The engineering thesis
The starting point for reverse engineering human biology is a practical question: what can the evidence available for one individual tell us about how that person's biology is functioning and how it may change? Genomic information describes potential influences on biological processes. Laboratory measurements and clinical observations describe selected aspects of the person's condition. Treatment history and repeated measurements add the temporal relationships needed to investigate why that condition persists, improves, or deteriorates. The proposed Nexus framework brings these sources into an explicit model whose explanations and forecasts can be examined against evidence.
Within this paper, reverse engineering means working backward from observations to candidate mechanisms, dependencies, and constraints. The task is to reconstruct enough of the relevant biology to answer a defined research question. This reconstruction is necessarily incomplete, but incompleteness need not make it uninformative. A model may be useful if it identifies which mechanisms are supported, preserves plausible alternatives, and predicts a specific outcome with uncertainty that is appropriate to the evidence. Its resolution should be determined by the question and available measurements, rather than by an ambition to represent every process in the body.
The phrase biological assembly language captures the engineering intuition behind this approach. Genetic variation can be examined for its relationship to molecular function, and molecular function can be connected to physiological observations through evidence about biological mechanisms. The analogy becomes misleading if it implies that the genome is a complete executable specification of the individual. Genotype interacts with regulation, development, environmental exposures, and other contextual factors. Consequently, the same genetic difference may have different implications in different circumstances, and sequence information alone cannot establish the patient's present biological state [3].
For Nexus, this distinction establishes the relationship between BioMatrix and the proposed Disposition Trajectory Engine. BioMatrix provides the analytical foundation for connecting genetic findings with laboratory biology and clinical context. The trajectory engine extends that foundation by introducing explicit representations of time, state transitions, future outcomes, and intervention scenarios. The proposed extension depends on the quality of the underlying integration, but its predictive performance must be established separately. A convincing explanation of current findings does not demonstrate an ability to forecast what will happen next.
The research therefore begins with an evidence chain. A genomic finding may support a possible molecular effect; that effect may be relevant to a physiological process; and an available measurement may provide information about that process. The clinical question determines whether the chain is useful. Each relationship needs an identifiable evidentiary basis and an assessment of its applicability to the individual. A laboratory abnormality accompanying a genetic finding does not, by itself, establish that the variant caused the abnormality. Routine laboratory results also cannot be treated as direct measurements of gene expression or intracellular pathway activity unless the assay supports that interpretation.
Repeated observations make this chain more informative by revealing persistence, direction of change, and timing relative to exposures or treatment. They can strengthen one explanation and weaken another while leaving causality unresolved. The proposed framework treats those differences as research information. It preserves competing mechanisms when the evidence cannot distinguish them and identifies further observations that could resolve the uncertainty. This approach is consistent with the focused development of cancer digital twins described by the National Cancer Institute, in which a specific patient-centered question guides the construction and integration of models [9].
Biological correction gives the reconstruction a measurable purpose. Here, correction means improvement in a specified dysfunctional process or clinically relevant outcome, assessed using an appropriate measurement over a defined interval. The target may concern physiological function, symptoms, disease control, recovery, or treatment tolerability. It must also account for adverse effects and burden. An improvement in a BioMatrix score can support interpretation, but the relationship between that score and a clinical benefit requires its own evidence. A movement toward a preferred laboratory range cannot independently demonstrate reduced disease risk or an overall treatment benefit.
The scope consequently includes two related research settings. In an individual without the disease being studied, the framework investigates susceptibility and the development of a defined outcome. In a patient with established disease, it examines the observed course and the applicability of treatment-related evidence. Oncology requires particular attention to the distinction between the patient's constitutional background and the acquired characteristics of the cancer [4]. These settings can share architecture, but they require different populations, endpoints, and validation procedures. Reconstructive correction refers to rebuilding an analytical understanding that supports investigation; it does not presume genome editing, tissue regeneration, or universal disease reversibility.
The scientific argument is testable through a small set of substantive comparisons. Genomic information should add value beyond clinical and laboratory inputs if it is to justify its role in a given forecast. Repeated measurements should improve on baseline-only predictions where temporal information matters. Mechanistic structure should contribute predictive accuracy, robustness, or an independently assessable explanation. Foundation-model assistance should improve evidence processing without increasing unsupported claims. Treatment-response models should identify meaningful departures from expectation early enough to support review. Failure in any comparison is informative and may justify simplifying the system. Together, these hypotheses connect the broad engineering ambition to an achievable program of empirical research.
Reconstructing the individual biological state
The patient state is the common representation through which genomic interpretation, current biology, and longitudinal prediction become connected. It describes the individual at a defined analytical cutoff and records how that description was assembled. A measurement has meaning within the circumstances of its collection, so the state must preserve specimen origin, timing, method, quality, and clinical context alongside the result itself. This establishes whether observations are comparable and whether they can support the proposed forecast. It also prevents the apparent completeness of a dataset from concealing uncertainty about what was actually measured.
