketing authorization, commensurate with the conclusions that would be drawn from a single pivotal in vivo study.

​G. Benchmarking Standards and Reference Compounds

Sponsors must identify the reference or “gold standard” compounds, including both positive and negative controls, that will be used to characterize the performance, sensitivity, and specificity of the proposed non-animal approach. Benchmarking data must be submitted alongside the NAM validation package to enable SFDA reviewers to independently assess the reliability of the methodology against known human outcomes.​​

All drug submissions must follow the electronic Common Technical Document (eCTD) format. Dossiers relying on NAMs or AI must ensure that the relevant modules—particularly Module 1 and Module 4—are structured to facilitate efficient regulatory review.

5.1. eCTD Module 4: Organization and Structure

Module 1 and Module 4 must contain standardized sections for non-clinical information. NAM and AI-derived data should be integrated into the eCTD structure as outlined in the following table:

Table 2. eCTD structure

The development, validation, and operational deployment of any methodology submitted under this framework must adhere to recognized international quality management systems and data integrity standards. The level of quality oversight applied should be commensurate with the methodology’s classification, intended regulatory purpose, and its defined CoU. Sponsors are responsible for demonstrating compliance with the applicable standards, which are not mutually exclusive and may apply concurrently depending on the nature of the methodology:

• Good Laboratory Practice (GLP): Compliance with GLP or recognized GLP-equivalent standards, for nonclinical safety studies intended to support regulatory submissions.

• Good In Vitro Method Practices (GIVIMP): Adherence to OECD GIVIMP guidelines to ensure the technical reproducibility, reliability, and inter-laboratory transferability of in vitro-derived results.

• Good Machine Learning Practice (GMLP): Governance across the end-to-end lifecycle of AI and machine learning models, encompassing data curation, model training, validation, and post-deployment monitoring

• ISO/IEC Standards: ISO/IEC 17025 and applicable International Organization for Standardization (ISO) standards governing laboratory competence and quality management systems, where relevant

• Good Clinical Practice (GCP) and Good Clinical Laboratory Practice (GCLP): Applicable to methodologies utilizing human-derived biological (such as PDO) or those supporting clinical investigations.

 i : Sponsors are advised that compliance with a single quality standard does not preclude the requirement to meet additional standards applicable to other components of the methodology. The SFDA may request formal evidence of quality compliance during scientific advice sessions or through inquiry waves.

​6.1. AI/ML MODEL TRANSPARENCY STANDARDS

Submissions incorporating AI or machine learning models — including those employed for toxicological prediction, ADME characterization, or biomarker identification — must comply with GMLP and applicable ISO standards governing computational model governance. A technical report documenting the AI/ML model must be included in Module 4 of the eCTD dossier, supported by the Non-clinical Overview in Module 2.4. Sponsors are required to complete and submit the SFDA AI/ML Submission Evaluation Checklist set out in Appendix A.2 alongside the Module 4 technical report. The checklist operationalizes the reporting elements of this section and the mandatory elements specified in Appendix A.1; Any missing or incomplete checklist information may be addressed during the regulatory assessment in accordance with the applicable SFDA submission and assessment procedures.

A. Model Architecture and Development

Sponsors must provide a granular and reproducible description of the algorithmic framework and the end-to-end data preprocessing pipeline, including:

• Specific algorithmic architecture (e.g., Deep Learning, Random Forest, GNNs) and the software versioning.

• Feature engineering and dimensionality reduction techniques, with a scientific rationale for the selection of molecular descriptors or biological inputs.

• Hyperparameters, random seed values, and environment configurations sufficient to enable independent replication of model outputs by SFDA reviewers.

B. Data Provenance and Integrity

The technical report must provide complete documentation of all datasets used for model training, validation, and independent testing, including:

• Source, composition, and known limitations of each dataset, with confirmation that training, validation, and testing sets were strictly independent.

• Assessment of chemical space coverage for molecular models, or demographic and biological diversity for clinical/biological datasets.

• Confirmation that training data was sourced from GLP-certified or equivalently audited repositories, with documented measures to prevent data contamination or fabrication.

• Local Representativeness: Where the intended CoU encompasses the Saudi patient population, sponsors must address dataset representativeness regarding relevant ethnic, genetic, and epidemiological characteristics.

