Skip to main content
Back to blog
crispr workflow evidence-first CRISPRos

CRISPR Workflow Walkthrough: Evidence-First Design in Action

Jeff Jaureguy ·

Evidence-first isn’t a philosophy — it’s an architecture decision that changes every step of a CRISPR workflow. Here’s what that looks like in practice with CRISPRos.

The Problem with Existing CRISPR Pipelines

A typical CRISPR experiment involves guide RNA design, off-target prediction, cloning strategy selection, sequencing analysis, and results reporting. Most teams stitch together 5–10 tools, copy results between spreadsheets, and write up findings in a Google Doc.

The result: a paper where the reader has no way to verify whether the reported off-target score came from the actual sequencing data or was transcribed incorrectly from a screenshot.

How CRISPRos Handles This

CRISPRos treats a CRISPR experiment as a compilation pipeline. Each stage produces typed artifacts that feed into the next stage. Nothing is generated by the LLM — the AI assists with code and configuration, but every output is deterministic.

Stage 1: Guide RNA Design

You provide a target gene and organism. CRISPRos calls the guide design engine, which wraps established algorithms (Cas-OFFinder, FlashFry) to score candidate guides.

The output isn’t a paragraph of text. It’s a structured artifact:

{
  "target_gene": "TP53",
  "organism": "Homo sapiens",
  "guides": [
    {
      "sequence": "GACUCCAGUGGUAAUCUAC",
      "position": "chr17:7577121",
      "on_target_score": 0.87,
      "off_target_sites": 3,
      "source_algorithm": "FlashFry v2.0.1"
    }
  ],
  "provenance": {
    "genome_build": "GRCh38",
    "timestamp": "2026-03-10T14:32:00Z"
  }
}

Every field is traceable. The source_algorithm tells you exactly which version produced the score. The genome_build tells you which reference was used.

Stage 2: Off-Target Analysis

CRISPRos runs off-target prediction against the genome reference, producing a scored list of potential cut sites. Each site includes:

  • Genomic coordinates
  • Mismatch count and positions
  • Chromatin accessibility score (if ATAC-seq data is provided)
  • Gene annotation (is the off-target in a coding region?)

This isn’t a summary generated by a language model. It’s a deterministic computation with a full provenance chain.

Stage 3: Experiment Report

The final report is compiled, not written. CRISPRos assembles figures, tables, and captions from the artifacts produced in stages 1 and 2. The AI assists with caption phrasing, but every number in the caption is bound to a source artifact.

If you change the input data and recompile, the report updates automatically. If a figure caption says “3 off-target sites detected,” that number is pulled directly from the off-target analysis — it cannot drift from the underlying data.

Why This Matters

The difference between “AI-assisted” and “evidence-first” isn’t about capability. Both approaches can produce a CRISPR experiment report. The difference is in what happens when something goes wrong:

  • AI-assisted: You re-read the generated text and hope you catch the error.
  • Evidence-first: You trace the claim to its source artifact. If the artifact is wrong, the error is visible. If the artifact is correct, the claim is correct.

In CRISPR research, where off-target effects have real consequences for patients, this distinction matters.

Try It

CRISPRos is in early access. We’re working with research groups in rare disease and oncology to validate the workflow.

CRISPRos is a Hordago Labs domain operating system for CRISPR gene editing.