Genomic evidence in its biological context
The Nexus genetic kit and whole genome sequencing enter this representation as complementary sources of evidence. Their interpretation begins with the documentation accompanying the actual assay and dataset. Kit findings are limited to the loci and variant classes supported by that assay. WGS permits broader interrogation, including coding and noncoding variation, but its analytical sensitivity varies across genomic regions and variant classes. The relevant capabilities belong to the complete laboratory and bioinformatics workflow, not to the label WGS alone [2]. The engine must retain these limits when deciding what a positive, negative, or unresolved finding means.
A normalized genomic record preserves the reference assembly, position, reference and alternate alleles, genotype, specimen identity, and available quality information. Where interpretation depends on a transcript, its identifier and version remain attached to the finding. Directly measured genotypes and computationally inferred genotypes remain distinguishable. This is particularly important when combining data from different assays: agreement is a useful consistency check, while disagreement requires examination of identity, allele representation, assembly, and analytical quality. The reconciliation should resolve the underlying evidence rather than select whichever result best fits a preferred explanation.
Negative evidence requires equal care. A position absent from a variant-only file has not necessarily been established as a reference genotype. Coverage and callable-region information are needed to distinguish a supported negative result from a position that was unassessed or unreliable. The state therefore preserves the difference between an available result and an assumed result. If an analysis requires a finding that the assay cannot reliably establish, that limitation follows the record into downstream interpretation and prediction rather than disappearing during normalization.
The biological origin of a finding also affects its meaning. Constitutional variation, confirmed inherited variation, acquired variation, and unresolved origin should be represented separately. Family inheritance cannot be inferred merely because a variant is constitutional. For predisposition research, pathogenicity classifications must retain their disease and inheritance context, while association-based risk estimates retain the populations and outcomes from which they were derived. A common risk-associated allele is not equivalent to a pathogenic variant causing a Mendelian disorder. A variant of uncertain significance remains uncertain and cannot, on its own, justify clinical intervention [1].
Cancer introduces a second genomic context within the same individual. Tumor-associated findings describe sampled disease material at a particular time, while constitutional findings describe a different aspect of the patient. Matched normal testing, where appropriate and available, can help distinguish their origins. Specimen location, collection date, testing method, and available tumor-content information are therefore essential parts of the record. Molecular differences within a cancer and changes during treatment limit how well an earlier sample represents later disease [5]. The state must preserve that temporal relevance instead of treating every molecular result as a permanent property of the patient.
Longitudinal reconstruction and evidentiary resolution
Laboratory biology provides the observations against which genomic hypotheses are examined. The original value, units, reference interval, laboratory, specimen type, and collection time remain available even when normalized values are added for computation. Clinical context may include developmental stage, fasting status, acute illness, medications, supplements, and proximity to treatment. Laboratory reference intervals and Nexus analytical thresholds have separate roles and should retain their respective provenance and intended use. This prevents an analytical preference from being mistaken for a laboratory-defined boundary or a validated clinical endpoint.
Clinical history gives these observations their sequence. Symptoms, diagnoses, procedures, interventions, interruptions, and reported responses should be reconstructed as dated events, with uncertainty retained when dates or administration are unclear. A prescription does not establish that an intervention was received. Equally, the time an event occurred and the time its documentation became available are different facts. Both are needed to reconstruct what a historical forecast could legitimately have used. A later report describing an earlier event cannot silently enter a prediction that was supposedly issued before the report existed.
Trends can then be calculated from comparable observations over their actual elapsed intervals. A change in assay, specimen, or collection conditions may interrupt comparability and should be examined before measurements are interpreted as one continuous trajectory. Treatment context belongs alongside the trend because a numerical change may occur during an intervention, after its completion, or during an unrelated illness. The derived result should identify the contributing measurements and transformation, allowing a reviewer to understand how the temporal interpretation was reached.
BioMatrix connects this longitudinal record to genomic evidence by organizing findings around relevant biological processes. Convergence occurs when different evidence streams support a compatible interpretation, provided their contributions are genuinely additional. Correlated laboratory findings or repeated annotations must not be counted as independent confirmations. A mismatch identifies a discrepancy between an evidence-based expectation and the observed state. It may point toward an incomplete mechanism, unmeasured exposure, treatment effect, timing difference, or data problem. Root-driver analysis uses these relationships to rank hypotheses with supporting and opposing evidence; the ranking itself does not establish a cause.
