Project summary

Neurological disease research is slow, expensive, and ethically constrained. Animal models fail to replicate human cognition and human trials are high-risk, limited in scale, and enormously costly. This project is a simulation framework that induces controlled, disease-specific degradation into open-weight LLM weights and activations to mimic how Alzheimer's, Parkinson's, schizophrenia, ALS, Huntington's, frontotemporal dementia, and Lewy body dementia affect the brain. We validate simulated behaviours against real patient datasets, then use mechanistic interpretability to identify and study the affected circuits and attempt to repair them, generating therapeutic hypotheses that can be handed directly to biolabs and pharma partners. The goal is to make LLM simulation a reproducible and ethical alternative testing ground that compresses the drug and therapy discovery cycle without relying solely on human or animal trials.

What are this project's goals? How will you achieve them?

Our core goal is to establish LLM weight degradation as a valid, reproducible proxy for studying neurological disease mechanisms across Alzheimer's, Parkinson's, schizophrenia, ALS, Huntington's, frontotemporal dementia, and Lewy body dementia. We will achieve this by designing biologically grounded degradation protocols for each disorder, running them across open-weight base models, and benchmarking simulated behavioural outputs against real patient datasets including ADNI, PPMI, and UK Biobank cognitive subsets. From there, we apply mechanistic interpretability tools including sparse autoencoders, activation patching, and circuit analysis to identify affected pathways, then run targeted circuit-level interventions to generate testable therapeutic hypotheses. All protocols, code, and datasets will be released openly, and findings will be submitted for peer review, with active outreach to biolabs and pharma partners throughout to ensure the work is grounded in real research needs.

How will this funding be used?

The $30,000 will cover compute costs for model experiments, clinical dataset access fees, and open-access publication and outreach expenses. All outputs will be released publicly.

Who is on your team? What's your track record on similar projects?

Me and my team at Aquin Labs research and build mechanistic interpretability tooling. Our team brings together builders and researchers across ML engs, open-source devs and researchers from Stanford, UCSC, Georgetown University, and neuroscience researchers contributing domain expertise on disease mechanisms and biological validation. Our core platform work, tracing model outputs to exact layers and weights and diagnosing failure modes across training runs, is directly continuous with the degradation and circuit analysis methodology this project requires.

What are the most likely causes and outcomes if this project fails?

The most likely cause of failure is that the behavioural gap between LLM degradation and actual neurological disease pathology proves too wide to produce clinically meaningful signals, making validation against patient datasets inconclusive. A secondary risk is that compute and iteration cycles exceed budget before degradation protocols are sufficiently refined. If the project fails, the most probable outcome is a negative result paper documenting where and why the analogy breaks down, which itself carries research value by clarifying the limits of LLM-based disease simulation and informing future attempts with better-constrained methodology.

How much money have you raised in the last 12 months, and from where?

$80,000 in grants from Emergent Ventures, Founders Inc, The Residency and Startup Programs in compute and cash for Aquin.app