Source-linked AI summary
Life Operators: a self-evolving framework for multiscale life modelling
Shuo Wang, Yike Guo
TL;DR
Medical AI can forecast records, but those forecasts do not by themselves identify biological state or establish what would happen under an alternative intervention. Life Operators define task-bounded roles and graphs for composing models; the framework's proposed self-evolution is cumulative, evidence-based revision, with broader models remaining a possibility to test.
Problem
Forecasting clinical records does not by itself identify task-relevant biological states or establish an individual patient's outcome under an alternative intervention.
Source: §1 · p. 2Method
Results
The paper presents evidence-gated self-evolution as cumulative revision of explicit, versioned models rather than unconstrained self-rewriting.
Source: §4 · p. 9Takeaways & Limitations
Validated operators and bridges could accumulate into broader multiscale models, but that possibility must be tested through bounded medical problems.
Source: §1 · p. 3The immediate objective is not a complete digital replica; broader models remain an empirical question to test through bounded medical problems.
Source: §1 · p. 3
Abstract
from arXiv · showhide
Medical AI is moving beyond recognition towards clinical dialogue and longitudinal prediction. Yet a central question remains: how would a patient's state change under intervention? Statistical models learn future observations, whereas mechanistic models describe selected processes. Neither provides a common framework for representing patient state, coupling scales or revising failed assumptions. We propose Life Operators: task-bounded mappings that define three scientific roles. Perception operators infer task-relevant biological states from multimodal observations, Evolution operators propagate these states under natural or intervention-conditioned dynamics, and Generation operators map them to measurable signals. Each role may be realised by equations, statistical models, neural networks or hybrids. Bridge operators connect components with different variables, scales and time steps. Selected operators and bridges form task-specific Operator Graphs containing the smallest set of states and mechanisms sufficient for a declared claim. This modular structure also makes scientific revision localisable. An AI co-scientist may propose changes to states, operators, bridges or graph structure, while independent evidence determines which variants are retained, restricted or retired. Over time, validated components could accumulate into broader multiscale models of the human body and provide a computational foundation for medical artificial superintelligence.
1 Introduction
Medical AI can recognize states and forecast records, but intervention questions require models that represent biological states and how interventions alter their dynamics. Life Operators offer a common, modular language for combining partial models around bounded medical questions while keeping their assumptions explicit.
1 Introduction
- Figure 1 aligns three medical questions with biological organization from molecular to individual scales and representative interventions, with bidirectional interactions between adjacent scales.Source: Figure 1 · p. 2
- Intervention questions require models in which interventions enter biological dynamics, because recognition can omit temporal change and forecasting can omit how treatment changes the process.For the same patient, watchful waiting, a drug, ablation or a device may lead to different future states and clinical paths.Source: §1 · pp. 1–2
- Predicted clinical records do not by themselves identify biological states or transition laws, and observed treatment labels do not establish individual counterfactual outcomes.Records reflect both patient change and care processes such as testing, treatment assignment, follow-up and coding [8].Source: §1 · p. 2
- Mechanistic models provide explicit variables and assumptions for selected processes, but local fidelity does not ensure patient specificity or valid composition across scales.Parameters may be weakly identifiable, boundary conditions incompletely observed, and effective relations unreliable when transferred across scales or interventions [9–11]; cardiac digital twins illustrate patient-specific parameterisation [12, 13].Source: §1 · p. 3
- Life Operators define a common language for components that infer, evolve or generate, while allowing equations, statistical models, neural networks or hybrids as implementations.Bridge operators support composition, and evidence-gated revision can change components without implying common algorithms or equal evidential status.Source: §1 · p. 3
- The framework starts from a declared endpoint, horizon, interventions and observations, then selects the smallest state and graph sufficient for that question.Accumulation into broader body models remains an empirical question to test through bounded medical problems; medical artificial superintelligence is described as a possible long-term consequence.Source: §1 · p. 3
2 Life Operators
Life Operators represent health and disease as partially observed dynamics, with a task-specific state linking observations, context and interventions to outcomes. Perception, Evolution and Generation specify the roles of inference, state propagation and signal formation, while Bayesian filtering keeps patient change distinct from changing knowledge.