The resulting state contains several kinds of knowledge. Observed elements are measurements, reported findings, or documented events. Calculated elements are reproducible derivations from those inputs. Inferred elements are interpretations dependent on evidence and assumptions, while simulated elements arise from a specified model and scenario. Unknown elements remain unresolved. Research status is an additional attribute rather than a substitute for these classes: an observed variant may have a research interpretation whose biological significance is uncertain. Keeping these dimensions separate allows the model to use available evidence without giving all information the same authority.
This representation is versioned whenever new evidence or a revised interpretation changes it. A new laboratory result may alter the estimate of the patient's condition; a new annotation may alter the understanding of an unchanged genomic finding. Those are different developments and must remain distinguishable. The same principle applies to missing information. Not measured, unavailable, failed quality review, not applicable, and below a reporting limit are different analytical states. Any imputation must be identifiable and evaluated for its effect on the result. A language model cannot invent values to complete the patient representation.
Data quality and traceability make this reconstruction usable for research. Identity checks, file integrity, unit consistency, duplicate detection, and comparison with source documents protect the intake process. Every derived finding retains a path back to its source and transformation. Readiness is then assessed for the particular question: incomplete evidence may permit a descriptive interpretation while preventing a numerical forecast. The patient state is complete enough for modeling only when the inputs needed by that model are supported at the required resolution. This question-specific standard turns heterogeneous records into a controlled starting point for the trajectory engine.
Modeling dysfunction and disease trajectories
Once a patient state has been reconstructed, the modeling problem is to estimate how it can change. A useful trajectory is attached to a specific outcome and interval. Disease onset, progression of established disease, treatment response, and recovery involve different endpoints even when they draw on overlapping information. The model must also distinguish an anticipated laboratory change from a clinical event. Genetic susceptibility provides one source of predictive information, but the transition from susceptibility to a future outcome depends on evidence about the individual, the relevant population, and the process being modeled.
Biological dependencies provide a basis for organizing those relationships. The proposed network connects genomic findings, molecular processes, physiological functions, exposures, and observations with explicit evidentiary qualifications. Experimental mechanisms, population associations, and patient-specific inferences retain separate meanings. Highly connected features do not automatically deserve greater causal weight, because connectivity can reflect research attention or annotation practices. Feedback and shared causes also matter: several abnormalities may reflect one process, while one abnormality may be consistent with several processes. Competing mechanisms remain available when the evidence cannot resolve them.
State estimation and longitudinal forecasting
The mathematical structure separates the underlying biological state from the measurements available to observe it. Let G represent constitutional genomic context, X at time t the modeled biological state, C the relevant context, A a recorded or modeled intervention, and Y the observations. The parameter set theta defines the transition and observation functions. A general formulation is:
The transition function describes change over the elapsed interval, while the observation function relates the state to measured quantities. Process variability and measurement error appear as separate terms because uncertainty about biological evolution differs from uncertainty in an assay or observation. These functions and error distributions must be specified for the particular application. The notation provides a structure for investigation, not estimated biological parameters or a claim that all relevant mechanisms have already been identified.
Constitutional genomic context does not replace time-varying tumor information. Acquired molecular findings enter the changing state or its observations with their specimen and temporal context. The state must also retain relevant history, such as cumulative exposure, whenever present measurements alone are insufficient to describe future behavior. In a sequential model, the chosen representation must contain the information on which the transition depends. Elapsed time requires explicit treatment when visits occur irregularly, and the forecast horizon must remain aligned with the question clinicians are trying to answer.
Because the complete state is not directly observed, the engine maintains a belief about plausible states given the history available at the cutoff. In the following expression, H denotes that history and b denotes the state distribution:
Forecasting propagates this distribution through the transition model. For an event endpoint, the result is a probability of the defined event within the selected interval. For a continuous endpoint, the result is a distribution over a future measurement with a corresponding prediction interval. Where repeated biomarkers and event timing are modeled together, longitudinal and time-to-event methods provide a methodological foundation [10]. The selected formulation still has to account for the population, observation process, and endpoint on which it will be evaluated.
New observations update the state estimate according to their consistency with the model and their measurement uncertainty. They may move the forecast, reduce uncertainty, or reveal an inconsistency that calls the model into question. The original forecast remains preserved so that the update does not erase what was predicted before the result became available. Updating one patient's state is also distinct from retraining the shared model. The former incorporates new evidence about that individual; the latter changes parameters or rules and requires its own version and evaluation.
Research on predictive cancer digital twins offers a relevant example of combining longitudinal imaging, mechanistic models, and Bayesian calibration to estimate tumor progression with uncertainty [8]. That work supports the methodological direction without demonstrating the broader Nexus application. Whole-cell computational research likewise illustrates reconstruction at a different biological scale: the model cited here concerns a bacterial organism, not a whole human [6]. These examples help define tractable modeling strategies while making clear that transfer across scales, diseases, and data types is an empirical question.
