AI can propose a molecule, rank a site, flag a safety case or estimate a treatment effect. Those outputs are not interchangeable. A discovery hypothesis may be inexpensive to reject in a laboratory; an eligibility or dosing error can affect a participant directly; a model used in a regulatory submission can alter the evidence on which benefit and risk are judged.
The governing concept is context of use: the specific question, data, user, decision and consequence for which the model is intended. “AI for drug development” is too broad to validate. Every model needs a bounded claim and an evidence package proportionate to the influence it has.
This guide reflects UK rules and regulator material available on 31 July 2026. The amended UK Clinical Trials Regulations took full effect on 28 April 2026. Requirements differ for old-rules and new-rules trials, and multinational programmes must also map EU, US and other jurisdictions rather than treating one assessment as globally sufficient.
Classify the Model by Its Decision
Start the inventory before discussing architecture.
| Context | Example output | Decision influenced | Main evidence risk |
|---|---|---|---|
| Target discovery | Ranked target or pathway | Which experiments receive resources | Circular use of literature or assay data |
| Molecule design | Candidate structure or property estimate | Which compounds are synthesised | Invalid domain or unreliable uncertainty |
| Non-clinical safety | Toxicity or pharmacokinetic prediction | Whether and how a candidate progresses | Missing mechanism or population coverage |
| Trial feasibility | Site or recruitment forecast | Site selection and capacity | Historic access bias treated as ability |
| Eligibility support | Extracted criterion match | Who receives manual screening | Omitted exclusion or stale source record |
| Endpoint derivation | Feature or score from trial data | Primary or secondary analysis | Uncontrolled measurement or algorithm change |
| Safety surveillance | Prioritised case or signal | Review order and escalation | Rare serious cases buried by ranking |
| Regulatory evidence | Model-derived estimate or dataset | Benefit-risk assessment | Poor traceability and unverifiable analysis |
Write a single context-of-use statement that includes the model’s prohibited uses. For example: “This tool extracts possible eligibility facts from source documents for trained-site review; it does not determine eligibility, alter protocol criteria or contact a potential participant.”
The adopted EMA reflection paper on AI in the medicinal-product lifecycle is an EU document rather than UK law, but it is a useful lifecycle reference for multinational developers. UK programmes should seek product-specific advice where an AI method may become material to regulatory decisions.
Build a Model Evidence Dossier
The dossier should let an independent reviewer reconstruct why the model was considered fit for its context. A compact structure is:
- intended context and decision owner;
- data provenance, permissions and representativeness;
- target, labels and reference standard;
- training, tuning and independent evaluation split;
- leakage, duplication and missingness analysis;
- metrics tied to decision consequences;
- subgroup, site and temporal performance;
- uncertainty, abstention and human review;
- software, model and environment configuration;
- cybersecurity and access control;
- change, retraining and rollback policy; and
- deviations, incidents and post-deployment monitoring.
Do not reduce the dossier to a model card supplied by the vendor. Connect claims to protocols, analysis plans, validation reports, data-management records and quality approvals. Our AI assurance evidence-pack guide gives a reusable index for that chain.
The MHRA’s June 2026 medicines AI sandbox announcement focuses on generating evidence for models that predict safety and pharmacokinetics. A sandbox is a controlled learning route, not an exemption or market approval. Organisations should not describe participation or thematic alignment as regulatory endorsement.
The MHRA offers scientific advice at different stages of medicine development, including pivotal-trial design. Advice is based on the questions and evidence submitted and is not legally binding on a future application, so record the assumptions and later changes.
Preserve GxP Data Integrity Through the Pipeline
An accurate prediction built from irreproducible or unauthorised data is not reliable evidence. Map the full data lifecycle: acquisition, transfer, transformation, labelling, analysis, output, review, retention and destruction.
MHRA GxP data-integrity guidance applies a risk-based approach across laboratory, clinical, manufacturing, distribution and pharmacovigilance contexts. Its core distinction is important for AI: controls that preserve integrity do not automatically prove data quality or model validity.
For every material dataset retain:
- origin, collection purpose and governing protocol;
- participant, sample, assay, instrument and site identifiers as applicable;
- source and processing timestamps;
- transformation code and parameter version;
- inclusion, exclusion and imputation rules;
- label source, reviewer and adjudication;
- access, export and correction history;
- training, validation and test membership;
- derived-feature definitions; and
- retention and deletion requirements.
Use immutable or controlled raw zones and reproducible transformation pipelines. Never overwrite a corrected label without preserving the original and reason. Generated synthetic data should be marked as synthetic through every derived dataset; it should not quietly re-enter an evaluation set as though independently observed.
Protect the independent test set from repeated tuning. If the team changes the model after seeing test performance, create a new truly independent evaluation or explain the resulting limitation. Check patient, sample and site duplication across splits, including aliases and longitudinal records.
Validate Discovery Models With Experimental Reality
Retrospective benchmark performance is a screen, not a discovery outcome. Define the model’s domain: chemical space, assay type, species or system, endpoint, measurement conditions and applicability limits.
A prospective evaluation can freeze the model and selection rule, nominate candidates and comparators, then test them under a pre-specified laboratory protocol. Record every proposed candidate, including failures. Otherwise, selective publication can turn a broad sequence of unsuccessful suggestions into one impressive anecdote.
Evaluate:
- enrichment against an appropriate baseline;
- assay reproducibility and confirmation;
- novelty without invalid chemistry;
- uncertainty calibration;
- performance outside common scaffolds or classes;
- sensitivity to preprocessing and assay batch;
- cost and time per confirmed useful result; and
- the rate at which the model abstains or leaves the domain.
