What Can an LLM Learn from 22 HD Research Papers?
Experiment #1
Experiment Card
git clone https://github.com/jravinder/hd-research-agent && cd hd-research-agent && pip install -r requirements.txt && python src/run_experiment.py
If we feed recent HD research papers to an open-source LLM running on a $1,000 edge device, what patterns does it find?
What We Did
1. Pull Papers
22 papers from PubMed across 5 queries: treatment, AI/ML, somatic expansion, repurposing, biomarkers.
2. Run LLM
Llama 3.1 8B on NVIDIA Jetson AGX Orin 64GB via Ollama. Local inference, zero cloud cost.
3. Extract
Targets, compounds, key findings, and repurposing signals from each abstract.
4. Hypothesize
Which FDA-approved drugs might work for HD?
Results
Top Hypotheses
IL-6 Pathway
Tocilizumab
IL-6 receptor blocker (arthritis drug). May reduce neuroinflammation from mutant huntingtin.
Glutamate
Riluzole
Already tested in HD with modest results. Model didn't know this. AI needs human verification.
TDP-43 / GSK-3β
Lithium
Novel TDP-43/HD connection. Cross-paper insight that's hard for humans to spot quickly.
What Worked
- check_circle
Cross-paper patterns
Found TDP-43/HD connection and post-translational modification targets.
- check_circle
Categorization
Correctly categorized 20/22 papers by research type.
- check_circle
Zero cost
Ran on a Jetson in a home office. Repeatable nightly.
What Didn't
- cancel
No trial history
Suggested Riluzole without knowing it was already tested.
- cancel
Biomarker confusion
Confused NfL (biomarker) with a causal target.
- cancel
2 parse failures
Abstracts too short for the model.
Limitations
- • Not reviewed by HD domain experts or clinicians
- • No experimental validation
- • Abstracts only, not full texts
- • Single model (Llama 3.1 8B), single run
- • Scores reflect model self-assessment, not clinical probability
Methodology
Data
PubMed E-utilities API, 5 queries, 90-day window
Model
Llama 3.1 8B, Ollama, NVIDIA Jetson AGX Orin 64GB
Analysis
Structured JSON extraction per paper, cross-paper synthesis