Sequential decisions and the Bellman formulation
A trajectory model becomes a decision model when candidate actions and their consequences are introduced. The Bellman equation relates the expected utility of a present action to the value of future decisions [12,13]. Its role in the Nexus framework begins after the state representation, action-dependent transitions, permitted choices, and objectives have been defined. A finite-horizon formulation using the uncertain patient state is:
The terminal condition specifies how the state is evaluated at the end of the horizon:
The value function represents the best expected cumulative utility available within the model. The permitted action set constrains the choices under consideration. Immediate utility evaluates consequences over the next decision interval, and the expectation accounts for possible future states and observations that update the belief. The terminal utility evaluates the final state. The discount factor controls the relative contribution of later consequences; a value of one supports an undiscounted finite-horizon objective. Any discounting requires an explicit rationale appropriate to the study.
These terms have substantive clinical implications. A utility function cannot combine benefit, toxicity, functional status, and treatment burden using arbitrary weights and still claim clinical meaning. The horizon also matters, because an apparently favorable short-term result may carry later consequences. Hard safety boundaries belong in the permitted action set, and the interpretation of tradeoffs requires clinical input. The equation identifies a way to organize sequential evaluation; it does not determine what matters to the patient or establish which interventions are acceptable.
The most demanding requirement is the validity of action-dependent transitions. Outcomes observed in differently treated patients may reflect differences in disease severity, selection, or other factors. They do not automatically estimate what would happen if treatment changed for the current patient. Causal inference is therefore necessary for interpreting intervention comparisons, with assumptions appropriate to the data and study design [11]. The initial implementation should evaluate defined clinician-selected scenarios. Broader optimization becomes defensible only when the estimated action effects and the resulting policy have adequate evidentiary support.
Uncertainty remains part of both forecasting and decision analysis. It arises from incomplete observations, parameter estimation, measurement error, biological variability, and the structure of the model itself. Sensitivity analysis should establish whether conclusions persist under plausible changes in assumptions, and alternative models should be examined when several mechanisms fit the evidence. A narrow numerical interval does not resolve a missing biological process. Forecasts should be restricted when the patient falls outside the supported population or when necessary inputs are unreliable. These limits identify what the next observation or study must resolve before a simulated future can inform a consequential decision.
Biological correction and treatment response
The purpose of intervention modeling is to examine whether changing a specified condition could improve a meaningful outcome for the individual. This connects the reconstruction of biology to a testable consequence. A correction target therefore includes a starting condition, an outcome, an assessment interval, and criteria for improvement. Disease control, physiological function, symptoms, recovery, and tolerability may all matter, but they are distinct dimensions. The clinical interpretation must address their relationship rather than allow improvement in one measurement to stand for an overall benefit.
An intermediate measurement can be useful when its relationship to the relevant clinical outcome is understood. A surrogate endpoint, however, requires evidence that supports its use in the particular context [18]. This distinction applies to BioMatrix system scores as well as individual laboratory values. The engine should expose the measurements contributing to a score and retain the difference between an analytical change and a demonstrated improvement in the patient's health. In oncology, a disease-response endpoint may need to be considered alongside organ function and treatment burden. Their relative importance is a clinical question that must be specified before scenario results are interpreted.
Modeling intervention effects and comparing scenarios
An intervention enters the model through an evidence-based hypothesis about the biological process it may influence. The hypothesis identifies the target, expected direction of effect, timing, and patient characteristics relevant to response. For pharmacological research, recorded exposure may include the agent, route, schedule, duration, modifications, and interruptions. Planned administration and documented administration remain distinct. If exposure is estimated, the estimate must rely on an appropriate method and sufficient inputs. Otherwise, uncertainty in what was received becomes uncertainty in the interpretation of the response.
The proposed mechanism may involve several pathways and timescales. Benefit may coexist with toxicity, compensation, or interactions with concurrent interventions. The model must state which effects it represents and which remain outside its scope. Foundation language models can help assemble mechanistic evidence, but the numerical effect sizes and transition parameters belong to the quantitative model and need identifiable sources. Evidence supporting the direction of a possible effect does not establish its magnitude for an individual. When only qualitative support is available, the result should remain a qualitative hypothesis or an explicitly assumption-dependent scenario.
Comparison requires a common patient-state version, analytical cutoff, endpoint, and horizon. Each scenario specifies its intervention and timing, assumptions about administration, and any rules for subsequent decisions. Contextual conditions that remain fixed and conditions allowed to change must be stated. This common basis allows a reviewer to identify why forecasts differ. Without it, an apparent intervention effect could arise from a different starting state, a different outcome definition, or additional information introduced into only one scenario.
