B3: Causal disease modelling

A correlation says "patients with low vitamin D have higher infection rates." This single observation is consistent with at least four distinct causal structures: (a) low vitamin D causes infections; (b) infections deplete vitamin D; (c) a confounder such as sun exposure, time outdoors, or baseline frailty causes both; or (d) the correlation is sampling noise. The observational data alone cannot distinguish them — all four produce the same correlation coefficient.

Current medical AI excels at finding such correlations from observational data but cannot, from that data alone, distinguish between these explanations. The causal question — "if I intervene to raise this patient's vitamin D, will their infection rate drop?" — is a question about a counterfactual world that the observational data never sampled. Answering it requires either randomised intervention (the basis of clinical trials) or causal inference methods that combine observational data with explicit, untestable-in-principle assumptions about the absence of unmeasured confounders.

The formal statement of the gap is Pearl's "ladder of causation": observational data lives on rung one (seeing), interventional questions on rung two (doing), counterfactual questions on rung three (imagining). A model trained purely on rung-one data cannot answer rung-two or rung-three questions without additional causal assumptions injected from outside the data. No amount of observational data, however large, climbs the ladder on its own.

For AIHS, the relevant question is almost always at rung two or three: given this patient's state, will this intervention improve it? Pattern-matching from observational data is necessary but fundamentally insufficient. This is the precise reason B3 is the central scientific gap in the entire study — it is not an engineering problem that scales away with more compute or more data of the kind we already collect.

RCTs provide gold-standard causal evidence but with three structural limitations that prevent them from solving B3 on their own.

They produce average effects, not individual effects. An RCT shows that "drug X reduces mortality 15% across the trial population." It does not directly tell you whether drug X will help this particular patient with their specific genetic background, comorbidities, and current physiological state. The average treatment effect can mask enormous heterogeneity — a drug that helps 30% of patients substantially, harms 10%, and does nothing for 60% can show the same average as a drug that helps everyone modestly. AIHS needs the individual prediction, which the trial's average does not provide.

They cover only the interventions and populations tested. The space of possible interventions, doses, timings, and combinations is astronomically larger than the space of interventions that have been or could ever be subjected to RCTs. Most clinical decisions involve some extrapolation beyond directly-tested evidence.

They are slow, expensive, and sometimes unethical or impossible. A definitive RCT takes years and tens to hundreds of millions of dollars. For rare diseases, individualised combination therapies, or situations where withholding treatment would be unethical, the relevant RCT will never be run.

Bridging from trial evidence to individual prediction requires one of: extensive subgroup analysis within trials (statistical power collapses as subgroups shrink); individual patient data meta-analysis across trials (rare, resource-intensive, and limited by what was originally measured); or causal models that extrapolate from trial evidence plus mechanistic knowledge plus the individual's data to a personalised counterfactual prediction.

The third approach is what B3 is fundamentally about. RCTs are an essential input — they provide the rung-two interventional anchors that pure observational data lacks — but they are an input to causal modelling, not a substitute for it.

These three terms are often conflated but refer to distinct things that combine in AIHS interpretation.

Mechanistic models start from known biology — biochemical reaction kinetics, physiological equations (e.g. the Hodgkin-Huxley equations for neuronal firing, Windkessel models for arterial flow), anatomical structure — and produce predictions from first principles. They have causal structure built in: the equations explicitly specify what causes what. Their strength is that they extrapolate correctly to situations outside any training data, because they encode actual mechanism. Their weakness is that they require extensive parameterisation and cannot represent the large fraction of biology we do not yet mechanistically understand.

Causal ML models infer causal structure from data, typically by combining observational data with interventional data (from perturbation experiments) and structural assumptions. They can capture relationships we don't mechanistically understand — they learn the causal graph rather than being handed it. Their weakness is that identifying causal direction reliably requires either very large interventional datasets or strong assumptions that may not hold, and they extrapolate poorly beyond the regime of their training data.

Digital twins (B4) are patient-specific instantiations of either or both model types, parameterised from an individual patient's measured data, used to simulate that patient's likely response to candidate interventions.

The most promising direction is hybrid: mechanistic structure where biology is known (cardiac electrophysiology, renal filtration, basic pharmacokinetics), causal-learned components where it isn't (most of immunology, most of cancer signalling, most of neuropsychiatry), with the whole parameterised per-patient. This is the trajectory of programmes like the Physiome Project and the Living Heart Project, and it is where AIHS interpretation should ultimately converge. The hybrid approach inherits the extrapolation reliability of mechanism where mechanism is known and the flexibility of learning where it isn't — but stitching the two together coherently, with consistent uncertainty propagation across the boundary, is itself an open research problem.

Two formal frameworks dominate, and they are largely inter-translatable.