2 Life Operators
- For a specified problem, health and disease are represented as open, stochastic, controlled and partially observed dynamics, with context, interventions and unresolved influences shaping state transitions.The formulation distinguishes the task-relevant state, transition map, intervention, context and unresolved stochastic influences.Source: §2 · p. 3
- The task-relevant state is an abstraction, not a complete reconstruction or lossless compression, and is sufficient when the observed past adds no material outcome information given state, allowed inputs and context.The sufficiency claim is limited to a stated domain and outcome horizon.Source: §2 · p. 4
- The same operator roles may be implemented by equations, numerical solvers, statistical transition models, neural networks or hybrids rather than a single algorithm class.Mechanistic implementations encode selected structure, while learned components may approximate unresolved relations or accelerate simulation and inversion [14–16].Source: §2 · p. 5
- Perception infers a posterior over task-relevant states, Evolution propagates states under context and relevant interventions, and Generation maps states to measurable-signal distributions.The roles respectively represent inference uncertainty, natural or intervention-conditioned transitions, and signal formation and acquisition.Source: §2 · pp. 4–5
- Bayesian filtering connects the three roles: Evolution supplies a transition prior, Generation an observation likelihood, and Perception the posterior update.The simplified recursion omits parameters and acquisition variables.Source: §2 · p. 5
- The filtering recursion separates patient change from changing knowledge and need not collapse natural variability, measurement error, parameter uncertainty and model-form uncertainty into one confidence score.Source: §2 · p. 5
3 Operator Graphs
Operator Graphs select components for a declared medical question and connect them through explicit interfaces or bridges. Their validity depends on task-specific evidence, and the HCM example lays out a bounded prediction graph and tests rather than a validated individual-response model.
3 Operator Graphs
- A task-specific Operator Graph selects Perception, Evolution and Generation components and connects them through compatible interfaces or bridge operators sufficient for a declared claim.The endpoint, time horizon, interventions and available observations determine the graph's states, components, bridges and scales.Source: §3 · pp. 5–6
- Bridge operators transform one component's output into the typed input required by another, carrying an effective parameter, source term, boundary condition or distribution with uncertainty.Scale Bridge mappings in upward and downward directions have different meanings and are not assumed to be inverses.Source: §3 · p. 6
- Scientific validity requires preserving the meaning, units, timing and uncertainty needed by the receiver; computational compatibility or speed alone is insufficient.Contracts also record component roles, assumptions, valid domains, evidence, versions and failure conditions.Source: §3 · p. 6
- Graphs should be validated against independent observations for the receiving component and declared task, then revised by changing the message, narrowing scope or activating a finer model if validation fails.Source: §3 · p. 6
- Graph expansion should answer an evidence need: missing observations call for measurement or wider uncertainty, missing intervention evidence calls for interventional evidence, and unhelpful components should be removed.Source: §3 · p. 6
- The HCM example proposes a bounded prediction of joint Valsalva LVOT-gradient and LVEF changes under a declared mavacamten regimen, not a complete digital twin.Its component tests include held-out pharmacokinetic profiles, human cellular data, independent bridge and organ observations, and blinded protocol-matched echocardiography; integration into an individual-response graph is not validated [17–19].Source: §3 · p. 7
4 Evidence-Gated Self-Evolution
Life Operators self-evolve through bounded, versioned revisions: AI proposes changes to model components, while independent evidence determines which changes persist. Explicit contracts and modularity help localise failures and allow accepted components to be reused.