The output should show expected outcomes together with uncertainty and sensitivity to the scenario assumptions. Depending on the question, it may compare event probabilities, time to an event, or the distribution of a future measurement. A definitive ranking is inappropriate when the model cannot reliably distinguish the alternatives. In that circumstance, the useful result may be a tradeoff or the identification of an observation that would reduce uncertainty. The comparison remains conditional on the transition model, and claims about the effect of changing treatment require the causal support discussed in the preceding chapter [11].
Oncology interpretation and longitudinal response
For cancer applications, the patient state must connect broader physiological condition with disease-specific evidence. Pathology, classification, extent of disease, molecular findings, imaging, treatment history, and documented response assessments define the context in which a proposed mechanism is relevant. Constitutional susceptibility and tumor-associated findings remain separate evidence streams. Each molecular result retains its specimen and timing because a biomarker may help identify a potentially relevant treatment without guaranteeing a response, and the tumor's characteristics may change during its course [5].
Treatment history establishes how the current state was reached. Earlier responses, adverse effects, modifications, interruptions, and concurrent interventions help interpret the observed trajectory. For diseases in which measurable residual disease is relevant, the assay, specimen, sensitivity, and reporting threshold remain attached to the result. A negative result means the test did not detect disease within its capabilities; it cannot be detached from those capabilities when used in modeling [19]. General laboratory improvement likewise cannot replace a formal disease-response assessment performed using the appropriate criteria and documented version.
Assessment of a treatment plan should bring these strands together. The central questions are whether the evidence applies to the patient's disease setting, whether the observed biology supports the proposed mechanism, whether relevant constraints are represented, and whether the actual course is consistent with expectations. The output should expose supportive and contradictory findings, unresolved information, and the temporal relevance of the evidence. It constitutes an assessment of applicability and observed response, with clinical conclusions remaining subject to the treating team's judgment and information that may lie outside the model.
Longitudinal monitoring tests the model against observations that were unavailable when the forecast was issued. A departure from expectation should be evaluated in relation to its magnitude, predictive interval, persistence, and clinical context. Potential explanations include altered exposure, a change in disease behavior, intercurrent illness, measurement problems, and model inadequacy. The engine should examine these possibilities before labeling the difference as resistance or deterioration. Research into evolutionary treatment strategies illustrates the importance of studying changing tumor behavior in a defined disease setting, but findings from a specific prostate cancer study do not establish transferability to other cancers or to the Nexus framework [7].
Recovery also requires interpretation across the outcomes selected at the outset. Improvement in one dimension should remain visible alongside deterioration in another, and temporary changes should be considered in relation to the expected treatment course. Alert thresholds need evaluation for false alerts, missed events, and useful lead time. A reviewable alert identifies the observation, the expected range or trajectory, the discrepancy, and the evidence that could clarify it. Updating the patient's forecast should preserve the original estimate and explain whether the change arose from new evidence or a revised model.
These applications make the evidentiary progression central to clinical translation. Mechanistic support explains why an effect may be plausible. Simulation estimates consequences under defined assumptions. Observed response records what happened after an intervention without, by itself, establishing why it happened. Demonstrated clinical benefit requires relevant outcomes, an appropriate comparator, and a study capable of supporting the claim. Predictive accuracy under observed care and the benefit of model-informed decisions are therefore separate research questions. A system may forecast response well while providing no improvement in care, so the clinical value of acting on its outputs requires its own evaluation.
Research architecture and evidence governance
The proposed architecture implements the preceding scientific argument through a shared patient state and explicit boundaries between interpretation, computation, and review. WGS analysis supplies normalized genomic findings with specimen context and analytical quality. BioMatrix relates those findings to laboratory biology and clinical history. The Disposition Trajectory Engine applies eligible quantitative models to estimate future outcomes or compare defined scenarios. Each component contributes structured evidence that remains inspectable as it moves through the workflow.
The interface between these components is consequential. A narrative summary may omit an uncertainty, compress several observations into one conclusion, or obscure the date on which information became available. The computational exchange should therefore carry structured records with source identifiers, state versions, methods, and time intervals. Narrative explanations accompany those records for the reader. Execution begins with readiness for the selected question, proceeds through reconstruction and interpretation, and then invokes models whose input and population requirements are satisfied. Missing prerequisites block the affected calculation while leaving supported analyses available.
Reasoning and quantitative execution
Foundation LLMs are proposed for interpreting heterogeneous clinical and scientific material. They can assist with extraction of documented events, organization of evidence, development of candidate mechanisms, examination of conflicting explanations, and translation of numerical results into readable findings. The design accommodates models from OpenAI and Anthropic through a common task interface. For independent review, each receives the same authorized evidence package and output requirements, with provider and model versions recorded. These roles are proposed architectural choices whose contribution must be evaluated.
