logo

Irendil: Translating the microbiome into the language of the human genome.

💡 Inspiration

In modern medicine, we are standing on the edge of a data cliff. Doctors like Dr. Nonclercq can see that a patient has a "healthy" microbiome based on bacterial abundance, but they are often left in the dark regarding how that patient’s specific DNA will actually respond to those bacteria.

We were inspired by a simple, frustrating truth: Two people can have the exact same gut bacteria but experience completely different health outcomes. We realized that the "missing link" was the regulatory wiring of the human genome. We built Irendil to be the bridge—an agentic platform that finally allows the microbiome to speak to the human genome.

🛠️ How We Built It

Irendil is a high-performance orchestration of cutting-edge AI and genomic modeling:

The Genomic Core: We utilized AlphaGenome, DeepMind’s 2026 sequence-to-function model. By feeding it 1MB windows of human DNA, we could predict the "Active Potential" of regulatory switches (promoters and enhancers).
The Microbiome Engine: We built a FastAPI backend that processes bacterial abundance data, mapping specific species like F. prausnitzii to their known metabolic pathways (e.g., Butyrate production).
The Agentic Interpreter: We used LangChain and LangSmith to orchestrate a Gemini 2.0 Flash agent. This agent performs the critical "translation" from raw genomic scores to clinical insights.

The core logic relies on converting $log_2$ fold-change predictions into linear, actionable data for physicians:

$$\text{Functional Reaction Strength} = 2^{\text{AlphaGenome Score}}$$

🧠 What We Learned

This project was a deep dive into the "Dark Genome." We learned that:

Context is Everything: A 10bp switch is useless without the 1MB of 3D-looping context that AlphaGenome provides to understand long-range enhancer interactions.
LLMs as Biologists: When properly constrained with Structured Outputs and strict system prompts, LLMs like Gemini are incredible at symbolic math and metabolic mapping.
Personalization is Non-Linear: Small SNPs (Single Nucleotide Polymorphisms) in a promoter region can lead to exponential differences in how a patient responds to a probiotic.

🚧 Challenges We Faced

The Data Bottleneck: Human genomic data is massive. We had to implement a "pre-loading" strategy for the 1MB DNA sequences to ensure the live demo remained snappy and responsive.
The JSON Struggle: During the hackathon, we fought with the "chatty" nature of LLMs. We eventually solved this by using Gemini’s Native JSON Mode and LangSmith to enforce strict schemas, ensuring our frontend didn't break during analysis.
Prompt Precision: Tuning the "Interpreter Agent" to correctly calculate $2^x$ without hallucinating required multiple iterations and testing within the LangSmith Playground.

🚀 The Impact

Irendil transforms a doctor from a "pattern seeker" into a "precision engineer." By providing a Holistic Summary and Clinical Insights at the genomic level, we empower surgeons and GPs to provide hyper-personalized care. Instead of guessing if a diet will work, Dr. Nonclercq can now see the "wiring" that makes a patient a super-responder or a non-responder.