Decision Biology Bridge: ASRA Phase 8 — From Grid Worlds to Perturbation–Response Reasoning

Abstract

Phases 1–7 of the Adaptive Scientific Reasoning Architecture (ASRA) built and hardened a transition-centric cognitive stack for interactive grid environments. From Phase 4 onward, the project drew conceptual parallels to Decision Biology — intervention–response reasoning under latent objectives — but remained game-bound in implementation.

We describe ASRA Phase 8 as the Decision Biology Bridge: an isomorphic extension mapping environment states to cell states, game actions to perturbations, and goal hypotheses to pathway survival objectives. The bridge reuses Phase 1 transition schema, Phase 4 intervention semantics, Phase 5 hypothesis ranking, and Phase 6 experiment sequencing on biological datasets — LINCS L1000, OmniPath, scPerturb, Cell Painting, and Human Cell Atlas context.

The research library lives in asra-arc/src/asra/decision_biology/; the Kaggle agent carries bridge identity without requiring biology data in the sandbox. A Decision Biology demo notebook provides the scientific narrative artifact for Phase 9.

---

1. The architectural gap Phase 8 closes

Phase 1–4   Intervention–response structure (games)
Phase 5     Latent objectives (win conditions)
Phase 6–7   Planning + robustness (games)
Phase 8     Same loop, biological domain
Phase 9     Unified research story

Phases 4–5 established the analogy:

game state  →  action  →  next state  →  progress toward hidden goal

Phase 8 makes it operational:

cell state  →  perturbation  →  next cell state  →  survival / pathway objective

Without Phase 8, ASRA remains an ARC competition project. With Phase 8, it becomes a Nature Foundation Models narrative: one architecture for adaptive reasoning in unknown dynamical systems — whether those systems are grid worlds or living cells.

flowchart LR
  subgraph Game["Phases 1–7"]
    GS[Grid state]
    GA[Game action]
    GH[Goal hypothesis]
  end
  subgraph Bio["Phase 8"]
    CS[Cell state]
    PT[Perturbation]
    PH[Pathway hypothesis]
  end
  GS -.->|isomorphic| CS
  GA -.->|isomorphic| PT
  GH -.->|isomorphic| PH

---

2. Theoretical stance: one transition calculus, two domains

ASRA's core commitment is transition-centric reasoning:

τ = (s, a, s′, r, terminal, metadata)

Phase 8 asserts that τ is domain-agnostic if state and action representations are appropriately defined. The cognitive operations — logging, semantics inference, hypothesis ranking, experiment planning — apply without modification to the form of reasoning.

Operation	Game	Biology
State identification	Grid hash	Gene signature hash
Action	ACTION1–5	Compound, CRISPR, cytokine
Effect	Cell diff	Differential expression
Reward	WIN, level	Viability, pathway activation
Latent objective	Win template	Survival pathway
Experiment	Discriminating action	Next perturbation in screen

This is not claiming that pixels equal genes. It claims that the reasoning loop — observe, intervene, compare, hypothesize, test — is shared between adaptive game play and adaptive experimental design.

Paradigm	Phase 8 stance
Separate bio ML pipeline	Complementary — ASRA adds reasoning layer
End-to-end neural biology FM	Deferred — Phase 8 is symbolic + graph
Hand-wavy analogy only	Rejected — operational schema + demo
Full clinical prediction	Rejected — demo scale with explicit limits
Shared transition schema	Adopted

---

3. Perturbation-as-action semantics

Phase 4 learned that ACTION3 might mean translate in one game and recolor in another. Phase 8 learns that BRD-K12343256 might mean inhibit_pathway for one cell line and no_response for another.

Perturbation semantics (v1):

Label	Biological reading
upregulate_pathway	Activates target pathway genes
inhibit_pathway	Suppresses pathway activity
bypass_node	Compensatory downstream activation
rescue_viability	Restores survival readout
no_response	Expression change below threshold

Semantics are inferred from edge diff summaries on BiologyStateGraph — the same mechanism Phase 4 uses on grid cell diffs.

---

4. Cell-state embedding and identification

Grid worlds use hash(grid) for state identity. Biological states require compact signatures:

1. Gene activity vector (LINCS L1000).
2. Top-k responsive genes → signature.
3. state_hash.py → cell_state_id.

Optional morphology features from Cell Painting enrich the signature — bridging Phase 2's visual object intuition to cellular morphology.

Human Cell Atlas provides context: which neighborhood of cell-state space a sample occupies, grounding perturbation interpretation.

---

5. Pathway hypotheses as biological objectives

Phase 5 goal templates (move_to_target, collect_tokens, …) become pathway survival hypotheses:

PathwayHypothesis(
    pathway_id="MAPK",
    objective="survival",
    preferred_perturbations=["trametinib", "CRISPR_MAP2K1"],
    confidence=0.72,
)

Ranking reuses Phase 5 logic:

score(h) = w_r · response_magnitude(h) + w_p · omnipath_prior(h) - w_c · contradiction(h)

Experiment planning selects the perturbation maximizing discrimination between top-2 pathway hypotheses — the biological analog of Phase 5's goal discrimination experiments.