An evidence package contains the relevant observations, retrieved source material, analytical question, and permitted scope. A generated claim either points to identifiable supporting material or remains labeled as a hypothesis. Source verification must examine whether the cited material supports the claim, rather than merely whether a citation exists. Undocumented patient facts, invented numerical parameters, and assumed normal values cannot enter the state through language generation. The aim is to improve the handling of evidence while retaining control over what becomes a patient fact or a computational input.
Quantitative execution follows a different contract. Every model has a defined endpoint, supported population, input requirements, horizon, parameter source, and validation status. A registry retains the implementation, configuration, development data, evaluation results, and known limitations. The appropriate model depends on the question and available evidence; a simple validated model may be preferable to an elaborate formulation whose parameters cannot be estimated reliably. Before execution, the system checks compatibility between the requested analysis and the model's documented scope.
The numerical output includes the estimate, uncertainty, assumptions, and execution metadata. Bellman-based evaluation also identifies the permitted actions, transition assumptions, utility function, and decision horizon. These are part of the result's meaning, so the narrative layer must preserve them. Reproducible computation allows an expert to determine whether a result follows from the stated model even when the biological adequacy of that model remains under investigation. This separates a computational error from disagreement about the underlying science.
Independent review and reproducibility
Cross-model review begins with independent analyses before either model sees the other's interpretation. Comparison then occurs at the level of claims and sources. Factual discrepancies, source conflicts, alternative interpretations, missing evidence, and numerical inconsistencies require different responses. A single agreement score would obscure those distinctions. Where disagreement persists, each model may examine the opposing claim and its evidence, but resolution should depend on verification, applicability, and calculation rather than repeated attempts to produce matching language.
Agreement is a review signal rather than independent proof. The workflow needs a stopping rule, and consequential unresolved claims proceed to expert review with their alternatives intact. Preserving the initial analyses and subsequent revisions makes it possible to assess whether reconciliation corrected an error, resolved an ambiguity, or simply concealed disagreement. Evaluation of this review process must include the quality of resolved claims and the consequences of errors that both models accept. The additional complexity is justified only if it improves a predefined research or review outcome.
Reproducibility requires a record of each run. The manifest identifies source-file hashes, the patient-state version, transformations, reference datasets, software versions, parameters, and model configuration. LLM steps additionally preserve the provider, model identifier, prompt version, evidence package, and actual output. Numerical calculations should repeat within documented tolerances; stochastic simulations retain seeds, sample counts, and convergence settings. Language-model outputs remain archived because a later request may produce a different interpretation or wording even when its inputs appear unchanged.
Corrections create new versions with documented reasons. An annotation update or parameter change must not silently alter a historical forecast, because evaluation depends on knowing what was predicted with the methods and information available at the time. Access controls and the audit trail should also make clear who reviewed, accepted, revised, or rejected consequential findings. These records support scientific scrutiny and the operational investigation of errors without requiring clinicians to examine every internal processing step during routine review.
The clinical presentation should bring the analytical question, relevant observations, principal findings, horizon, uncertainty, and unresolved issues into one readable account. A reviewer should be able to trace a conclusion to its evidence, identify the strongest contradiction, and see what additional information could change it. Genomic, oncology, and quantitative questions should reach the appropriate expertise. The final interpretation records the reviewer's rationale and accurately reflects the underlying results. This architecture makes the evidence accessible enough to challenge and the computation explicit enough to reproduce, which are prerequisites for the validation program that follows.
Validation and clinical translation
The framework's scientific value depends on performance against outcomes that were not available when its forecasts were generated. Implementation establishes that the components can execute together. Predictive validity establishes how well the output describes subsequent observations. Clinical usefulness concerns whether the information assists a meaningful decision, and demonstrated benefit concerns what happens when it is used in care. The research program must evaluate each claim at the appropriate stage rather than treat successful software operation as evidence for all of them.
Study design and predictive evaluation
The initial study should address one primary endpoint in a clearly defined population. Selection depends on clinical relevance, available longitudinal data, adequate outcome events, and the ability to reconstruct when inputs became available. The protocol defines eligibility, prediction start time, required information, forecast horizon, outcome ascertainment, and follow-up. For oncology, cancer type, disease setting, and treatment context are necessary parts of that definition. Predisposition forecasting needs its own population at risk and rules for establishing incident disease. Exploratory endpoints remain distinguished from the primary question.
A suitable comparator is specified before evaluation, such as an established clinical model or a model based on routine clinical and laboratory data. Sample size is justified by event frequency, development requirements, and the desired precision of performance estimates. Multiple observations from one patient do not increase the number of independent patients. Missing follow-up, censoring, and competing events need a prespecified strategy. TRIPOD+AI provides reporting guidance that makes these design choices and results assessable, while leaving the substantive validity of the model to the evidence [16].
