GLMP's current methodological contribution is layered hybridization: reconciling literature-derived regulatory logic with database-grounded entity completeness, instead of treating them as competing choices. A literature-first flowchart makes a circuit's Boolean logic legible — AND/OR/NOT gates, explicit branch conditions — but tends to leave real biosynthesis steps implicit (a repressor "appears" without a drawn path from gene to protein). A RegulonDB-style wiring diagram has the opposite strength: complete regulatory entities, but no explicit gate logic. The hybrid keeps both — worked out in detail for the E. coli lac operon as a case study, comparing three deliberately different encodings side by side and recommending a four-step workflow (regulatory backbone → interpretive overlay → parts audit → identifier layer) for building one without collapsing distinct abstraction levels into an unreadable graph.
Read the full methods paper: Mermaid Flowcharts for Smarter Perturbation Design.
The flowchart that started it
How this works
-
1
LLMs generate logic-gate flowcharts for regulatory circuits across organisms.
-
2
A qualified molecular biologist validates that the flowcharts accurately reflect observed molecular behavior. This is the pivotal open step. Neither PI is a molecular biologist, so the judgment has to come from outside; a student can attempt a first-pass review, but a biologist's sign-off is what this step requires.
Your contribution here matters.
-
3
Computational methods extract and verify the logic directly from DNA sequences using motif scanning and a custom parser.
-
4
Validated results are written up and scaled to advanced methods including Evo 2 and single-cell regulatory inference.
-
5
At scale, the theory is either confirmed or refined — and the knowledge engine is built on solid ground.
The research arc
Theory → framework → empirical test → synthesis → methods.
-
Primitive Relations, Computational Complexity, and a Conjecture on the Genomic Computational Class
Paper I — foundational typology; the five-class complexity ladder.
-
The Genome as Computer
Paper II — a 22-primitive logical vocabulary; the transcriptome as runtime state; nine falsifiable predictions.
-
Circuit Class Predicts Virtual Cell Model Accuracy
Paper III — 780 genes, K562 Perturb-seq: Class III genes are systematically harder to predict (t = −3.55, p = 0.0015).
-
Genomic Regulatory Complexity and the Limits of Perturbation Prediction
Synthesis — the biology-facing bioRxiv draft.
-
Mermaid Flowcharts for Smarter Perturbation Design
Methods — under active revision. A deposited, DOI-cited version (currently v1.6) is on Zenodo: 10.5281/zenodo.20831780.
Papers and links
Also see
Circuit diagrams for review
Resources
- GitHub
- Annotation Review
- Validation Package
-
Full catalog — 217 processes
The Circuit Class (I–V) labels in this catalog are a first-pass stratification, and parts of it are already known to be wrong. Ongoing expert review is finding these errors case by case; once that review completes, the classification will be rebuilt from what it teaches us, not patched piecemeal. A marked limit is a finding, not a failure — this is exactly the kind of check the project is built to survive.
GLMP in the suite
GLMP is a working engine of the CopernicusAI Knowledge Engine — it consumes the shared Knowledge Base (corpus, embeddings, search) and contributes its decoded circuits back into it.