Read the biology before fitting: P2C cuts worst-form error from 13.096 to 1.095 as noise grows
Synopsis
Recovering ordinary differential equations from sparse, noisy biological observations tends to select structures that fit the data equally well while encoding different mechanisms. P2C first turns mechanistic prose from the literature into typed constraints tied to source spans, then rejects candidate equations that violate them before any numerical fitting. Across four published inflammation and immune-regulation models, adding the text constraints reduced the worst form error from 1.397 to 0.499 on clean data and from 13.096 to 1.095 under Gaussian noise perturbation.
Interpretation
Equation discovery gains a structural channel from scientific text: biological mechanisms are often communicated in natural language and pathway diagrams rather than complete mathematics, and P2C converts those descriptions into hard restrictions on candidate ODE structures before numerical fitting. Earlier LLM-based equation-discovery methods are largely data-driven and assume observations are clean and representative enough to identify the dynamics; in biology, where measurements are sparse, noisy, heterogeneous and expensive, multiple equations can fit the same observations while encoding different mechanisms. On four published ODE systems of acute inflammation and immune regulation, progressively degrading the observations reduced the worst form error from 1.397 to 0.499 on clean data and from 13.096 to 1.095 under Gaussian noise perturbation when the text constraints were added; over the tested noise range, P2C also had lower form error and trajectory error than several LLM baselines.
The pipeline separates what is inferred from language from what is inferred from data: prose is first translated into a typed set of predicates covering interaction sign, saturation, structural absence, joint interaction, source terms and positivity, and candidate equations must pass symbolic checks before parameter fitting. Each constraint is grounded to a source span and passes entailment checking, consistency across independent extraction runs, and logical-consistency filtering; the LLM that proposes candidate equations is blinded to the numerical observations and the constraint set, seeing only the system description, state variables and feedback on violated constraints. Verification uses SymPy to differentiate, take limits and substitute symbolically; when a partial-derivative sign cannot be established symbolically, it is evaluated at 400 random points with positive states and parameters and accepted only if the sign is uniform, which is numerical evidence rather than proof. Parameter fitting uses a weak-form objective that moves differentiation from the noisy trajectory onto smooth test functions instead of differentiating observations directly.
Perspective
The demonstrated scope is four published ODE systems of inflammation and immune regulation, each paired with a manually written pathway description and sentence-level reference constraints, with observations generated by numerical integration and subjected to Gaussian noise of varying strength; a separate demonstration uses measured dose-response data from splenic focused-ultrasound treatment of inflammatory arthritis, 50 samples in total, with the 20-minute group held out. The authors propose extending P2C to larger biological systems and richer forms of mechanistic knowledge.
The constraint schema captures only a restricted subset of mechanistic properties and filters rather than uniquely determines equation structures; the method assumes the supplied mechanistic description is a reasonably coherent summary of the relevant knowledge, and robustness to conflicting literature, ambiguous or unsupported statements is not evaluated, since the grounding checks verify consistency with the provided description rather than its scientific correctness. The authors also note the controlled benchmark tests whether text constraints improve discovery under known mechanisms and does not establish generalization across the broader literature, substantially larger dynamical systems or other domains.
