HD + AI Landscape
Who's applying AI to Huntington's Disease, what approaches exist, and what makes this hard.
groups Who's Working on AI + HD
| Who | What They Do | AI Type | Stage | Open? |
|---|---|---|---|---|
| Isomorphic Labs | DeepMind spinoff. IsoDDE drug design engine. First AI-designed cancer drug entering Phase 1 | Protein-ligand prediction | Phase 1 | No |
| Aitia + UCB | Causal AI digital twins on CHDI TRACK-HD data (30K patients). Novel HD target identified | Causal AI | Active | No |
| Novartis | Generative AI designed 15M candidate compounds for HD, synthesized 60 | Generative chemistry | Optimization | No |
| Healx | AI drug repurposing for rare diseases. Clinical trials in 24 months vs 10-15 years | Drug repurposing | Active | No |
| Insilico Medicine | End-to-end AI drug design. Rentosertib (first AI drug) validated in Nature Medicine | Generative biology | Clinical | No |
| Unlearn | Digital twins from 30K Enroll-HD patients | ML forecasting | Active | No |
| SOM Biotech | AI-discovered bevantolol (SOM3355) for chorea | Drug repurposing | Phase IIb | No |
| NVIDIA BioNeMo | Open platform for drug discovery. Clara models, $1B co-lab with Eli Lilly | Foundation models | Platform | Open platform |
| IBM + CHDI | ML model of 9 HD disease states | Progression modeling | Published | Partially |
| BDASeq (academic) | AI transcriptomics, 394 druggable HD genes | Transcriptomic AI | Published | Paper only |
| TxGNN (Harvard) | Foundation model for drug repurposing. 17K diseases, 8K therapeutics. Zero-shot | Graph neural network | Published | Yes (MIT) |
| TxGemma (Google) | Open Gemma models for drug discovery. Text + molecular structures | Multimodal LLM | Available | Yes |
| AlphaFold (DeepMind) | HTT protein structure predicted | Protein folding | Available | Yes |
| OpenMed AI | 475+ biomedical NLP models. Disease, chemical, genomic entity extraction. 29.7M downloads | Biomedical NER | Active | Yes (Apache 2.0) |
| HuggingFace ML Intern | Open-source autonomous ML agent. Reads papers, trains models, ships to HF Hub. Used for HD drug target research. | Agentic ML | Active | Yes (MIT) |
| OpenAI Health | GPT-5.2 for clinical care. Deployed at MSK, Stanford Medicine, UCSF | Clinical LLM | Deployed | No |
| HD Research Hub (us) | Autonomous agents, drug repurposing, full-text RAG, somatic expansion drug screen | Agentic AI + RAG | Live | Yes (MIT) |
What we bring to the table
- Open-source autonomous research agents for HD (to our limited understanding)
- Full-text knowledge base with semantic search and multilingual chatbot
- Auto-refreshing research dashboard combining PubMed + ClinicalTrials.gov + HDBuzz
- Published AI experiments with full methodology, limitations, and raw data
- Somatic CAG expansion drug screen across 5 validated targets (MSH3, FAN1, PMS1, MLH1, LIG1)
Most players above keep their models, data, and findings proprietary. We publish everything: code, prompts, raw results, and limitations. That's the point.
Open-source tools available for HD research
These are tools anyone can use today. Part of our mission is highlighting what's accessible.
Harvard. Drug repurposing foundation model. 17K diseases, zero-shot predictions. GitHub
Google. Open Gemma models for drug discovery. Text + molecular + protein. Docs
DeepMind. 200M+ protein structures including HTT. Database
Open platform for biological AI. Clara models, drug design. Platform
475+ biomedical NLP models. Entity extraction for diseases, chemicals, genes. Apache 2.0. Site | HuggingFace
warning Why This Is Hard
Blood-Brain Barrier
Most drugs can't reach the brain. Gene therapies require brain surgery. Small molecules need to cross the BBB. This limits which AI-generated candidates are actually viable.
15-19 Year Timeline
CNS drugs take 15-19 years from discovery to approval on average. Longer than any other drug class. AI can accelerate parts of this, but clinical trials still take years.
98% Failure Rate
For neurodegenerative disease clinical trials, the attrition rate is ~98%. Most candidates that look promising in the lab fail in humans. AI can help identify failures earlier, but can't eliminate them.
Limited Data
HD is a rare disease (~30K patients globally in studies). AI models need data to learn from. Enroll-HD (30K participants) is the largest dataset, but it's tiny compared to what deep learning typically needs.
Disease Complexity
HD involves protein aggregation, transcriptional dysregulation, mitochondrial dysfunction, neuroinflammation, and somatic DNA expansion. No single target addresses all mechanisms. AI must reason across all of them.
AI Hallucination Risk
LLMs can generate plausible-sounding hypotheses that are scientifically wrong. Our Experiment #1 suggested Riluzole without knowing it was already tested. Every AI output needs expert review.
Regulatory Uncertainty
The FDA hasn't established clear guidelines for AI-generated drug candidates. AMT-130 (the most promising HD therapy) is stuck in regulatory debate despite 75% disease slowing in trials.
Commercial Incentive
HD affects ~30K people in the US. Pharma companies prioritize diseases with larger patient populations. This is why open-source and non-commercial research matters.
auto_awesome Where AI Can Make a Difference
Literature Synthesis
Thousands of HD papers published yearly. AI can read them all, find cross-paper patterns, surface connections humans miss. This is what our agents do.
We're doing this now.
Drug Repurposing
Screen thousands of FDA-approved drugs against HD targets. Faster path because safety is already proven. AI can generate and score hypotheses at scale.
We're doing this now.
Clinical Trial Design
Digital twins (Unlearn) can simulate patient controls. AI can identify optimal endpoints, predict enrollment challenges, and stratify patient populations.
Others are leading this.
Molecular Design
Generative AI designing brain-penetrant molecules (Novartis). AlphaFold predicting protein structures. Computational chemistry at a scale impossible manually.
Requires wet-lab validation.
Biomarker Discovery
ML on patient data to find early indicators of disease onset and progression. NfL is established. AI can find new ones from multi-omics data.
Active research area.
Accessibility
Making research understandable to patients, families, and non-specialists. Multilingual chatbots, plain-language summaries, curated learning paths.
We're doing this now.