Skip to main content
EasternDawnPortfolioPio

Liability, evidence, design, review, experiment

Scientific Agent · Peptide optimization

PioTurn one peptide into a reviewable optimization decision

Pio connects task validation, liability scanning, evidence retrieval, candidate generation, separate activity/stability/developability review, explicit ranking, and next-experiment planning while preserving provenance and uncertainty.

Read the method Inspect evidence
PyMOL parent conformer of AALKAAAE with lysine 4 and glutamate 8 proposed bridge anchors highlighted
Illustrative case / conformation lock

A proposed bridge begins as an inspectable hypothesis

The parent PyMOL conformer highlights K4 and E8 as proposed anchors for a lactam bridge while keeping the unmodified molecular object visible.

Evidence boundary. The bridge is absent from this conformer and its structural or biological effect is unvalidated.
Design object
One peptide plus context
Primary objective
Stability-aware optimization
Review lanes
Activity · stability · developability
Evidence state
Typed and reviewable

01 / Abstract and project definition

Evidence-Aware Peptide Optimization

Pio is an evidence-aware scientific agent for peptide optimization. It begins with one sequence, the serum or protease environment, the desired stability outcome, protected motifs, chemistry limits, and assay context. Downstream work waits until this task contract is complete enough to review.

The agent then maps liabilities, retrieves analogue and literature support, generates constraint-aware edits in a stable modification language, and reviews each proposal through separate activity-preservation, stability, and developability scorecards. Ranking keeps hard failures, observed support, inferred support, introduced risk, uncertainty, and the next discriminating experiment visible.

02 / Scientific problem

A plausible modification can improve one liability while introducing a new activity or developability risk.

Peptide optimization combines sequence rules, assay context, analogue evidence, chemical feasibility, and uncertain biological mechanisms. Generic suggestions are difficult to audit when protected motifs, evidence locators, reviewer disagreement, and experiment design disappear into prose. Pio uses typed runtime objects and narrow review contracts to preserve these boundaries.

01 / Constraint

Protect functional context

Required motifs, maximum edit count, allowed chemistry, forbidden transformations, and assay frame are validated before design.

02 / Evidence

Separate observation and inference

Source-linked support, mechanistic rationale, retrieval gaps, and confidence remain distinct fields.

03 / Decision

Expose benefit and risk together

Every candidate states its intended benefit, introduced risk, review disagreements, and next experiment.

03 / Method overview

A thin harness governs routing and state while registered scientific skills own bounded domain work.

The harness validates the task, selects a route, controls stage transitions, persists inspectable objects, dispatches registered skills, prunes weak candidates, ranks survivors, and builds the final report. Ten scientific capabilities cover intake, liability mapping, retrieval, evidence packing, modification strategy, generation, three reviews, and delivery.

01

Task contract and intent router

Normalize the peptide, environment, objective, constraints, modifications, and assay fields before selecting a workflow route.

Output · Validated task card
02

Liability and evidence lanes

Run sequence-level checks and retrieval lanes, then normalize results into provenance-bearing evidence objects.

Output · Liability map and evidence packet
03

Design and multi-review

Translate liabilities into allowed strategies, generate candidate edits, and score activity, stability, and developability separately.

Output · Reviewed candidates
04

Prune, rank, and report

Demote hard failures, apply visible weighted dimensions, preserve uncertainty, and connect each shortlist entry to a discriminating experiment.

Output · Decision report

04 / Architecture

Every handoff has an explicit scientific contract

Inputs, operations, evidence, and outputs stay named so the source of each decision can be reviewed.

  1. Control

    Task and state machine

    Validated task objects determine allowed stage transitions and resume context.

  2. Runtime

    Registered skill contracts

    Each skill declares triggers, input and output schemas, callable tools, memory writes, eval hooks, and fallback behavior.

  3. Review

    Evaluator–optimizer loop

    Three scorecards preserve disagreement, prune blocked candidates, and focus redesign on a bounded survivor pool.

  4. Memory

    Four governed classes

    Working, evidence, user, and learning memory follow separate write and retention policies.

  5. Control surface

    Trace, guardrails, and evals

    Stage events, decisions, rejections, provenance, regression checks, and acceptance criteria remain inspectable.

05 / End-to-end workflow

From scientific input to an evidence-bearing decision

Each stage consumes a bounded object and leaves a reviewable artifact for the next stage.

  1. 01

    Intake

    Normalize the peptide, environment, objective, hard constraints, allowed modifications, and assay context.

    Task card
  2. 02

    Scan

    Map likely serum or protease liabilities, protected regions, and safely editable positions.

    Liability map
  3. 03

    Retrieve

    Search exact sequence, near-neighbor, motif, environment, and mechanism lanes while retaining gaps and errors.

    Raw evidence
  4. 04

    Normalize

    Convert retrieval results into typed evidence with source, effect, environment, confidence, and uncertainty.

    Evidence packet
  5. 05

    Design

    Choose allowed strategies and generate constraint-aware candidate edits with intended benefit and introduced risk.

    Candidate set
  6. 06

    Review

    Apply separate activity-preservation, stability, and developability scorecards.

    Review bundle
  7. 07

    Refine

    Prune hard failures and weak candidates, then apply bounded redesign to a focused pool.

