Clinical In Silico Studies in Drug Development and Therapy: Simplified Guidance for Authors, Reviewers and Editors

Clinical in silico studies are used in drug development and to assess therapies.1 They include pharmacokinetic and pharmacodynamic analyses, dose and schedule selection, disease-progression models, evidence synthesis, virtual populations and clinical-trial simulations. By combining evidence from several sources, they can compare treatment strategies under a common protocol and identify questions for future studies. However, they cannot create information that is absent from the source evidence or replace well-designed clinical trials.2–6

For editors and reviewers, the central question is simple: do the evidence, model and validation support the clinical claim? Common problems include a poorly defined question, selective or untraceable inputs, calibration described as validation, very large virtual cohorts treated as independent patients, incomplete analysis of uncertainty, and unsupported claims of superiority, equivalence or safety. These problems may be missed if the clinical and computational parts are reviewed separately.

This editorial gives practical guidance for authors and reviewers of clinical in silico studies. It follows the risk-based principles of the final ICH M15 guideline and complements the journal’s guidance on statistical analysis and computational methods in drug design and discovery.2,7,8 The evidence required should reflect the question, the intended use of the model, its influence and the consequences of a wrong conclusion. Authors should include a dedicated In Silico Methods section and a versioned model analysis plan or equivalent protocol covering the requirements below.

These standards should be applied in proportion to the intended use. They are not a rigid checklist, and an exploratory study should not be excluded simply because independent external data, patient-level data or openly available code are unavailable. The manuscript must still explain how the model was built, which evidence it used, how its performance was tested and which conclusions it can support. Requirements for advance planning, independent validation, uncertainty analysis and reproducibility should become stricter as the model has greater influence or an incorrect conclusion could have more serious clinical consequences.

Define the Clinical Question and Intended Use Before Modelling

State the target population, intervention, comparator, outcome, time frame and treatment or monitoring strategy. For a causal or counterfactual comparison, an estimand-like description should state exactly what is being compared and how events after treatment begins are handled.2,9 Classify the model as exploratory, decision-supporting or confirmatory. An exploratory model may need less independent evidence than one used to claim comparative efficacy, safety, non-inferiority or equivalence.

Describe the Evidence Base

Each empirical input should be traceable to its source, population, regimen, time point, outcome definition, estimate and uncertainty. Reports from the same trial should be linked and not counted as independent evidence. State which evidence was used for model development, parameter estimation, calibration, validation, plausibility checks and sensitivity analyses. If one source serves more than one purpose, declare each role and explain the resulting dependence. Evidence used to fit or calibrate the model is not independent validation. Where possible, test the effect of excluding or down-weighting sources with overlapping roles. Cross-trial comparisons should describe important differences in eligibility, baseline risk, regimens, outcome timing, rescue or retreatment rules and measurement definitions. Authors should normally provide a concise evidence-source and use table, in the main manuscript or supplementary material, with one row for each trial or major data source. Table 1 lists the minimum fields. This makes overlapping evidence and repeated use of the same source easier to identify without requiring a separate table for every model parameter.

Table 1.

Minimum Fields for an Evidence-Source and Use Table

Field	Minimum Information to Report for Each Source
Source identification	Full citation or dataset identifier; link all reports arising from the same trial or dataset.
Clinical context	Population and eligibility criteria, intervention, comparator, regimen, sample size and assessment time.
Information used	Outcome definition, estimate, uncertainty measure and any reconstruction or transformation.
Role in the model	Development, parameter estimation, calibration, validation, external plausibility assessment and/or sensitivity analysis.
Independence and overlap	Whether participants, datasets or reports overlap, and whether the source serves more than one model role.
Main limitations	Risk of bias, missing information, indirectness, dependence and other limitations that could affect the model.

Plan the Analysis in Advance Where Possible

A versioned model analysis plan or equivalent protocol should usually be prepared before the final comparison. It should state the outcomes, evidence hierarchy, model structure, assumptions, acceptance criteria, uncertainty and sensitivity analyses, and decision rules. Authors should state whether the plan was prepared before examining the results and provide a dated or versioned copy where available. Decisions made after seeing the results should be identified, and material changes should be recorded. Exploratory work begun without a formal plan can still be considered. Prospective registration is not required for publication, but undisclosed outcome selection or model changes after the results are seen are unacceptable.

Distinguish Verification, Calibration, Validation and Applicability

Verification checks that the equations, algorithms, data transformations and code were implemented correctly. Calibration estimates unknown parameters by matching model outputs to selected data. Validation tests performance against data not used to build or calibrate the relevant model component. Applicability asks whether those data match the intended population, intervention, outcome and conditions.2,4–6,10 A good fit to calibration data is not external validation, and cross-validation is an internal method. Independent external validation is preferred and should normally be expected for confirmatory, evidence-replacing or high-consequence claims. It may not be possible for exploratory or method-development studies. In those cases, authors should combine verification with internal validation, held-out or leave-one-source-out tests where possible, sensitivity analyses, benchmark reproduction and checks of external clinical plausibility. Methods such as posterior predictive checks should be described accurately as internal evidence. Conclusions must remain within the context supported by the evidence.