4 Evidence-Gated Self-Evolution
- Versioned components make failures traceable to a specific operator or graph element, so changes can be localised rather than retraining an undifferentiated system.Components may be added, replaced, restricted, or retired while the rest of the graph remains fixed.Source: §4 · pp. 7–8
- Accumulated data and discrepancies can prompt AI agents to propose new operators, bridges, or graph structures, each with a scientific contract.Contracts specify inputs, outputs, assumptions, uncertainty, domain, and failure conditions.Source: §4 · pp. 8–9
- Candidate model variants may change state definitions, operator libraries, bridge libraries, or graph structure rather than merely adjust parameters.The model represents state definitions, operator libraries, bridge libraries, and graph structure as explicit components.Source: §4 · p. 9
- Independent data or discriminating experiments select among the current model and proposed revisions, so generating a proposal does not make revision compulsory.When evidence cannot distinguish candidates, retain the current model, narrow its domain, or carry structural uncertainty forward.Source: §4 · p. 9
- Accepted operators enter a versioned library for recomposition, making self-evolution cumulative evidence-based revision rather than unconstrained self-rewriting.Task-specific use can expose discrepancies that motivate revisions to the shared scientific model.Source: §4 · pp. 8–9
5 Biomedical Applications
The paper describes Operator Graphs as potential tools for designing person-specific therapies, exploring candidate interventions and trials, and adapting care as patient observations accumulate. These applications would focus, not replace, laboratory experiments and clinical trials.
Personalised therapeutics
- An Operator Graph could design a person-specific therapeutic by generating and comparing candidate molecules, therapies, or regimens across doses, schedules, and delivery strategies.The graph could evaluate efficacy, toxicity, resistance, and long-term outcomes; mRNA neoantigen vaccines are one example using mutation-derived neoantigens and HLA information.Source: §5 · pp. 9–10
Virtual development and trials
- Operator Graphs could explore candidate drugs, devices, and procedures across virtual cohorts to identify response and toxicity subgroups, biomarkers, and informative clinical comparisons.In silico development would help focus physical experiments and trials on candidates and populations most likely to be informative.Source: §5 · p. 10
Adaptive precision care
- For an individual patient, repeated state updates could support comparing interventions and adjusting care as observations reveal deviations from expected trajectories.Comparisons could include intermediate biological changes, time to benefit, adverse effects, recovery paths, and future treatment options.Source: §5 · p. 10
Biomedical applications
- These applications would make laboratory experiments and clinical trials more targeted, not eliminate them.The proposed focus is on selecting candidates, patient groups, and intervention contrasts most worth testing.Source: §5 · p. 10
6 Towards Life Modelling Infrastructure
The proposed infrastructure matches evidence to claims, tests validity after components are composed, and supports transparent reuse across sites and populations. A staged path begins with bounded operators and minimal graphs rather than a complete digital human.
Evidence matched to claims
- Evidence should match each operator's or graph's intended claim, with intervention-response claims requiring interventional evidence or an explicit causal identification strategy.Clinical claims additionally require evidence of benefit and safety in the intended workflow.Source: §6 · p. 10
Validity after composition
- Valid operators do not guarantee a valid composed graph, so validation should examine uncertainty transport, intervention response, long-horizon stability, and calibration.Added components are justified only when they improve the task, explain a failure, or change an experimental or clinical decision.Source: §6 · p. 11
Infrastructure for reuse and revision
- Reusable operators and bridges require open, versioned registries with machine-readable contracts documenting interfaces, assumptions, domains, provenance, evidence status, and known failures.Shared standards, reference implementations, and benchmarks would support reproducibility and comparison.Source: §6 · p. 11
International collaboration for validation
- International cohorts and federated analysis can support evaluation across sites and populations where person-level data cannot be pooled.Diverse cohorts can test transportability and reveal missing or non-transferable operators.Source: §6 · p. 11
A staged path
- A staged path begins with independently measurable bounded operators, validates bridges at both scales, and builds minimal graphs around explicit interventions and prospective endpoints.Shared registries and cohort networks can support benchmarks, external validation, and evidence-gated reuse.Source: §6 · p. 11