Version: 1.0
Locked before: Phase 3 candidate generation run
This document pre-registers the selection rule used to nominate candidates from the Phase 3 generation batch. The rule and all thresholds are locked in advance and may not be changed after the generation run produces scores.
All scores produced by this foundry are heuristic physicochemical proxies. They are not validated biological predictors. No antimicrobial activity has been demonstrated in vitro or in vivo. These rules select candidates for possible future expert review and assay — they do not predict whether any peptide will be active, safe, or useful.
As defined in configs/phase3.yaml (separate from the benchmark config pipeline.yaml):
| Dimension | Weight |
|---|---|
| activity | 0.35 |
| safety | 0.30 |
| synthesis | 0.20 |
| novelty | 0.15 |
Ensemble score = weighted sum of the four dimensions.
A candidate passes all of the following gates, evaluated in order:
| Gate | Threshold | Rationale |
|---|---|---|
| Length | 8–35 aa | Pipeline filter; shorter too short, longer harder to synthesize |
| Canonical AAs only | No non-standard residues | Synthesis feasibility |
| Ensemble score | ≥ 0.50 | Minimum quality bar across all dimensions |
| Safety score | ≥ 0.60 | max_safety_risk = 0.40; hemolysis/toxicity proxies |
| Novelty score | ≥ 0.05 | Minimum 1-mutation difference from any reference seed |
| Activity score | ≥ 0.40 | Must meet minimum predicted activity signal |
Note on novelty threshold: Phase 3 candidates are generated by 1–3 conservative substitutions from AMP-like template seeds. A min_novelty of 0.05 ensures every selected candidate differs by at least one amino acid from the nearest seed, while allowing the generator to fully explore the near-seed neighborhood. A higher threshold (e.g. 0.20, used in pipeline.yaml for benchmark purposes) would exclude most or all near-seed variants and is not appropriate here.
After gate filtering, candidates are ranked by ensemble score (descending). Greedy diverse selection is applied: each candidate added to the batch must have pairwise normalized Levenshtein similarity < 0.80 with all previously selected candidates.
Target batch size: 50–100 candidates.
The Phase 3 batch passes if:
- ≥ 50 candidates clear all gates above
- ≥ 5 distinct template seeds are represented (diversity of origin)
- No selected candidate has ensemble score < 0.50
- No selected candidate has safety score < 0.60
If fewer than 50 candidates pass, the generation parameters (n_double, n_charge_enhance) may be increased and the batch re-run. This will be documented transparently.
- It does not claim any candidate will show antimicrobial activity.
- It does not claim any candidate is safe for any use.
- It does not constitute a drug development programme.
- It does not replace expert biological evaluation.
The lab is the judge. This rule only determines which computational candidates are nominated for possible future evaluation.