Pearl's structural causal models (SCMs) represent a system as a directed acyclic graph of variables, where each variable is a deterministic function of its parents plus noise. The do-calculus is a set of three rules that specify when an interventional quantity (the effect of setting a variable, written do(X)) can be computed from observational data given the graph structure. The framework is mathematically complete: if an effect is identifiable from the available data and assumed graph, do-calculus will find the formula; if it isn't, do-calculus proves it isn't.

The potential outcomes (Neyman-Rubin) framework represents causation through counterfactual random variables — for each unit, the outcome it would have under treatment and the outcome it would have under control, only one of which is ever observed (the "fundamental problem of causal inference"). It underpins most of biostatistics and clinical trial design.

Both frameworks are well-developed. The difficulty in biology is not the mathematics but the assumptions the mathematics requires:

The graph is usually unknown. SCMs require knowing (or learning) the causal graph. In biology the graph is enormous, incompletely characterised, and full of feedback loops that violate the "acyclic" requirement of standard SCMs (cyclic and dynamic extensions exist but are harder to work with).

Unmeasured confounding is pervasive. Both frameworks lean heavily on the assumption that all common causes are measured. In biology, there are always unmeasured molecular variables, making most observational causal claims fragile.

Biology is nonlinear, redundant, and adaptive. Systems compensate for perturbations (knock out one gene and a paralogue takes over), which means the effect of an intervention depends on the state of the rest of the system in ways that simple causal graphs do not capture well.

So the gap is not that we lack a theory of causation — we have a rigorous one. The gap is that the theory's preconditions (known graph, measured confounders, stable mechanisms) are exactly the things biology denies us. This is why B3 is a genuine scientific problem rather than a matter of applying existing tools.

There is no guarantee that the AlphaFold playbook — large curated dataset plus a clever architecture — transfers to causal disease modelling, because causality is not present in observational data the way 3D structure is present in protein sequences. A breakthrough would more likely come from one of three directions, and we should be honest that we cannot currently predict which (if any) will deliver.

Direction one: perturbation biology at transformative scale. If perturbation methods (Perturb-seq, prime-editing screens, optogenetic and chemogenetic perturbation, base-editing tiling screens) scale by several more orders of magnitude, the resulting interventional datasets might be large enough that causal structure becomes learnable directly. The recognisable milestone would be a model that, trained on perturbation data up to time T, correctly predicts the outcome of perturbations it has never seen — validated prospectively, not just on held-out historical data.

Direction two: hybrid mechanistic-learned models that compose reliably. The recognisable milestone would be a model that combines mechanistic modules (where biology is known) with learned modules (where it isn't) and produces calibrated counterfactual predictions across the boundary — demonstrably extrapolating correctly to intervention regimes outside its training distribution, which pure learned models cannot do.

Direction three: new theory. A genuinely new causal-inference framework better suited to feedback-rich, partially-observed, adaptive systems than current SCM and potential-outcome approaches. This is the least predictable and potentially most important direction. We would recognise it the way the field recognised the original do-calculus: a formalism that makes previously-intractable identification problems tractable.

The honest assessment is that the breakthrough may require all three — more interventional data, better hybrid architectures, and new theory — and that it may take decades or may, for the hardest disease classes, not arrive at all. This irreducible uncertainty is precisely why the Phase II decision gate is structured around B3 viability: the programme must be designed to deliver value whether or not B3 fully yields.

Partial success is the most likely outcome, and the study is designed around it rather than treating it as failure. Partial B3 means reliable causal modelling for some disease classes (those where mechanisms are well-characterised and interventional data is rich) and unreliable or unavailable causal modelling for others (rare, complex, poorly-characterised, or heavily feedback-driven conditions).

Deploying this partial capability safely depends on the system knowing which regime it is in — which puts well-calibrated uncertainty (see A4) at the centre of safe deployment. The architecture would route decisions differently based on causal confidence:

High causal confidence (well-characterised disease, rich interventional evidence, in-distribution patient): the system can propose interventions with counterfactual justification, and — within the safety architecture's bounds (C6) — act with limited human oversight.

Low causal confidence (rare condition, sparse evidence, out-of-distribution patient, heavy feedback effects): the system explicitly abstains from autonomous causal recommendations, falls back to correlational pattern-matching presented as such, and routes the decision to a clinician with the uncertainty made transparent.

This regime-aware deployment means partial B3 is genuinely useful rather than dangerous — the danger comes from a system that applies causal confidence uniformly regardless of whether the underlying causal model is trustworthy for the case at hand. The clinical value covers most of common medicine (where causal mechanisms are well-known) while honestly flagging the long tail of conditions where AIHS remains advisory rather than autonomous.

The Phase II and Phase IV roadmaps both treat this partial-capability scenario as a successful programme outcome, not a fallback — the universal-capability scenario is the upside case, not the baseline expectation.