OmniPath supplies structural priors: hypotheses inconsistent with known signaling topology receive penalties.

---

6. Biological transition graph

BiologyStateGraph reuses ASRA's edge schema from Phase 1 state_graph.py:

• Nodes: cell_state_id, gene activities, pathway context, visit count.
• Edges: perturbation name, replicate count, average reward (viability), num_changed_genes.

The graph is the world model for biological planning — sparse, built from observed experiments, exactly like Phase 6's BFS input.

LINCS ingest → transitions JSONL → BiologyStateGraph → pathway hypotheses

---

7. Datasets and empirical scope

Dataset	Role
LINCS L1000	Core perturbation–response transitions
OmniPath	Signaling graph priors
scPerturb	Single-cell CRISPR before/after
Cell Painting	Morphological state features
HCA	Cell-type context grounding

Phase 8 eval is demo-scale:

• Response direction accuracy (up/down regulation).
• Pathway rank @k on held-out perturbations.
• Schema validation against Phase 1 transition schema.
• OmniPath prior ablation (does graph structure help ranking?).

Explicit non-claims: clinical outcome prediction, full-atlas embedding SOTA, replacement of differential expression pipelines.

---

8. Architecture

Library (asra-arc/src/asra/decision_biology/):

biology_state_graph.py   — Transition graph (reused schema)
state_hash.py            — Cell-state hashing
lincs_loader.py          — LINCS ingest
lincs_experiment.py      — Experiment runner
omnipath_loader.py       — Pathway graph
pathway_simulator.py     — Lightweight dynamics
pathway_hypothesis.py    — Hypothesis ranking
experiment.py            — Perturbation orchestration
cell_embedding.py        — Morphology + expression (planned)

flowchart TB
  LINCS[LINCS L1000] --> Loader[lincs_loader]
  Loader --> Graph[BiologyStateGraph]
  OmniPath[OmniPath] --> PH[pathway_hypothesis]
  Graph --> Sem[Perturbation semantics]
  Sem --> PH
  PH --> Demo[Decision Biology demo]

Kaggle agent (asra-v0.9-phase8):

• Full Phase 7 robust planning stack unchanged.
• Bridge metadata: bio=bridge_active in reasoning string.
• No network calls; no biology data dependency in sandbox.

---

9. Agent integration

Version	Tag	Layer added
Phase 7	asra-v0.85-phase7	Robustness
Phase 8	asra-v0.9-phase8	Biology bridge identity + schema hooks

The competition agent demonstrates architectural continuity; the demo notebook demonstrates scientific extension.

Build:

cd kaggle-notebooks/phase8
python3 build_phase8_kaggle_notebook.py
python3 asra_phase8_my_agent.py --self-test

---

10. Closing the loop with Phases 1–7

Phase	Biology reuse
1	Transition schema, graph structure
4	Effect signatures → perturbation semantics
5	Hypothesis ranking → pathway hypotheses
5	Experiment planner → next perturbation
6	Strategy library → intervention protocols
7	Failure analyzer → non-responder clusters

The module reuse count is a Phase 8 metric: how much game code runs on biology data without modification?

---

11. Position in the Nature Foundation Models program

ASRA Phase 8 is the scientific extension chapter:

ARC Prize 2026  →  proves adaptive reasoning in unknown environments
Decision Biology →  proves same architecture in perturbation–response science
Future FM work   →  neural encoders atop transition-centric scaffold

Phase 9 unifies these into a single research story for submission, GitHub, and deck.

---

12. Open problems

1. Neural cell-state encoders — when signature hashing saturates.
2. Multi-omic integration — joint expression + morphology + chromatin.
3. Causal identifiability — beyond observational LINCS edges.
4. Active learning — closed-loop perturbation selection at scale.
5. Phase 9 narrative — architecture diagram with dual domain.

---

13. Conclusion

ASRA Phase 8 is not a pivot away from games — it is a proof of architectural generality. The same transition log that powers ARC-AGI-3 reasoning powers LINCS perturbation analysis when states become cell signatures and actions become interventions.

The Decision Biology bridge makes the Phase 4–5 analogy executable: pathway hypotheses are goal hypotheses; perturbation screens are experiment plans; biological transition graphs are state graphs.

Transition-centric adaptive reasoning remains the core; Phase 8 shows that core extends to living systems — the scientific mission of the Nature Foundation Models research program.

---

References (conceptual)

• ASRA roadmap — ASRA-roadmap-datasets.md (Phase 8: LINCS, OmniPath, scPerturb, HCA)
• Phase 5 article — goal inference as objective uncertainty
• Phase 8 specification — phase8-decision-biology-bridge.md
• LINCS L1000 — perturbation–response corpus
• OmniPath — pathway structure priors
• Phase 8 implementation — phase8-implementation.md