What it does
BioLattice transforms the fragmented "information silos" of biomedical research into a navigable, 3D map of evidence. Instead of providing a flat list of search results, the platform reads through thousands of scientific abstracts to uncover the hidden connections between diet, genetics, and disease. By turning a massive library of unstructured text into an interactive web of relationships, it allows researchers to visualize exactly how a single dietary factor might influence a genetic marker to trigger a specific health outcome.
How I built it
We engineered a "Triple-Engine" pipeline designed to bridge the gap between raw human language and structured computer logic. The process begins with SciBERT, an AI model specifically trained on scientific papers, which acts as the "Reader" to identify key entities and the relationships between them. For high-speed navigation, we implemented FAISS as our "Searcher" to provide semantic context, ensuring the system understands the meaning behind a query even when the terminology differs. Finally, a Neo4j graph database serves as the "Map," storing these connections as nodes and edges to power a 3D frontend that allows users to physically traverse the data landscape.
Challenges I ran into
The most significant hurdle was the "Hairball Problem," where early versions of the live database returned so many connections that the 3D map became a cluttered, unusable mess. To solve this, we had to build custom ranking logic to ensure only the most significant, evidence-backed links were rendered. We also faced a major technical gap in our "Disconnected Ingestion" phase; because the search engine and the graph database were built separately, they initially failed to speak the same language. We had to develop a unified ID mapping system to ensure that a search for a specific protein would accurately light up the correct node in the 3D web.
Accomplishments that I'm proud of
We are particularly proud of successfully moving the project from a mocked-up demo with hardcoded answers to a fully functional, live retrieval engine. Seeing the 3D visualization transform from a monochromatic "unknown" graph into a color-coded, searchable map—where genetic markers are instantly distinguishable from dietary factors—was a major milestone for the team. We managed to create a system that can take 10,000 disconnected data points and organize them into a clear, hierarchical path of evidence that is both scientifically accurate and visually intuitive.
What I learned
Through this process, we learned that metadata is just as vital as the primary data itself; without proper entity classification to tell the system what a node represents, a visualization is merely eye candy rather than a functional tool. We also realized that even the most sophisticated AI extraction is limited if the "plumbing" between databases isn't rock solid. The real magic of the project lies in the connection between the vector search and the graph database, which requires rigorous structural alignment to function effectively.
What's next for BioLattice
The next phase for BioLattice involves the full integration of the SciBERT ingestion pipeline to replace our initial keyword-based scripts, which will provide much higher accuracy for relationship labeling. We also plan to implement a dynamic "Graph Ranking" system based on citation counts so that the thickness of the lines in our map represents the actual scientific consensus found in the literature. Eventually, we aim to expand the frontend functionality to allow users to click any connection and immediately view the specific scientific abstract that proves the link.
Log in or sign up for Devpost to join the conversation.