Acolytics

Inspiration

Modern genomics moves fast: new biological questions demand new sequencing protocols, and every protocol tweak creates new library structures, new metadata conventions, and new failure modes.

When a pilot experiment comes back with issues (e.g., 1.9k cells instead of 45–50k), the bottleneck is not downstream analysis — it is upstream diagnosis. Existing software mostly reports generic metrics and often can’t explain why a run failed, especially for custom assays. General LLM copilots help with code and writing, but they struggle with real-world sequencing context with customization and can’t reliably reason from raw data organization and assay-specific constraints. We built aco to turn biological experiment failures into fast, evidence-driven troubleshooting that accelerates technology development.

What it does

Aco is a genomics-native, agentic bioinformatics platform focused on upstream raw sequencing quality control (QC) and diagnosis. It helps teams:

• Interpret highly customized complex, assay-specific read/library structures (barcodes, UMIs, read geometry)
• Diagnose likely failure modes from messy real-world inputs (folders, FASTQs/BAMs, sample sheets, protocol notes)
• Produce actionable QC summaries and next-step suggestions that support rapid protocol iteration.
• Power a “human-in-the-loop” workflow where results are auditable and usable for decision-making.

How we built it

We built a lightweight end-to-end demo that connects:

• A command line + UI experience for running QC/diagnosis workflows
• A Gemini 3 API-powered agent layer for structured interpretation and troubleshooting logic
• A structured workflow that mirrors how bioinformatics teams actually work:
  • Understand: capture experiment context and assay constraints
  • Analyze: generate a concrete diagnosis plan and run targeted checks
  • Summarize: produce human-readable outputs (figures/tables/report-ready summaries)

Challenges we ran into

• There are many genomic sequencing protocols; general QC assumptions break easily
• Multiple failure reasons can produce similar symptoms (low yield, barcode issues, demux errors, metadata mismatches)
• We needed a clean, privacy-safe recording (no real usernames/paths) while still showing realistic HPC-like workflows

Accomplishments that we're proud of

• Built a compelling end-to-end demo narrative: a failed pilot run -> one-shot prompting falls short -> aco performs structured QC/diagnosis
• Designed a genomics-native framing that focuses on upstream troubleshooting and protocol iteration, not generic downstream analysis
• Established a scalable direction: representing assays as structured schemas so agents can generalize across new data types.
• Created a clear UX concept (CLI + UI) that’s practical for real bioinformatics teams

What we learned

If there is enough prior information in the prompt, the agents perform much better. However, not every assay, especially the one generated from a new technology, has prior accurate analysis. The latest LLMs still struggle in writing de novo coding scripts for processing such kind of data.

What's next for Acolytics

• Build an agentic assay knowledge base that automatically collects protocols into machine-readable library structure schemas and ranks assays by popularity/usage
• More assay-aware QC modules
• We need more systematic benchmarks on real failure cases

Built With

• fastapi
• gemini3
• git
• node.js
• python
• react
• tailwind