Novelty Audit Guide — OpenAMP Foundry

Last updated: 2026-06-29
Audit level: STRONG (APD6 + DRAMP 3.0 + UniProt + ESCAPE NeurIPS-2025 + dbAMP 3.0)
DB size: 51,503 unique clean standard-AA sequences (5–100 aa)

Why Novelty Auditing Matters

Every AMP candidate must pass a full novelty audit before wet-lab synthesis to:

Avoid IP risk — DRAMP patent sequences are flagged; any candidate with ≥60% identity to a patent sequence receives POSSIBLE_PATENT_RISK and a SYNTHESIS_HOLD.
Ensure scientific validity — a "novel" label on a KNOWN_VARIANT of a published AMP is misleading.
Protect the budget — synthesis of a patent-risk or known candidate wastes money.

The single biggest mistake in OpenAMP Wave 0.5 was relying on the novelty-check-broad CLI tool, which uses a 72-sequence curated reference — far too small to catch patent proximity. See docs/postmortems/2026-06-29-seed020-novelty-patent-risk.md for the full incident.

Rule: Always run scripts/novelty/run_expanded_novelty_audit.py before adding any candidate to the panel or ordering synthesis.

Database Sources

All FASTA files are committed to data/novelty_db/.

File	Source	Sequences	Patent-flagged	Notes
`apd6_natural.fasta`	APD6	3,306	No	Canonical natural AMPs, curated
`apd6_animal.fasta`	APD6 2024	2,580	No	Animal AMPs with known activity
`apd6_plant.fasta`	APD6 2024	268	No	Plant defensins & thionins
`apd6_bacteria.fasta`	APD6 2024	410	No	Bacteriocins, lantibiotics
`dramp_general.fasta`	DRAMP 3.0	11,687	No	General AMP collection
`dramp_patent.fasta`	DRAMP 3.0	18,715	YES	Patent-protected AMPs — IP risk detection
`dramp_specific.fasta`	DRAMP 3.0	6,321	No	Clinical, stability, structural entries
`uniprot_amps_reviewed.fasta`	UniProt KW-0929 ≤100aa, reviewed	2,673	No	Swiss-Prot curated
`uniprot_amps_unreviewed.fasta`	UniProt KW-0929 ≤60aa, unreviewed	1,692	No	TrEMBL broad coverage
`escape_amps.fasta`	ESCAPE NeurIPS-2025	3,542	No	21k AMPs from 27 repos; 3,542 clean standard-AA extracted
`dbAMP3.fasta`	dbAMP 3.0	35,599	No	Largest single AMP DB; 29,984 clean standard-AA extracted
`dbaasp-peptides.fasta`	DBAASP	1,988	No	Database of Antimicrobial Activity and Structure of Peptides
`apd6_human.fasta`	APD6 2024 (human subset)	154	No	Human host-defense peptides (cathelicidins, defensins, histatins)
`uniprot_amps.fasta`	UniProt KW-0929 (combined pull)	2,348	No	Separate UniProt download; fully overlaps with reviewed+unreviewed above

After deduplication: 51,503 unique clean standard-AA sequences (5–100 aa)
(DBAASP, APD6 human, and UniProt combined added 0 net new sequences — all subsumed by existing DB.)

Databases NOT yet included

Database	Size	Status	Why excluded
UniProt nr (full)	Billions	Scope	Beyond AMP-specific novelty; use BLASTp for global search

DBAASP activity / hemolytic data

data/novelty_db/activity-against-target-species.csv and hemolytic-and-cytotoxic-activities.csv are DBAASP metadata files (MIC values + hemolytic data). They are reference data for panel annotation, not sequence novelty files. Use them when interpreting activity predictions.

For global novelty (pre-grant): Run BLASTp against NCBI nr and Google Patents / Lens.org in addition to this audit.

Updating the databases

# Re-download APD6 (run annually)
curl -sL "https://aps.unmc.edu/assets/sequences/naturalAMPs_APD2024a.fasta" \
  -o data/novelty_db/apd6_natural.fasta
curl -sL "https://aps.unmc.edu/assets/sequences/animalAMPs_APD2024a.fasta" \
  -o data/novelty_db/apd6_animal.fasta
# (plant and bacteria similarly)

# Re-download DRAMP 3.0 (run annually)
BASE="https://dramp.cpu-bioinfor.org/downloads/download.php?filename=download_data/DRAMP3.0_new"
curl -sL "${BASE}/general_amps.fasta" -o data/novelty_db/dramp_general.fasta
curl -sL "${BASE}/patent_amps.fasta"  -o data/novelty_db/dramp_patent.fasta
curl -sL "${BASE}/specific_amps.fasta" -o data/novelty_db/dramp_specific.fasta

# Re-download UniProt (run annually) — use the download-novelty-dbs paginator
python3 scripts/novelty/download_novelty_dbs.py --uniprot

# Re-download ESCAPE (run when Harvard Dataverse dataset is updated)
python3 scripts/novelty/download_novelty_dbs.py --escape

Audit Methodology

Algorithm

BLOSUM62 local alignment (BioPython PairwiseAligner):

gap_open   = -11
gap_extend = -1
mode       = local
identity   = (exact matches in aligned blocks) / query_length

This catches conservative substitutions (K↔R, I↔L↔V, F↔W↔Y) that simple edit distance misses.

