Synaptix

Landing page
loading screen
profile
Home page
Alzheimer's Graph
Galantamine Prediction
Analytics dashboard
Research Agents

Inspiration

Drug development costs $2.6 billion and takes 10-15 years. Meanwhile, 7,000+ rare diseases have no treatment at all. But drug repurposing — finding new uses for existing drugs — has already changed medicine. Viagra was a heart drug before it treated ED. Aspirin was a painkiller before it prevented heart attacks. In 2020, BenevolentAI used knowledge graph AI to discover baricitinib could treat COVID-19 in just 3 days — but they had $300M in funding and 200 employees.

Before AI, drug repurposing happened by accident (Viagra took years to notice), by one researcher's hunch (Aspirin for heart attacks took 30 years to validate), or by well-funded AI companies inaccessible to most researchers.

We asked: what if any researcher, anywhere, could have that same power — for free?

What it does

Synaptix is an AI-powered drug repurposing platform. Researchers type a disease — including diseases with no known cure like ALS, Huntington's, and Glioblastoma — and instantly receive ranked drug candidates with biological reasoning paths, plain-English explanations, and AI-generated research briefs.

We systematically analyzed every disease in a 97,238-entity biomedical knowledge graph and identified 1,434 diseases with zero approved treatment. For the top 100, we generated 500+ new drug-disease candidates that didn't exist before Synaptix.

Beyond predictions, Synaptix is a living platform:

Researchers contribute new findings with evidence links
Peer review — agree/disagree/uncertain voting on discoveries
Four AI research agents continuously scan PubMed, ClinicalTrials.gov, and PubChem to validate and expand the knowledge graph in real-time
Context-aware research assistant powered by Gemini that understands the current disease, selected drug, and active pathway

The next drug repurposing discovery shouldn't take 10 years of accidents. It should take 10 seconds of Synaptix.

How we built it

Dual-Signal ML Engine: We loaded the Drug Repurposing Knowledge Graph (DRKG) — 97,238 biological entities and 5.8 million relationships merged from 6 medical databases (DrugBank, Hetionet, GNBR, STRING, IntAct, DGIdb). Using pretrained TransE embeddings, every drug, gene, and disease is a point in 400-dimensional space. We score drug-disease pairs by how closely a drug aligns with a disease through the "treats" relationship vector. All 24,313 drug embeddings are indexed in Actian VectorAI DB with HNSW for sub-millisecond nearest-neighbor retrieval — critical for interactive exploration.

Molecular Similarity: RDKit computes Morgan fingerprints (2048-bit, radius-2) for 8,807 drugs. Candidates are scored by Tanimoto similarity to known treatments — an independent chemical signal fused with the knowledge graph signal (70% KG + 30% molecular). A drug must pass both biological AND chemical tests to rank high. This dual-signal approach is something no existing tool combines.

Explainability: Google Gemini 2.0 Flash powers three layers of explanation — plain-English drug explanations, structured 4-section research briefs (Biological Plausibility, Existing Evidence, Next Steps, Risks) for no-cure diseases, and automated extraction of drug-disease relationships from PubMed abstracts in our AI agents. A context-aware chat assistant lets researchers ask follow-up questions naturally.

Living Knowledge Graph: SQLite-backed collaborative graph where researchers contribute edges with evidence, peer-review discoveries, and four AI agents (PubMed scanner, ClinicalTrials.gov monitor, PubChem bioactivity scanner, Hypothesis Validator) continuously enrich the data. Every new edge makes every future prediction smarter.

Frontend: React 19 with a retro Windows 95 aesthetic featuring a cinematic CRT boot sequence, 3D force-directed knowledge graph with interactive legend filtering and node inspection, and a floating Gemini-powered research assistant.

Deployment: Dockerized FastAPI backend deployed on Vultr Cloud Compute with nginx reverse proxy. 21 diseases pre-cached at startup (~50s) for instant predictions (<0.15s). Live at http://144.202.54.117.

Challenges we ran into

20-minute startup → 50 seconds: DRKG's 5.8M edges created a massive adjacency list (~2GB in memory). BFS path finding initially took 20+ minutes at startup computing paths for every drug-disease pair. We solved this with lazy computation — paths are computed on-demand when a user clicks a drug (~4s each), not at startup. This 24x speedup made the platform usable.

DrugBank IDs aren't human-readable: DRKG uses identifiers like "Compound::DB00331" instead of "Metformin." We built a multi-source name resolution pipeline using DrugBank metadata, PubChem API, and NLM MeSH API to map every compound to names researchers actually recognize.

Explainability vs. speed: TransE scores alone don't help researchers — they need to understand WHY. Building BFS path finding through 5.8M edges, generating Gemini explanations, and creating research briefs for no-cure diseases added significant computational overhead. We solved this with aggressive caching, lazy loading, and background async startup.

Honest constraints: COVID-19 doesn't exist as a disease entity in DRKG (built before the pandemic). Rather than showing broken results, we removed it entirely. Intellectual honesty matters more than feature count.

Accomplishments that we're proud of

All 4 FDA-approved Alzheimer's drugs ranked in the top 4 positions — with zero fine-tuning
84.6% of known repurposed drugs rediscovered in the top 500
1,434 diseases with zero treatment identified across the entire knowledge graph
500+ new drug-disease candidates generated for diseases with no approved cure
Dual-signal scoring (knowledge graph + molecular similarity) that no existing tool combines
A living knowledge graph that grows with every human contribution and AI agent discovery
Four autonomous research agents scanning PubMed, ClinicalTrials.gov, and PubChem in real-time
Full-stack deployment on Vultr Cloud — live and accessible at http://144.202.54.117
Built entirely in 36 hours

What we learned

Knowledge graph embeddings are surprisingly powerful — pretrained TransE vectors with zero fine-tuning rediscovered most known repurposed drugs. The signal is already in the data.
Explainability matters more than accuracy. Researchers don't trust a score — they trust a biological reasoning path they can verify. Every prediction needs a "why."
The real impact isn't well-studied diseases. Predicting drugs for Alzheimer's (which has treatments) validates the model. Predicting candidates for Huntington's Disease (which has none) is where Synaptix actually matters.
A platform that grows is fundamentally more valuable than a static tool. The living graph with peer review and AI agents transforms Synaptix from a prediction engine into a collaborative research infrastructure.
Drug discovery before AI relied on luck. Viagra's repurposing was an accident. Aspirin for heart attacks took 30 years. The fact that a hackathon team can systematically screen 1,434 untreated diseases in under a minute shows how much AI can accelerate this field.

What's next for Synaptix

Clinical validation — partnering with university research labs to experimentally test our top predictions
Doctor-facing features — upload a drug inventory CSV, match drugs to treatable diseases, highlight what's already in stock
Continuous learning — automated re-training of TransE embeddings as the living graph grows with new contributions
Expanded data feeds — real-time ingestion from PubMed, ChEMBL, and UniProt as new research is published
Autonomous lab integration — connecting predictions directly to robotic wet labs for experimental validation
The vision: GitHub for drug discovery — where every researcher, every AI agent, every lab experiment contributes to a living knowledge graph that gets smarter every day. The next cure won't be found by one person. It'll be found by thousands of researchers and AI agents working together on a platform like Synaptix.