This paper, recently published by researchers at Purdue University, introduces a novel framework called P2C (Prose-to-Constraints). It tackles a fundamental problem in systems biology: how do we discover the mathematical "laws" governing a biological system when our data is messy, but our textbooks are full of qualitative descriptions?
TL;DR
P2C is an agentic framework that uses Large Language Models (LLMs) to translate scientific prose into formal mathematical constraints, which are then used to guide the discovery of Ordinary Differential Equations (ODEs). By "reading" the mechanism before "fitting" the data, P2C significantly improves the accuracy of model discovery, especially when experimental observations are sparse or corrupted by noise.
What Problem Does This Paper Address?
In fields like physics, data is often "clean" enough for symbolic regression—algorithms that search for the simplest mathematical expression to fit a curve. However, biology is "sloppy." Measurements are often sparse, expensive, and noisy.
The authors point out a critical failure in purely data-driven methods: Equifinality. Multiple, wildly different equations can fit the same noisy data points perfectly. A model might look great on a graph but represent a biological impossibility (e.g., a population that grows to infinity or a concentration that becomes negative).
While we have decades of mechanistic knowledge locked in scientific literature—descriptions of which proteins activate others or where a response saturates—most current AI models for equation discovery treat this text as a "loose prompt" rather than a source of hard mathematical truth.
Methods
The P2C framework operates in five distinct stages:
- Text2Constraint Extraction: An LLM reads a description of a biological pathway and translates it into a set of formal, typed predicates. These include properties like SIGN (does X increase Y?), ABSENT (does X ignore Y?), SAT (does the effect reach a ceiling?), and POS (must the value stay positive?).
- ODE Discovery: A generator LLM proposes candidate ODE structures based on the variables and the general description.
- Symbolic Verification: This is the "gatekeeper" stage. Using the SymPy library, P2C checks if the proposed equations mathematically satisfy the extracted constraints. If an equation violates a constraint (e.g., the text says "A inhibits B" but the equation shows a positive derivative), it is rejected before any fitting happens.
- Parameter Fitting: For the survivors, P2C fits numerical parameters to the data. Crucially, it uses a "weak-form" objective, which integrates the equations to avoid the "noise amplification" that happens when you try to calculate derivatives from jittery data.
- Model Scoring: Candidates are ranked by their trajectory error on held-out validation data.
Datasets
The authors evaluated P2C on two fronts:
- Synthetic Benchmarks: Four established biological ODE systems representing acute inflammation and immune regulation (named after their original authors: Dunster, Baker, Kumar, and Reynolds). These were tested under varying levels of Gaussian noise ($\sigma = 0, 0.05, 0.1$).
- Real-World Application: A dose-response dataset from a study on splenic focused-ultrasound (pFUS) treatment for inflammatory arthritis. This dataset is notoriously difficult because it contains only 50 samples.
Key Insights and Ideas
- The "Noise-Robustness" Threshold: The most striking insight is that textual constraints provide the most value when the data is at its worst. On perfectly clean data, standard regression is often "enough." But as noise increases, the unconstrained baselines fall apart, while P2C remains stable because the "biology" (the text) hasn't changed.
- From Prompting to Translating: Instead of asking an LLM to "write the equation," P2C asks the LLM to "translate the rules." This makes the AI’s contribution verifiable and attributable to specific sentences in the source text.
- Pruning vs. Searching: By using symbolic verification, P2C effectively prunes the search space. It doesn't just find a model that fits the dots; it finds a model that fits the logic of the system.
Key Results
- Error Reduction: On the most challenging noisy data ($\sigma = 0.1$), P2C reduced the "form error" (a measure of how wrong the equation structure is) from 13.096 down to 1.095 compared to unconstrained methods.
- Baseline Comparison: P2C consistently outperformed state-of-the-art LLM-based discovery models like LLM-SR, DoLQ, and EDL across all noise levels.
- Biological Plausibility: In the arthritis application, P2C recovered a bounded intervention-response equation that correctly predicted that ultrasound benefits have a "ceiling." Baselines often produced equations that predicted biologically impossible outcomes, like negative ankle swelling.
Limitations and Future Work
The authors acknowledge several limitations:
- Scale: The current benchmark focuses on relatively small systems (2–4 variables). Scaling to genome-wide networks remains a challenge.
- Trusting the Text: P2C assumes the provided scientific description is accurate. If the literature is conflicting or wrong, the constraints will lead the model astray.
- Constraint Variety: The current schema only captures a subset of possible mechanistic properties.
Future work aims to extend P2C to larger systems and incorporate richer forms of knowledge, such as pathway diagrams or more complex logical relationships.
Reference: Banerjee, P., Rahman, M. M., & Wachs, J. P. (2026). Read Before You Fit: Text-Derived Constraints for Discovering Differential Equations from Noisy Biological Data. arXiv:2611.19779v1 [cs.AI].
Summary updated