GLMP — From Square One

We call DNA "the code of life," but usually just as a metaphor. GLMP takes the metaphor seriously and asks a concrete question: can the logic of a gene — the rules that decide when it turns on and off — be read directly from the DNA sequence, the way you'd read a few lines of a program?

Genes don't fire at random. A stretch of regulatory DNA behaves like a tiny circuit: make this protein when a nutrient is present AND a second signal is absent. Those are real logic gates — AND, OR, NOT — built out of molecules. GLMP reads those gates straight from the genome, draws each gene circuit as a flowchart, and then checks that reading against what biologists have actually established.

This is the inverse of synthetic biology. Work like Voigt et al. (Science 2016) compiles human-written logic into new DNA; GLMP reads the logic evolution has already written into natural regulatory DNA. Same grammar, opposite direction.

Why it matters

The logic of gene regulation is how a cell makes decisions — how one fertilized egg becomes a nerve cell or a muscle cell, and how those decisions go wrong in disease. If we can reliably read and verify that logic at scale, we build a foundation for understanding how the instructions inside our cells work, and eventually for seeing where they break. It sits exactly where computer science meets biology and medicine.

Where it stands

This is a working research program, not a proposal. It already draws logic-flowchart models for hundreds of gene circuits across organisms, runs a computational decoder that reads regulatory logic directly from DNA sequence, and is backed by a growing research knowledge base. Today the decoder reliably reads the most fundamental piece of that logic — the "off switch," where a protein blocks a gene — in bacteria, checked against gold-standard reference data. It is deliberately careful: it claims only what the sequence actually supports, and treats a clearly marked limit as a finding rather than something to hide. The frontier from here — reading the "on switches," handling more complex logic, and extending from bacteria toward human cells — is wide open.

Where you come in

The project needs exactly the mix of skills in this room, and there is real work to own. There are two tracks:

• Computational validation — check the decoder's readings against curated biological databases. A concrete, self-contained job that turns programming skill into scientific evidence.
• Biological judgment — connect what the code says to what molecular biology knows. Neither current lead is a molecular biologist (Gary is a logician, Konstantinos a bioinformatician), so there is a genuine open seat for biological expertise.

The detailed asks are at the end of this briefing. Everything in between is how the project actually works.

The project began with a picture: a molecular regulatory process drawn as a logical circuit, before any decoder or database. The anchor example is the lac operon in E. coli — one of the most studied systems in molecular biology, and the worked example for everything that follows.

The purple diamonds are logic gates; the orange node is the source of the repressor protein; the green node is the output. The whole regulatory logic of the lac operon is captured in one diagram — and this is what GLMP formalizes at scale. Note the flowchart is the biology: the cell really does compute "operator free AND CAP assist." What the decoder can read from sequence alone is a separate question, and an honest one — see "Where the decoder stands," below.

Steps 3 through 5 are contingent on step 2 being answered in the affirmative. That is not a weakness — it is how science works.

1. 1
   Have LLMs generate flowcharts of molecular processes We use large language models to produce Mermaid logic-gate flowcharts for regulatory circuits across multiple organisms and disciplines. Each flowchart assigns gate types (AND, OR, NOT) to transcription factor binding relationships and classifies the circuit by complexity class (I–V). We currently have 217 processes in a structured catalog and are actively scaling toward a target of 1,000+ circuits. Batch flowchart generation runs automatically overnight on dedicated compute hardware, adding new circuits continuously. In parallel, a curated collection of embedded research papers is being built to ground each circuit in the primary literature.

   Status: active and scaling — 217 circuits today, targeting 1,000+.
2. 2
   Have a qualified molecular biologist validate the flowcharts A subject-matter expert reviews the flowcharts and annotations against the primary literature, confirming or flagging each logic-gate assignment. This is the pivotal open step — and an honest limitation of the team as constituted: neither Gary (a logician) nor Prof. Krampis (a bioinformatician) is a molecular biologist, so the definitive biological judgment must come from outside. A student on the biology track can and should attempt this review as a learning exercise; their draft is then taken to a molecular biologist for sign-off. Everything in steps 3–5 rests on this being answered affirmatively.

   ⚑ Pivotal step — this is the open question. Your contribution here is the foundation everything else rests on.
3. 3
   Apply computer-assisted methods to extract the circuit logic from DNA Assuming the flowcharts are valid representations of real biology, we undertake computational methods for verifying, extracting, and teasing out the logical structure directly from DNA sequences — using FIMO motif scanning with validated custom prokaryotic PWMs for E. coli (JASPAR retained for yeast and phage only), and a logic parser that classifies circuits by their DNA-level topology. This is what Gary, Krampis, and our AI collaborators are engaged in now.

   Status: active — pipeline built, first circuits decoded.
4. 4
   Publish results and scale to advanced validation methods If step 2 confirms the flowcharts are biologically faithful, and step 3 produces credible and publishable results, we will write up our findings and advance to higher-level validation — including cross-validation against foundation models trained on genomic sequences (such as Evo 2 from the Arc Institute) and integration with single-cell regulatory inference tools such as RegVelo.

