AI Pharma Paths

Interfaces first; capabilities by phase.

Five paths map common pharma-AI layers, but ISSP requires every AI output to land on a calibrated + UQ model version with auditable artifacts. v1: intervention strategy search only — no molecule generation.

§ 1

Five paths × platform landings

Path	Landing	v1 boundary
Design	intervention_script.yaml	Strategies only; no molecules
Understand	drug_profile.json	Literature/public placeholders; structural pipeline later
Integrate	curve/table digitization	Time-series alignment only; no omics
Discover	small in-house graph	No external large-graph linking
DMTA	experiment_request.json / assay_result.json	Interfaces only; no hardware

§ 2

Design · strategies, not molecules

Search space: timing, dose constraints and combination rules → intervention_script.yaml evaluated in simulation for robustness.

Structure example

intervention_script.yaml (structure example)

scenario: virus_A
objective: minimize T_clear subject_to inflammation_AUC < bound
actions:
  - type: antiviral
    start_day: "<searchable>"
    constraints: ["declared_in_COU"]
model_version_required: "calibrated+UQ"

§ 3

Understand · parameter placeholders

drug_profile.json holds literature/public parameters with source notes; atomic structural pipelines are deferred to keep v1 scoped.

§ 4

Integrate · time-series alignment

v1 multimodal shrinks to curve/table digitization and time-series alignment with mandatory digitization_error_band. No omics commitment.

§ 5

Discover · small graph

A small in-house graph for three scenarios supports mechanism/intervention lookup — not external large-graph entity linking.

§ 6

DMTA · reserved interfaces

Define experiment_request / assay_result shapes only — no lab hardware.

Structure example

experiment_request.json (structure example)

{
  "hypothesis_from_run": "<run_id>",
  "requested_assays": ["<public_assay_name>"],
  "success_criteria_ref": "COU.validation"
}

§ 7

Hard boundaries

AI suggestions are not dosing or treatment advice; never market “optimal strategies” on models without UQ.