C. Explainability and Biological Interpretability 

Sponsors must provide a robust analysis of the features driving model outputs. Technical documentation must demonstrate that predictive logic is mechanistically consistent with established toxicological pathways or physiological mechanisms, rather than statistical artefacts. Interpretability tools such as SHAP (SHapley Additive exPlanations) or Layer-wise Relevance Propagation (LRP) are encouraged. The SFDA will not accept opaque “black box” predictions as primary regulatory evidence.

D. Robustness and Generalizability

Evidence of model performance under stressed and out-of-distribution conditions must be submitted, including:

• Out-of-Distribution (OOD): Performance evaluation when applied to compounds or data points outside the training domain.

• Sensitivity Analyses: Quantification of the impact of minor input variations (input noise) on prediction stability.

• The Applicability Domain of the model must be explicitly defined and clearly communicated as a boundary condition on the CoU.

E. Lifecycle and Algorithm Change Protocols (ACP)

Sponsors must submit documented protocols governing the full product lifecycle of the AI/ML model, including:

• Version control procedures and quantitative thresholds triggering model retraining (e.g., performance metrics or the availability of a materially new dataset).

• A “Human-in-the-Loop” strategy specifying the role of qualified scientific personnel in verifying safety-relevant predictions prior to regulatory use.

• Confirmation that a change in model architecture constitutes a material modification requiring a new technical assessment submission or formal SFDA notification.

F. Expert Qualification and Declaration (Module 1)

In accordance with SFDA eCTD requirements, the Non-clinical Expert responsible for the Module 1.4 Non-clinical must provide a signed declaration attesting to their qualifications. For dossiers incorporating AI-driven evidence, this declaration must additionally confirm that:

• The AI model’s outputs have been cross-validated against established biological reality.

• The risk of opaque, uninterpretable predictions (“Black Box” risk) has been identified, assessed, and sufficiently mitigated within the proposed CoU.​

The SFDA applies a multi-dimensional approach to evidence integration that moves beyond binary pass/fail assessment toward a holistic, scientifically grounded characterization of benefit–risk. 

• Evidence Integration and Weight of Evidence (WoE): Sponsors are encouraged to employ robust statistical frameworks — including Bayesian inference — to formally integrate historical data (e.g., GLP-compliant historical controls) with current NAM outputs and in vivo results. WoE documentation must transparently demonstrate how integrated evidence supports clinical safety. Isolated NAM results presented without contextualization within a broader evidentiary framework will not be considered sufficient as standalone primary evidence.

• Translational Bridging and Population Extrapolation: Advanced mechanistic modeling is the preferred approach for bridging data across populations, species, or age groups where traditional empirical datasets are unavailable. Integrating adult clinical data with human-relevant NAM modeling — including PBPK and Quantitative Systems Toxicology (QST) framew​orks — is appropriate to support dose selection and safety extrapolation for pediatric populations, rare disease cohorts, and ethnic or genetic subgroups of relevance to the Kingdom of Saudi Arabia.

• Predictive Modeling for Decision Support: ML and algorithmic simulation tools to identify potential risks or therapeutic signals prior to clinical entry. Classified as Decision Support Tools, these models complement, rather than replace, traditional frequentist statistical analyses. Sponsors should demonstrate that outputs provide mechanistically interpretable insights into drug–target interactions and systemic physiological responses, adhering to the explainability standards set out in Section 6.1(c).

The SFDA’s evaluation of NAM- and AI-integrated dossiers follows a risk-based approach. The regulatory outcome of any assessment is determined by the methodology’s validation status, demonstrated biological relevance, and the coherence of its CoU. The procedural pathways described in this section apply to all submissions incorporating NAM or AI-derived evidence.

8.1. Inquiry Waves / Interim Deficiency Notice

Where submitted NAM data is deemed insufficient to support a definitive safety or efficacy conclusion, the SFDA may issue an inquiry in accordance with the applicable inquiry waves and procedures outlined in the Regulatory Guidance for Submissions and Applications. Such inquiries constitute interim procedural actions and do not represent a negative regulatory opinion on the application. 

Scientific Rationale: The SFDA will specify, with particularity, the identified gaps in the CoU definition, the biological relevance of the proposed model, or the adequacy of the validation dataset. Generic deficiency notices will not be issued; each inquiry wave will reference the specific technical requirements with which compliance is required.

• Sponsor Action: The sponsor may be required to provide supplemental data, including cross-validation studies using traditional in vivo models, expanded sensitivity analyses for computational outputs, or additional biological plausibility evidence. The inquiry wave response timeline and format will be specified in the notice.

