EVIDENCE GRAPH
Governed multi-model engineering
Design protein function with an evidence chain you can inspect.
Define a design objective, orchestrate generation and sequence engines, challenge designs with independent structure models, optimize developability, then select on explicit trade-offs with a permanent Design Passport.
01Campaign-nativeIntent and constraints persist
02Multi-modelGeneration and validation separated
03Human-reviewedNo silent candidate approval
04Passport-backedMethods and provenance preserved
Evidence classes remain distinct throughout the campaign.
ACTIVE CAMPAIGNNo campaign selectedSelect a campaign to begin.
CAMPAIGN COMMAND CENTER
Programme state at a glance
SCIENTIFIC INTENT
Objective & constraints
REVIEW STATE
Human decisions remain explicit
COMPUTE STATE
Jobs started from this workspace
DESIGN EXPLORER
Search, filter and compare design alternatives.
INTERACTIVE PARETO VIEW
Explore trade-offs without collapsing them into one score.
Pareto frontOther designSelected designOnly persisted comparable metrics are plotted.
SEQUENCE INTELLIGENCE
Select a design to inspect its sequence.
MUTATION MAP
—Parent → design differences
LIABILITY SUMMARY
Deterministic sequence-only signals
ANNOTATED SEQUENCE
Residue-level inspection
N-glycodeamidationAsp isomer.hydrophobiclow complexitymutation
STRUCTURE WORKSPACE
3D context for the selected design.
No structure available
The viewer never substitutes a predicted fold for a missing persisted structure.
Select a design with a persisted structure URI or use a campaign target structure.
EVIDENCE LEDGER
Keep measured, predicted, computed and reviewed evidence distinct.
METRIC MATRIX
Select a design
ADD EVIDENCE
Persist a metric with its evidence class.
COMPUTE ORCHESTRATION
Submit only to configured engines, then preserve the job record.
NEW JOB
Explicit execution contract
JOB TIMELINE
Queued → running → completed / failed
SCIENTIFIC RUNTIME
Endpoint health, execution readiness and qualification stay separate.
01 · THE ENGINEERING ARC
One traceable path from scientific intent to selection.
Every transition preserves input hashes, model/version settings, evidence class, applicability limits and review state. The workflow adapts to the question without losing the provenance chain.
02 · DEFINE THE CAMPAIGN
Begin with the engineering objective—not the model.
Choose the scientific question BayesARC must solve. Engine selection happens only after target context, constraints and campaign intent are explicit.
CAMPAIGN TEMPLATES
1
Question firstMode selection frames what “better” means.
2
Context attachedTarget, structure and cofactors travel with the record.
3
Trade-offs explicitSelection is multi-objective, not a hidden composite.
No silent execution
Creating a campaign does not submit sequence or structure data to an external engine. Each compute job requires separate confirmation.
03 · DESIGN MODES
Twelve engineering questions. One governed campaign model.
Binder, enzyme, antibody and sequence optimization workflows retain their specific inputs and validation requirements instead of being flattened into generic generation.
MODE PRINCIPLE
Different questions deserve different evidence contracts.
Each mode defines required inputs, intended outcomes and downstream checks. The model is an implementation detail; the scientific question is the durable object.
Inputs required Outcomes recorded Review remains human
04 · DESIGN RECORDS
Register a design. Separate deterministic QC from predicted biology.
Sequence-only screening is labelled computed evidence. Structure, affinity and stability predictions retain their originating model, version and applicability context.
ACTIVE CAMPAIGN
No campaign selected
No design campaign is active
Create or select a campaign to register protein designs and begin the evidence record.
COMPUTED SEQUENCE QC
Awaiting a design
No sequence-derived values yet
No values are invented before a sequence has been submitted and analysed.
05 · MULTI-OBJECTIVE SELECTION
Expose the trade-offs. Never bury them in one score.
Binding, stability, solubility, specificity, fold confidence and novelty are maximized; immunogenicity, aggregation and sequence liabilities are minimized.
Binding ↑Stability ↑Solubility ↑Specificity ↑Fold confidence ↑Immunogenicity ↓Aggregation ↓Liabilities ↓Novelty ↑
Awaiting campaign designs
Select a campaign to load persisted designs and its Pareto frontier.
06 · SCIENTIFIC RUNTIME
Catalogued ≠ configured ≠ qualified.
Render is the campaign control plane. GPU models remain isolated behind service endpoints and cannot appear executable until a real runtime is configured.
Qualifiedpassed context gate
Configuredendpoint available
Catalogued onlyknown, not executable
!
Runtime truth is a product feature. A catalogued model is not shown as runnable, and a configured endpoint is not shown as scientifically qualified until the applicable qualification gate has passed.
EVIDENCEPredicted is not measured.
Every output carries an explicit evidence class and originating method.
QUALIFICATIONAvailability is not applicability.
Task-specific benchmarking and context-of-use review control promotion.
AUTHORITYQualified humans select.
BayesARC preserves the review record; it does not autonomously approve a candidate.
FROM SEQUENCE TO DECISION
Define a campaign↑