   Status: methods paper in preparation.
5. 5
   If accomplished at scale — proceed to the Big Picture Goals If steps 1 through 4 are completed at scale across hundreds or thousands of regulatory circuits, we can consider the core theory plausibly correct and proceed with confidence to the Big Picture Goals — which we will describe in a future conversation once the foundation is established.

The flowcharts don't just record which logic gates a circuit uses; they place each circuit on a ladder of increasing regulatory complexity. The intent is a five-class scheme, from the simplest computable logic to the most open-ended:

• Class I — Feed-forward logic. Input flows to output through logic gates (NOT, AND, OR) with no loops. Fully determined by its inputs, and in principle fully predictable. Most classic bacterial operons live here.
• Class II — Self-regulation / negative feedback. A damping loop that holds a system near a set point — homeostasis and noise buffering.
• Class III — Bistable switches. Positive feedback or mutual repression creates two stable states; the circuit stays flipped after the signal is gone. The basis of cell-fate decisions.
• Class IV — Oscillators. Feedback with a built-in delay produces sustained rhythms rather than a fixed state — circadian clocks are the classic case.
• Class V — Self-modifying circuits. The circuit rewrites its own regulatory architecture, for instance through chromatin change. The most expressive tier, and the least predictable.

We are direct about the status of this scheme: defining the class boundaries precisely — and settling how the logic we read from sequence relates to the biological complexity of the whole circuit — is active work, and one of the near-term tasks for this collaboration. That is why the table below carries two separate columns, and why some entries read "I/II": they mark circuits whose classification sits on a boundary we are still drawing.

The lac operon is the sharp case. At the sequence level it reads as Class I feed-forward logic — a repressor (NOT) and, when present, an activator. Yet biologically it runs dual control: activation by CAP together with de-repression of LacI. Whether dual control of that kind should define its own place on the ladder — and how to decide such questions in general — is not a detail we mean to smooth over. It is close to the central research question, and we would rather state our tentative reads and work them out than reach for the tidier label. The full formal treatment is in Papers I–III (see "Papers and further reading," below); the practical outcome is what this collaboration is here to help establish.

The GLMP DNA Decoder uses FIMO with validated custom prokaryotic PWMs for E. coli (JASPAR for yeast/phage) to identify transcription factor binding sites in promoter sequences, then infers logical relationships from their spatial arrangement. It has been tested on five E. coli regulatory circuits and one yeast circuit as an exploratory first eukaryotic attempt.

The decoder reads repression (NOT) reliably but cannot yet confirm a cooperative activation (AND) from sequence alone; for single-activator promoters like lac, spacing-based AND detection is structurally unreliable, so the honest DNA-topology label is I/II. Biological class II — CAP-dependent activation plus repression, the logic shown in the flowchart above — remains the biological ground truth the validation team is asked to confirm. Across all circuits the decoder currently makes zero AND calls, and says so rather than inventing one.
* The lac operon's logic class is the specific open question this collaboration is asked to weigh in on. Our decoder returns I/II — it detects LacI's NOT gate confidently but does not find evidence for a combinatorial AND, because CRP and lacO1 are ~70 bp apart, outside cooperative spacing. Whether that arrangement nonetheless constitutes AND logic is a biological judgment we are seeking.
† AraC is absent from JASPAR (a prokaryote-specific TF); a custom matrix is in development.

Two tracks, matched to background:

Both tracks work from a single self-contained package — data files, decode outputs, RegulonDB reference data, a per-track task brief, and a shared report template — available for download with no credentials or database access required:

• Computation task brief: task-brief-computation.md
• Biology task materials: lac operon annotation review
• Full validation package: README

Contributions will be acknowledged in the methods paper currently in preparation. Co-authorship is on the table depending on the significance of the contribution.

An external molecular biologist's judgment is required here, not optional: because neither Gary nor Prof. Krampis is a molecular biologist, the definitive call on each logic-gate assignment comes from outside the two of us — potentially sourced through the John Jay–Hunter network. A student's draft review is a valuable first pass and a learning exercise, complementary to that outside sign-off, not a substitute for it.

Start here (written for biologists)

Synthesis paper — the framework brought together in one accessible overview.

Methods (published)

Methods paper — LLM-generated Mermaid flowcharts, curated databases, and the lac operon as a worked example. Zenodo: 10.5281/zenodo.20831780.

Full technical papers (reference only)

Paper I — Foundational typology — the typed-logic foundations and the circuit-class typology.

Paper II — Genome as computer — Boolean primitives mapped to molecular mechanisms, and the five-class complexity ladder.

Paper III — Empirical sequel — applying the framework to real circuits and data.

Background (accessible)

From Inspiration to AI: Biology as Visual Programming — Medium

Is the Genome Like a Computer Program? — ResearchGate

Gary Welz — Google Scholar profile