8.2. Non-Acceptance of Methodology (Rejection of NAM Evidence)

The SFDA may reject the inclusion of a specific NAM as a primary line of evidence where it fails to meet the technical and scientific benchmarks defined in this guidance. A determination of non-acceptance will be communicated in writing, with specific reference to the deficiency identified.

• Criteria for Non-Acceptance: Non-acceptance may be determined on the basis of insufficient model validation, inadequate explainability of AI-derived outputs, failure to establish a credible human-biological bridge, or a material deficiency in data integrity or provenance documentation.

• Impact on the Application: Non-acceptance of a specific NAM methodology does not, in itself, result in rejection of the Marketing Authorization Application (MAA). Provided the sponsor is able to substantiate the relevant safety or efficacy claim through alternative established methodologies, the remainder of the application may continue to be evaluated without prejudice.

 i.

Sponsors are advised to engage with the SFDA through Scientific Advice prior to submitting NAM-based primary evidence. Non-acceptance at the time of MAA review is more likely where the methodology has not been prospectively discussed and agreed with the SFDA.

8.3. Safety Disclosure and Data Confidentiality

 Safety Disclosure: Where safety-relevant findings are identified through NAMs or AI models, sponsors shall report such findings to the SFDA in accordance with the applicable regulatory safety reporting requirements. Such findings should be appropriately documented and considered in the overall safety assessment.

Data Confidentiality: All proprietary information submitted in connection with an application — including AI model architectures, training datasets, and proprietary in vitro protocols — shall be handled in accordance with the applicable SFDA confidentiality and data protection requirements. 

8.4. Technical Re-Assessment and Appeal

Sponsors retain the right to appeal a negative regulatory decision, including where the decision of rejection the registration was based, in whole or in part, on the interpretation of NAM- or AI-derived evidence. The appeal process is illustrated in the SFDA Policy of Appeal to Drug Sector Decisions.

8.5. Post-Decision Communication

Following the issuance of a final decision, the SFDA maintains structured channels for post-action engagement:

• Post-Action Technical Briefing: Sponsors may request a post-action technical briefing to review the specific scientific bases for the Authority’s determination and to identify lessons learned for future submissions. This session is a forward-looking scientific alignment exercise and does not constitute a mechanism for reopening or challenging the final decision. 

• Public Transparency Reporting: In the interest of advancing the scientific community’s understanding of regulatory standards for emerging technologies, the SFDA may publish a Summary of Non-Acceptance documenting the technical basis upon which a specific methodology was determined not to be fit for regulatory purpose. All such publications will be carefully reviewed to ensure that proprietary information, including model architecture details and training datasets, is appropriately redacted prior to public release.​

Appendix A: AI/ML Model Reporting and Submission Evaluation Standards

This appendix sets out the minimum reporting requirements and the corresponding SFDA evaluation checklist for AI/ML-based evidence submitted in Module 4 of an eCTD dossier. Compliance with these requirements is a prerequisite for substantive review of AI-derived evidence.

A.1 Mandatory Reporting Elements (Module 4 Technical Report)

1.  Regulatory Objective and Context of Use: A precise statement of the regulatory question the model is intended to answer (e.g., replacing the in vivo Ames genotoxicity assay; predicting drug-induced liver injury risk), with a defined and bounded CoU.

2.  Dataset Summary: The number and source of compounds or data points, an assessment of chemical diversity or biological representativeness, and the specific endpoint for which the data were sourced.

3.  Model Architecture and Version: The specific algorithmic architecture (e.g., Graph Neural Network, QSAR, Random Forest), the software library and version, and all hyperparameter and seed values required for replication.

4.  Performance Metrics: Accuracy, sensitivity, specificity, and root mean square error (RMSE) for regression models, reported separately for internal cross-validation and independent external test sets.

5.  Validation Methodology: A description of the internal validation approach (e.g., k-fold cross-validation) and the composition and source of the independent external test set.

6.  Uncertainty Quantification: The defined Applicability Domain of the model and prediction confidence intervals for individual outputs.

7.  Versioning and Lifecycle Plan: Defined criteria and technical thresholds for future model updates, retraining events, and the regulatory notification obligations triggered by material architectural changes.

A.2 SFDA AI/ML Submission Evaluation Checklist

The following abbreviations and defined terms are used throughout this document. Where a definition is provided, it shall apply for the purposes of interpreting this guidance unless the context expressly requires otherwise.