Autonomous Labs Project

In progress
Home / Projects / Autonomous Labs Project

Autonomous Labs Project

Working Thesis

DisMech plus OpenScientist can become an auditable experiment-suggestion layer for disease biology: DisMech stores a computable pathograph, OpenScientist searches and ranks mechanistic gaps, and a protocol layer turns selected gaps into standardized experiments that can be reviewed by humans and executed by university automation cores or cloud labs.

This is not a claim that AI should autonomously invent therapies. The useful near-term claim is narrower: AI can maintain a disease-mechanism graph, identify weak causal edges, propose bounded experiments, and package those experiments into reproducible, machine-readable protocols.

UNC Lineberger Anchor

The UNC Lineberger "Priming the pump for new cancer treatments" story is a good case study because the named investigators cover the whole loop:

Researcher DisMech/OpenScientist role Automation relevance
Ian Davis Disease-mechanism graph for Ewing sarcoma, especially EWS-FLI1 chromatin rewiring Defines causal nodes, models, and chromatin readouts
Samantha Pattenden Chromatin assay development and screening HT-FAIRE converts chromatin accessibility into an automated plate assay
David Drewry Chemical probe and medicinal chemistry infrastructure Supplies annotated compound libraries and probe optimization logic
Pengda Liu Protein-modification and targeted-degradation biology Provides degrader-style perturbations such as TF-PROTAC concepts
Lindsey James Chemical biology of chromatin regulators and degraders Shows how degrader discovery can be tied to selectivity and phenotype assays

Ewing Sarcoma Demonstrator

The strongest disease-specific demonstration is an Ewing chromatin-accessibility loop:

  1. DisMech encodes the Ewing pathograph: EWS-FLI1 fusion, GGAA enhancer reprogramming, ETV6 counter-regulation, NuRD/CHD4 repression, core regulatory circuitry, replication stress, and STAG2 modifiers.
  2. OpenScientist identifies a concrete gap: which EWS-FLI1-dependent chromatin states are causal dependencies rather than passenger accessibility changes?
  3. The platform proposes a standardized experiment: automated HT-FAIRE, ATAC-qPCR, or low-input ATAC-seq across Ewing models, perturbing epigenetic compounds, degraders, EWS-FLI1 controls, and ETV6 controls.
  4. A protocol compiler emits a reviewed protocol compatible with laboratory automation standards or a university cloud-lab/core facility.
  5. Results return as evidence: hit compounds, chromatin-state changes, transcriptional rescue, toxicity separation, and updated pathograph edges.

The corresponding curation target in kb/disorders/Ewing_Sarcoma.yaml is gap_ewing_chromatin_reversal_screen.

Protocol And Execution Layer

Candidate execution standards and systems:

Commercial Cloud-Lab Catalogues: What Is Actually Orderable

The execution layer above lists candidate standards. This section records what a specific commercial catalogue turned out to contain, because the answer changed which demonstrator is reachable.

Ginkgo Cloud Lab (https://cloud.ginkgo.bio/protocols, read 2026-10-01) publishes 18 protocols. They are almost entirely protein biochemistry: cell-free expression validation and optimisation, HiBiT and A280 quantification, Strep-tag and His-tag purification, LabChip purity, a SYPRO Orange thermal shift assay, Echo-MS detection of an enzymatic reaction, minibinder expression with SPR target onboarding, IVT mRNA/circRNA synthesis, E. coli and Pichia expression, and assay-onboarding services. Input is typically DNA sequences submitted as a CSV.

Commercial terms — price, turnaround, throughput — are deliberately recorded nowhere in this repository. They change with no signal that they have, and this project is not a reseller's catalogue; a curator settles them with the provider when something is actually ordered.

The Ewing chromatin demonstrator above is not executable on this catalogue. Nothing in it does chromatin accessibility, cell-line culture, or a sequencing readout, so HT-FAIRE, ATAC-qPCR and low-input ATAC-seq are all out of scope; that loop needs the UNC automation core it was designed around. This is not a defect in the demonstrator — it is a statement about which rung of the execution layer each catalogue serves. A commercial cloud lab of this shape reaches protein-level mechanism questions: variant functional impact, expression, folding, binding, and enzyme turnover.

So the first entry curated against this catalogue is a protein-biochemistry gap rather than the Ewing one: Prolidase_Deficiency → pd_allele_panel_abundance_stability_catalysis (see Experiment.executable_protocols, and the CLAUDE.md section of the same name). Clinical practice reports one "residual prolidase activity" figure per patient, measured in erythrocyte or fibroblast lysate, and the entry's open gap records that severity is uncoupled from it. Three catalogue protocols decompose that figure into abundance (HiBiT), folding (thermal shift) and catalysis (Echo-MS on an imidodipeptide substrate), which is a decomposition the clinical assay cannot make. That amount-versus-activity distinction is the axis deferred in design decisions §12, which names the residual-activity genotype-severity question among its motivating cases.

Open questions before this is a closed loop:

Pre-Execution Verification

ExecutableProtocol records that a run is possible. It deliberately says nothing about whether a particular run should proceed, which is the gate the section below is about. The candidate mechanism for that gate is clauz3 with its autolabs domain library: an agent-authored plan calls trusted side-effecting functions, carries @clauz3.guarantee(...) decorators, and clauz3 prove discharges those guarantees with Z3 before anything executes. Its existing lims domain already carries the contract shapes this project needs — instrument allowlists, cumulative budget bounds, and a no_hazard_sequence biosecurity contract on the oligo-ordering flow, which is exactly the step a variant-panel experiment reaches when its DNA templates are ordered.

That work is not in this repository and is not a dependency of the schema block: a cloud-lab contract domain would be a change to clauz3-tools-autolabs, scoped there and provable against mock effects with no provider access at all. Recorded here so the division is explicit — dismech states what a gap would take; the verifier decides whether a plan honouring that is allowed to run; neither is the human review this project requires before either.

Safety And Governance

The platform should keep human review as a hard gate before execution. Required checks include PI approval, institutional biosafety review where relevant, model-system and reagent provenance, protocol versioning, dose/exposure bounds, biosecurity screening, data-management plans, and curator review before any new result updates the knowledge base.

Near-Term KB Ideas

Sources