Retrospective analysis must reproduce the information available at each prediction cutoff. Later notes, molecular reports, and outcome summaries cannot be included merely because they refer to an earlier period. Training, model selection, calibration, and final evaluation require appropriately separated data, with patient records, related individuals where relevant, and duplicate specimens handled to avoid dependence that compromises evaluation. Imputation, feature selection, normalization, and threshold selection are fitted within development data. The final test data remain unavailable for iterative prompt revision, parameter adjustment, or model selection.
Temporal evaluation examines performance on later patients, while external evaluation examines transfer to another institution or setting. Both are important but answer different questions. LLM-assisted processing adds a further concern: retrieval packages must exclude future patient information and evaluation labels. If historical records are interpreted using current scientific knowledge, the study must describe that design rather than imply that the complete system was available at the historical date. These controls preserve the distinction between a credible retrospective test and an explanation assembled with knowledge of the outcome.
Prospective evaluation should initially run in silent mode, with forecasts generated but not used to direct care. Each result is timestamped and locked with its state version, inputs, model, endpoint, horizon, and uncertainty before the outcome becomes available. New predictions and corrections create new records. Failed runs, missing inputs, withheld predictions, and delays remain part of the study because they affect practical coverage. Outcome assessment follows predefined criteria and, where feasible, occurs without access to the prediction. Treatment changes during follow-up are retained so the forecast can be interpreted against the course actually observed.
Performance measures must match the endpoint. Event prediction requires assessment of discrimination, calibration, and performance at relevant prespecified thresholds. Sensitivity, specificity, and predictive values need interpretation alongside event frequency and the consequences of errors. Calibration tests whether predicted probabilities correspond to observed frequencies and is essential when an output is intended to communicate risk [14]. Continuous forecasts require error measures and prediction-interval coverage. Time-to-event analysis accounts for censoring and competing events, and repeated predictions require methods that recognize within-patient dependence.
Practical usefulness adds another set of measures: lead time, false-alert burden, missed events, clinician review time, and the proportion of eligible patients receiving a usable result. Decision-curve analysis can examine net benefit where the decision and threshold assumptions are appropriate [15]. Uncertainty intervals and relevant subgroup analyses should accompany aggregate results. Age, ancestry, disease subtype, treatment setting, and site may matter, but small samples cannot establish equivalent performance. The system's abstention rate and performance among accepted cases should be reported together so that selective prediction does not conceal limited coverage.
Component comparisons then determine whether the architecture contributes more than its baseline. Prespecified analyses can add genomics, mechanistic constraints, LLM-assisted evidence processing, and cross-model review in a controlled sequence using the same endpoints and data partitions. Tuning opportunities, computational effort, and evidence access should be comparable or explicitly documented. Extraction accuracy, source fidelity, unsupported claims, and interpretation consistency measure the quality of intermediate work. Downstream performance establishes whether those changes matter to forecasting or clinical review. A component that adds cost and failure modes without measured benefit may appropriately be simplified or removed.
Governance and the development pathway
Clinician-visible evaluation introduces questions that silent prediction cannot answer. The study must examine how users understand the output, whether it changes decisions, and whether new errors arise through its use. DECIDE-AI provides reporting guidance for this early clinical stage, including the human and workflow context in which an AI system operates [17]. A model that predicts response accurately has not thereby established that acting on its recommendations improves outcomes. Comparative evaluation of model-informed care, with randomization where appropriate and feasible, is needed to support that stronger claim.
Each stage requires a documented intended use, accountable clinical leadership, and the applicable institutional research review. Consent, data-use permissions, ethics review, and regulatory assessment must be resolved for the particular study before enrollment or clinical deployment. Research approval does not establish effectiveness or authorize unrelated uses. Oncology outputs retain clinician authority. The proposed pediatric implementation excludes autonomous treatment changes and patient-specific dosing recommendations, requires developmentally appropriate interpretation and specialist oversight, and treats adult-to-pediatric transfer as a separate validation question. Parental permission and child assent are addressed where applicable.
Privacy is part of the study design and operating architecture. Direct identifiers should be minimized, access restricted by role, and external processing documented. Genomic information remains sensitive after direct identifiers are removed. Provider access, retention, and processing arrangements require verification for the actual workflow. Monitoring must also detect changes in data, performance, or clinical use after release. Incident investigation, restrictions on affected outputs, and rollback procedures make it possible to respond when a consequential error or unsupported application is identified.
The development pathway begins with research specification and data readiness. The intended use, cohort, endpoint, permissions, data dictionary, and analysis plan are settled before model development is evaluated. The next milestone is a reproducible baseline with validated transformations, a versioned patient state, and traceable numerical execution. Synthetic cases can establish workflow behavior and expose technical failures, but clinical predictive performance depends on patient outcomes. Retrospective evaluation then combines locked testing, leakage audits, component comparisons, and analysis of the conditions under which the system fails or abstains.