Do not claim that a model “shortened development” from the time required to generate candidates alone. The meaningful pathway includes synthesis, assay, replication, safety, manufacturing, clinical evidence and regulatory review.
Use Trial AI as Controlled Trial Infrastructure
For an AI system used in a clinical trial, identify whether it is simply operational support, part of trial conduct, a measurement method, a medical device or part of the evidence submitted. Those statuses can overlap and change as intended use changes.
The amended regulations came fully into effect on 28 April 2026. MHRA transition guidance, updated on 15 July, distinguishes applications submitted before that date from new-rules trials submitted on or after it. Sponsors must apply the correct transitional provisions rather than rewriting an ongoing trial informally.
ICH E6(R3) GCP was implemented in the UK alongside the reform. MHRA implementation guidance expects sponsors to document the quality-system assessment needed for compliance and to conduct a documented impact assessment for ongoing trials. The UK-specific E6(R3) annotations clarify how the guideline applies in the UK.
Put the AI function into protocol, vendor, data, monitoring and change-control arrangements according to its criticality. If it derives an endpoint, pre-specify the algorithm, input quality, missing-data handling and analysis. If an algorithm must change during the trial, assess the protocol, comparability, validation and approval consequences before deployment.
Protect Recruitment and Eligibility Decisions
Site-selection models can learn where past recruitment was easiest rather than where eligible participants live or can access research. A high predicted recruitment score may encode historic exclusion, transport barriers or differences in referral.
Review site ranking against:
- epidemiology and protocol eligibility;
- local standard of care and competing studies;
- staffing, equipment and pharmacy capability;
- prior recruitment denominator, not only total recruits;
- participant burden and travel;
- underserved populations and accessibility;
- data completeness and recency; and
- manual justification for inclusion or exclusion.
Eligibility extraction should quote the source fact and date, not only produce “eligible”. A trained reviewer confirms each criterion against the current protocol and source record. Ambiguous, missing or conflicting information must remain unresolved. No outreach occurs merely because a model inferred a diagnosis or characteristic.
Health and genetic information is special-category personal data. Use a documented lawful basis and Article 9 condition, role-based access, minimisation and appropriate research safeguards. The general control pattern in automated decisions and human review applies, but consent and protocol-specific research governance still require specialist review.
Keep Endpoints Stable and Auditable
AI-derived imaging, speech, wearable or digital biomarkers can become critical measurements. Validate the measurement, not only the classifier.
Specify device and acquisition procedure, quality thresholds, preprocessing, model version, scoring range, repeatability, inter-reader or reference comparison, site transfer and missing-data handling. Test whether the score changes with equipment, language, lighting, movement, disease severity or other plausible conditions.
Separate endpoint computation from treatment information where blinding requires it. Log every run and prevent silent recomputation after model updates. Where historical and updated algorithms coexist, define whether bridging is possible and preserve both outputs.
An endpoint model should have an analysis-ready audit record containing input identity, hash or controlled reference, software environment, output, quality flags, timestamp and reviewer action. Generated narrative explanations are not a substitute for the specified endpoint definition.
Prioritise Safety Without Filtering It Away
AI can help deduplicate, code or prioritise adverse-event cases, but high throughput does not justify hiding low-scored reports. Seriousness, expectedness, causality and reporting obligations remain governed by the applicable pharmacovigilance process.
Use conservative rules to protect known urgent categories. Run the model in parallel with current intake before allowing it to affect queues. Measure missed serious cases, time to first review, duplicate handling, coding disagreement, model abstention and performance by product, source, language and population.
Every case remains retrievable, and the original report remains intact. Reviewer corrections should improve controlled dictionaries or future evaluation rather than mutate historical evidence invisibly. Establish a direct incident path when the model or integration may have delayed safety action.
Make Transparency Part of the Trial Record
The 2026 reforms added legal transparency requirements for relevant UK CTIMPs. HRA registration guidance explains that new-rules trials must be registered before the first participant is recruited or within the specified period after approval, whichever is sooner. HRA also requires publication of a results summary for applicable trials within the relevant timeframe (publishing trial results).
Do not let an AI workflow create a private parallel account of the study. Register material methods and endpoints consistently with the protocol and registry. Explain model-derived outcomes in results and lay summaries without implying that a prediction is observed fact. Preserve negative and inconclusive findings.
Release Gates Across the Lifecycle
| Gate | Required evidence |
|---|---|
| Context | One decision, named owner, regulatory role and prohibited uses |
| Data | Provenance, permissions, integrity, representativeness and independent split |
| Technical | Consequence-based metrics, calibration, subgroup results and uncertainty |
| Process | Protocol or procedure integration, trained review, audit trail and fallback |
| Regulatory | Correct trial regime, GCP impact, advice and approvals documented |
| Live | Incident, drift and workload monitoring; locked version; tested rollback |
| Scale | Benefits reproduced prospectively without breaching safety, integrity or inclusion floors |
Require full traceability for every model-derived result used in a regulated decision. Set zero tolerance for silent source-data overwrite, unapproved endpoint changes and automatically excluded safety cases. Require every eligibility recommendation and material safety signal to receive the defined human review.
Pause when data provenance is incomplete, the model leaves its validated domain, a site or subgroup falls below its agreed floor, the review queue exceeds capacity or a supplier update cannot be reconstructed. Efficiency never compensates for an invalid evidence chain.
Pharma AI is valuable when it makes a scientific question more testable and the resulting evidence more reliable. It becomes dangerous when a plausible output outruns experimental confirmation, participant protection or the regulated record.