Address Important Sources of Uncertainty

Consider uncertainty in parameters, variation between patients and studies, uncertainty about the model structure and its transfer to the target population, and numerical or Monte Carlo error. Uncertainty introduced by reconstructed standard errors, digitised figures, assumed correlations, transformed outcomes or evidence weighting should be carried through the model or tested in sensitivity analyses. These analyses should examine reasonable alternatives for evidence sources, priors, model forms, correlation structures, protocol thresholds and adherence assumptions. If reasonable choices reverse the direction of a result, report the result as uncertain rather than select the preferred specification.4–6,11 Focus on uncertainty that could change the interpretation or decision; authors need not model every theoretical source. Important omissions should be explained, and the depth of sensitivity analysis should match the intended claim.

Do Not Treat Monte Carlo Precision as Clinical Evidence

A simulation may generate thousands or millions of virtual patients, but this does not add clinical information. Virtual cohort size controls computational precision; it is not a clinical sample size. Simulation probabilities, uncertainty intervals, predictive distributions and Monte Carlo error estimates may be reported if their meaning is clear. Treating virtual patients as independently observed participants in hypothesis tests is pseudoreplication and is not acceptable. It is like treating repeated measurements from one experimental unit as separate subjects, a practice rejected in the journal’s statistical guidance.7 P values are appropriate only when their inferential basis and effective sample size come from the observed evidence, not from an arbitrarily enlarged virtual cohort.

Make Independent Appraisal Possible

Provide the model analysis plan, source registry, equations or algorithm description, parameter distributions and correlations, acceptance criteria, verification results, validation or plausibility assessments, sensitivity analyses and machine-readable main outputs. No particular software package is required: established platforms and bespoke code are acceptable when their choice is justified and their implementation can be checked. The final analysis code should be available to reviewers when ethical, legal and commercial restrictions permit. It should identify the required inputs, parameter settings, exact software and package versions, and random seeds, and should reproduce the main tables, figures and results. Public deposition at publication is encouraged. If code or data cannot be shared, authors should explain why and provide another way to assess the model, such as equations, pseudocode, parameter tables, validation outputs, worked examples or confidential reviewer access. Restricted access alone should not prevent publication if the model remains sufficiently auditable. Use reporting guidance suited to the model: STRESS for empirical simulation and TRIPOD+AI when a clinical prediction model is the main contribution.12,13 A completed checklist does not by itself establish validity.

Keep Conclusions within the Validated Context

The title, abstract when required, and conclusions should identify the work as in silico and state whether it uses aggregate data, patient-level data, a mechanistic model or a hybrid approach. An aggregate-data simulation cannot reproduce head-to-head randomisation. Standardising the protocol may address known regimen differences, but it cannot remove unmeasured differences between studies or ecological bias. Failure to detect a difference does not prove equivalence. Use terms such as superiority, non-inferiority, equivalence, safety and clinical effectiveness only when the design and validation support them. Do not estimate rare adverse events from efficacy-focused virtual cohorts without a validated safety model and suitable event data.

Report Model-Based Results as Model-Based

Authors may report the direction, size and uncertainty of simulated treatment differences, but should not describe them as observed treatment effects. Exploratory comparisons can generate hypotheses and help to design future trials. They should not become claims of superiority, equivalence, non-inferiority or safety unless the design and validation support those claims.

Apply Proportionate Editorial Triage

A manuscript may enter peer review when the clinical question and intended use are clear, the evidence sources and their roles are described, the model can be assessed, and the conclusions match the validation performed. Reporting gaps that can be corrected should usually lead to clarification or revision rather than immediate rejection. Rejection is appropriate when key inputs are untraceable or selectively chosen, the model cannot be meaningfully assessed, calibration is presented as independent validation, important uncertainty is ignored, virtual cohort size is used to manufacture statistical significance, or uncorrectable efficacy or safety claims exceed the evidence. The reviewer and editorial team should include clinical and methodological expertise suited to the model and intended claim.

Clinical in silico studies use existing evidence to address questions that conventional comparisons cannot answer. Their credibility depends on the quality and origin of the evidence, a clear intended use, separation of calibration from validation, adequate uncertainty analysis and restrained conclusions—not on the number of virtual patients. An exploratory model can test assumptions and show what evidence a future trial should collect. Exploratory work is not, by definition, low-quality work.1

Computational modelling does not replace clinical evidence, but it is not inherently less scientific than other methods. It has a legitimate role in drug development and therapy assessment when its assumptions, limits and influence are clear. Clinical, pharmacological, statistical and modelling expertise should be combined, with specialist advice when a model may affect patient care or a drug-development decision. By setting clear but proportionate standards, Drug Design, Development and Therapy can provide an appropriate home for these studies and help define the field through well-reported applications.

Funding Statement

No funding was received for this work.

Disclosure

The authors report no conflicts of interest in this work.