Only after those criteria are met should the fixed system proceed to prospective silent evaluation. That stage tests both forecasting and operational reliability on incoming cases. Supervised clinical evaluation follows under an appropriate protocol, examining interpretation, decision effects, and safety. Comparative evaluation can then determine whether model-informed care improves relevant outcomes. Expansion to new populations, endpoints, or clinical settings requires separately documented evidence. Each milestone should have explicit exit criteria and a valid path for revision or discontinuation when the evidence does not support advancement.
Scientific collaboration is most valuable where it contributes independent knowledge and testing. Well-characterized longitudinal cohorts, genomic interpretation, quantitative oncology, causal inference, and independent outcome assessment address different weaknesses that one team may not resolve alone. External challenge testing and replication should therefore be part of the program wherever feasible. The proposed Nexus contribution is a connected framework in which genomic evidence, measured biology, and sequential reasoning can be tested together. Its eventual scope should be defined by the questions it can answer reliably for identifiable patient populations, with a documented path from explanation to prediction and from prediction to demonstrated benefit.
References
These sources support the scientific concepts and methodological foundations of the framework. Their findings do not establish predictive performance or clinical effectiveness for the proposed Nexus implementation.
Richards S, Aziz N, Bale S, et al. Standards and guidelines for the interpretation of sequence variants: a joint consensus recommendation of the American College of Medical Genetics and Genomics and the Association for Molecular Pathology. Genetics in Medicine. 2015;17:405–424. doi:10.1038/gim.2015.30
Marshall CR, Chowdhury S, Taft RJ, et al. Best practices for the analytical validation of clinical whole-genome sequencing intended for the diagnosis of germline disease. npj Genomic Medicine. 2020;5:47. doi:10.1038/s41525-020-00154-9
National Human Genome Research Institute. Gene–environment interaction. Talking Glossary of Genomic and Genetic Terms. Institutional reference. Institutional source
National Cancer Institute. Genetic testing for inherited cancer risk. Institutional fact sheet. Institutional source
National Cancer Institute. Biomarker testing for cancer treatment. Institutional clinical information resource. Institutional source
Karr JR, Sanghvi JC, Macklin DN, et al. A whole-cell computational model predicts phenotype from genotype. Cell. 2012;150:389–401. Publication
Zhang J, Cunningham JJ, Brown JS, Gatenby RA. Integrating evolutionary dynamics into treatment of metastatic castrate-resistant prostate cancer. Nature Communications. 2017;8:1816. doi:10.1038/s41467-017-01968-5
Pash G, Villa U, Hormuth DA II, Yankeelov TE, Willcox K. Predictive digital twins with quantified uncertainty for patient-specific decision making in oncology. arXiv. 2025. Preprint, version 1. arXiv:2505.08927. Preprint v1
Stahlberg E. Digital twins for cancer: not if, but when, how, and why? National Cancer Institute, Center for Biomedical Informatics and Information Technology. January 6, 2025. Institutional research commentary. Institutional source
Rizopoulos D. Joint Models for Longitudinal and Time-to-Event Data: With Applications in R. CRC Press; 2012. Methodological monograph. doi:10.1201/b12208
Hernán MA, Robins JM. Causal Inference: What If. Chapman & Hall/CRC; 2020. Methodological book with an author-maintained online edition. Author resource
Bellman R. A Markovian decision process. RAND Corporation; 1957. Research paper P-1066. Primary source
Sutton RS, Barto AG. Reinforcement Learning: An Introduction. 2nd ed. MIT Press; 2018. Methodological textbook covering value functions, Bellman equations, and sequential decision methods. Author resource
Van Calster B, McLernon DJ, van Smeden M, Wynants L, Steyerberg EW. Calibration: the Achilles heel of predictive analytics. BMC Medicine. 2019;17:230. doi:10.1186/s12916-019-1466-7
Vickers AJ, Elkin EB. Decision curve analysis: a novel method for evaluating prediction models. Medical Decision Making. 2006;26:565–574. Publication
Collins GS, et al. TRIPOD+AI statement: updated guidance for reporting clinical prediction models that use regression or machine learning methods. BMJ. 2024;385:e078378. doi:10.1136/bmj-2023-078378
Vasey B, Nagendran M, Campbell B, et al. Reporting guideline for the early-stage clinical evaluation of decision support systems driven by artificial intelligence: DECIDE-AI. Nature Medicine. 2022;28:924–933. doi:10.1038/s41591-022-01772-9
National Cancer Institute. Surrogate endpoint. NCI Dictionary of Cancer Terms. Institutional terminology reference. Institutional source
National Cancer Institute. Minimal residual disease. NCI Dictionary of Cancer Terms. Institutional terminology reference. Institutional source