    Survivors
  8. 08

    Rank

    Assemble rank-ready features, demote constraint failures, and apply visible weighted dimensions.

    Ordered shortlist
  9. 09

    Deliver

    Build a comparison report with provenance, confidence, risks, uncertainty, and next experiments.

    Decision report

06 / Experimental design

Questions, protocols, and comparison boundaries remain attached

Every study below describes what was varied, what was measured, and where interpretation must stop.

Illustrative design pattern

Terminal shielding

For ACDFGKLR, Pio marks the N-terminus as an editable liability while keeping the DF context protected. N-terminal acetylation is proposed as a hypothesis with target-engagement risk made explicit.

Sequence
ACDFGKLR
Proposed edit
N-terminal acetylation
Protected context
D3–F4
Next test
Serum time course + potency
Illustrative design pattern

Protease cleavage escape

For GLPVKRGI, a putative Arg6/Gly7 cleavage context motivates a D-Arg6 hypothesis. Stereochemical effects on conformation and activity remain a stated risk.

Sequence
GLPVKRGI
Proposed edit
D-Arg6
Liability context
R6–G7
Next test
Liquid chromatography–mass spectrometry (LC–MS) cleavage panel
Illustrative design pattern

Conformation lock

For AALKAAAE, a K4–E8 lactam bridge is proposed to favor a local helical state. Synthetic burden, solubility, and receptor geometry remain visible uncertainties.

Sequence
AALKAAAE
Proposed edit
K4–E8 lactam bridge
Review focus
Structure and synthesis
Next test
CD + serum + activity

07 / Results and evidence

Read the signal with its evidence boundary

Quantitative summaries, structural views, and scientific interpretation remain separate layers.

10

Registered scientific skills

Intake, evidence, design, review, and delivery capabilities

3

Independent review lanes

Activity preservation, stability, and developability

4

Governed memory classes

Working, evidence, user, and learnings memory

0

Hidden evidence-backed scores

A score remains unassigned when validation support is absent

01

The current product evidence is a traceable decision system

Pio implements task validation, staged runtime objects, ten registered scientific capabilities, three bounded reviewers, hard-constraint demotion, explicit weighted ranking, trace events, guardrails, and regression or acceptance checks.

Boundary. Software and evaluation coverage describe system behavior; they do not establish biological success for a proposed peptide edit.
02

Illustrative cases keep the parent structure and the proposed edit distinct

Each molecular panel renders the parent peptide conformer, highlights the editable site and protected context, and states when the proposed chemistry or stereochemistry is absent from the image. Benefit, introduced risk, evidence status, and next experiment travel together.

Boundary. The conformers illustrate review geometry. No sequence-specific effect is validated by these panels.
03

Recommendations remain typed by evidence strength

Observed support retains a source locator, inferred support remains labeled as mechanistic or analogue context, and missing or conflicting support lowers confidence. The final report carries uncertainty beside the ranking basis.

Boundary. Retrieval and reasoning support a testable decision; they cannot replace the discriminating experiment.
PyMOL parent conformer of ACDFGKLR with the N-terminal edit site and protected DF context highlighted
Illustrative case / terminal shield

Protect the functional context while exposing the editable terminus

The parent conformer marks the proposed N-terminal edit and the DF context that the task contract protects.

Evidence boundary. Acetylation is not modeled and no sequence-specific serum or activity effect is shown.
PyMOL parent conformer of GLPVKRGI with residue 6 and the Arg6 Gly7 cleavage context highlighted
Illustrative case / cleavage escape

A protease hypothesis remains linked to its local sequence context

The view marks residue 6 and the R6–G7 neighborhood used to motivate a D-Arg substitution hypothesis.

Evidence boundary. D-stereochemistry is not modeled and protease escape remains experimentally unverified.
PyMOL parent conformer of WQFGLM with the C-terminal methionine and adjacent leucine context highlighted
Illustrative case / tail cap

Terminal protection carries charge and binding consequences

The parent conformer separates the proposed C-terminal amidation site from the adjacent structural context.

Evidence boundary. Amidation is not modeled and its stability, solubility, and activity effects remain unknown.

Reading key

Scientific abbreviations

LC–MS
Liquid chromatography–mass spectrometry, proposed here for peptide cleavage mapping.
CD
Circular dichroism spectroscopy, proposed here to test changes in peptide secondary structure.
Evidence object
A normalized record containing source type, locator, effect, environment, confidence, and uncertainty.

08 / Limitations and provenance

What the current evidence can establish

The page describes implemented agent behavior and clearly labeled illustrative cases. It introduces no biological result absent from the project evidence.

01

Illustrative cases are unvalidated

The case atlas demonstrates the review contract and proposed experiments; it contains no sequence-specific efficacy claim.

02

Retrieval depends on configured backends

Live literature, patent, sequence, and structure lookup can return partial coverage or explicit errors when sources are unavailable.

03

Analysis tools include bounded proxies

Current protease, aggregation, immunogenicity, and structure-context interfaces do not replace full external modeling or experimental assays.

04

Current orchestration is sequential

Multiple worker lanes are represented as governed branches, while the current runtime executes them in sequence.