MAPK / SCIENTIFIC TOOLS & AGENT SYSTEMS
Same mutation. Different responses.
What can a simplified model explain?
A mechanistic simulation and tool-validation project exploring cell-to-cell variation in ERK suppression.
In progress ? Planning and feasibility investigation, according to the repository. Components below describe the intended system, not completed results.
01 / The question
The mutation is shared.
The modeled cells are not identical.
Under the same KRAS mutation and drug conditions, how much variation in simulated ERK suppression can arise from differences in signaling protein abundance?
Hold conditions fixed
Compare the same mutation and treatment scenario across a shared virtual-cell population.
Vary model inputs
Sample documented protein totals and effective parameters while preserving state constraints.
Compare suppression
Analyze modeled ERK-response distributions and sensitivity to the sampled inputs.
Conceptual illustrations only. No experimental or simulated result is presented above.
02 / Planned architecture
Build the tools first.
Then evaluate the agent using them.
The simulator is designed to run independently of a trained molecular model. Experimental lookup and model predictions will share a source-labeled interface.
RF baseline / GCN comparison
ODE ? virtual-cell population
Budgeted search and evaluation
Planned biochemical IC50 data audit using BindingDB and ChEMBL. Compare fingerprint + Random Forest against GCN on the same frozen scaffold split.
Planned scipy implementation of the Kholodenko2000 simplified cascade, checked against an independently executed reference model.
03 / Validation strategy
Useful outputs need
visible assumptions and failure checks.
Invalid structures or
unsupported predictions.
Numerical failures or
invalid state constraints.
Agent claims beyond
the available evidence.
Parsing rejection and applicability warnings.
Reference trajectories and state-bound checks.
Evidence checks and injected-failure evaluation.
Does the agent add value?
Compare random selection, pIC50 sorting, simple budgeted search, and a ReAct agent. Use exhaustive execution of the same frozen tools as the internal ranking reference.
If simple pIC50 sorting solves the task, report that result. The ranking reference tests tool use within the simulator, not real-world drug efficacy.
04 / Scope & evidence
A research hypothesis.
An engineering project in progress.
The README reports a React frontend scaffold and empty backend implementation files. Data feasibility, reference execution, and tool implementations are planned deliverables.
ERK suppression is a model response, not cell death or clinical benefit. The inhibition mapping is phenomenological, and population assumptions constrain every result.