New Framework Lets AI Systems Build Better Disease Models with Full Audit Trails

When developing computational models to predict how cells respond to genetic or chemical changes, researchers face a web of interconnected decisions about data representation, algorithmic architecture, and training methods. A new framework called CellScientist tackles this complexity by harnessing large language models to propose and refine these models—while maintaining a complete, auditable record of the entire development process.

The Challenge of Automatic Model Design#

Building effective cellular perturbation-response models demands careful coordination across multiple technical layers. Language models can generate plausible candidate designs, but unguided revision risks producing broken implementations, altering the core problem definition, or eliminating components that actually work. The research team created CellScientist to address these risks by wrapping the LLM design process in a formal protocol.

How CellScientist Works#

The framework operates as a constrained workflow that converts feedback from actually running and testing models into structured revision steps. Rather than allowing arbitrary changes, CellScientist:

• Records each design state and its performance outcomes
• Directs identified problems to the specific model components responsible for them
• Applies repairs locally while preserving the overall task definition
• Maintains full traceability from initial proposal through final model

This approach produces two distinct outputs: a validated predictor ready for deployment, plus an inspectable development history showing improvements, setbacks, and failed experiments.

Validation and Performance#

In a controlled study with matched computational budgets, the team compared structured revision under two different language model backbones. The results show that systematic revision outperforms unstructured approaches when computational resources are limited. Additional evaluations tested whether models trained on one patient cohort generalized to an independently collected group, validating external predictive utility.

Transparency Through History#

Operational audits revealed concrete benefits of the auditable approach: the revision history showed fewer redundant proposals, contract verification caught potential violations before implementation, and systematic discrepancy routing enabled targeted repairs. Researchers extended the method to transcriptomic data and single-cell datasets, broadening its application across different types of biological response measurements.