Novelty Thresholds

Identity	Class	Meaning
≥99%	`EXACT_MATCH_OR_FRAGMENT`	Identical or contained within a known sequence
≥80%	`KNOWN_VARIANT`	Minor point mutations from a known AMP
≥60%	`CLOSE_RELATIVE`	Same structural family, meaningful differences
≥40%	`RELATED_NOVEL`	Related but sufficiently distinct
<40%	`HIGH_CONFIDENCE_NOVEL`	Genuinely new sequence space

Patent Risk Classification

Condition	Flag
Best hit in `dramp_patent` AND identity ≥60%	`POSSIBLE_PATENT_RISK` → SYNTHESIS HOLD
Best hit in `dramp_patent` AND identity ≥40%	`LOW_PATENT_RISK` → consult IP counsel
Best hit in `dramp_patent` AND identity <40%	`CLEAR`
Best hit not in patent DB	`CLEAR`

Running the Audit

Standard usage

# From repo root, always use .venv/bin/python3
.venv/bin/python3 scripts/novelty/run_expanded_novelty_audit.py \
    --input outputs/my_candidates.csv \
    --out outputs/my_novelty_audit.csv

# Or from a FASTA
.venv/bin/python3 scripts/novelty/run_expanded_novelty_audit.py \
    --fasta outputs/my_candidates.fasta \
    --out outputs/my_novelty_audit.csv

Input CSV must have columns: candidate_id, sequence

Interpreting results

✓ HIGH_CONFIDENCE_NOVEL + CLEAR  →  approved for predictor submission and wet lab
✓ RELATED_NOVEL + CLEAR          →  approved; note 40-60% similarity in documents
⚠ CLOSE_RELATIVE + CLEAR         →  approved; requires explicit prior-art disclosure
⚠ CLOSE_RELATIVE + POSSIBLE_PATENT_RISK  →  SYNTHESIS HOLD; IP review first
⚠ KNOWN_VARIANT / EXACT_MATCH   →  exclude from novelty claims; use as control only

Before ordering synthesis

# 1. Run audit
.venv/bin/python3 scripts/novelty/run_expanded_novelty_audit.py --input candidates.csv

# 2. Confirm output CSV — all selected candidates must show CLEAR patent_risk
grep "POSSIBLE_PATENT_RISK\|REVIEW_REQUIRED" outputs/expanded_novelty_audit_*.csv

# 3. Never use novelty-check-broad alone — it uses only 72 sequences

What Audit Level Do We Have?

Level	Requirements	Status
Minimum (pre-wet-lab)	APD6 + DRAMP	✅
Strong (pre-grant)	+ dbAMP + UniProt + ESCAPE	✅ All included
Best-in-class	+ NCBI pataa (local BLASTp) + NCBI nr (web API) + Lens.org	⚠ Run `scripts/novelty/run_patent_blastp.py`

The current audit is STRONG level and sufficient for wet-lab synthesis decisions.

For patent filings or grant submissions, additionally run:

BLASTp against NCBI nr (web: https://blast.ncbi.nlm.nih.gov/Blast.cgi?PROGRAM=blastp&PAGE_TYPE=BlastSearch)
Exact-sequence search on Lens.org or Google Patents

De Novo Campaign Results (2026-06-29)

All de novo candidates generated by scripts/research/generate_denovo_novel_amps.py (seed=42, 30 targets found in 15,315 attempts) were re-audited against the expanded 51,503-sequence DB.

Progressive audit results as each database was added:

Result	34,555 (pre-ESCAPE)	37,005 (+ESCAPE)	51,503 (+dbAMP3)
HIGH_CONFIDENCE_NOVEL + CLEAR	23	19	8
RELATED_NOVEL + CLEAR	7	11	22
POSSIBLE_PATENT_RISK	0	0	0
SYNTHESIS_HOLD	0	0	0

dbAMP3 added 14,498 new unique sequences (largest single addition). With all 5 major AMP databases covered, 8 candidates remain genuinely HIGH_CONFIDENCE_NOVEL at <40% to all 51,503 known AMP sequences, with zero patent risk.

outputs/denovo_top12_final.fasta — 2 submission batches, all CLEAR:

Batch A (8 × HIGH_CONFIDENCE_NOVEL): DENOVO_009/010/005/011/029/016/007/024 — submit first
Batch B (4 × RELATED_NOVEL, 40–44%): DENOVO_026/025/001/008 — submit if A insufficient

Prior wave novelty audit findings (outputs/wave0_pilot_novelty_audit.csv): All Wave 0 "HIGH_CONFIDENCE_NOVEL" labels were wrong — those were KNOWN_VARIANT (>80%) or CLOSE_RELATIVE when checked against the full DB.

Agents / Automated Use

When an agent generates AMP candidates:

Write sequences to a CSV with candidate_id, sequence columns.
Run: .venv/bin/python3 scripts/novelty/run_expanded_novelty_audit.py --input <csv> --out outputs/audit_<timestamp>.csv
Parse output — reject any row where patent_risk != CLEAR.
For wet-lab submission, require novelty_class IN (HIGH_CONFIDENCE_NOVEL, RELATED_NOVEL).
Never cite novelty-check-broad (72-seq) output in a submission document.

Navigation Menu

Repository navigation

NOVELTY_AUDIT_GUIDE.md

NOVELTY_AUDIT_GUIDE.md

Novelty Audit Guide — OpenAMP Foundry

Why Novelty Auditing Matters

Database Sources

Databases NOT yet included

DBAASP activity / hemolytic data

Updating the databases

Audit Methodology

Algorithm

Novelty Thresholds

Patent Risk Classification

Running the Audit

Standard usage

Interpreting results

Before ordering synthesis

What Audit Level Do We Have?

De Novo Campaign Results (2026-06-29)

Agents / Automated Use

FilesExpand file tree

NOVELTY_AUDIT_GUIDE.md

Latest commit

History

NOVELTY_AUDIT_GUIDE.md

File metadata and controls

Novelty Audit Guide — OpenAMP Foundry

Why Novelty Auditing Matters

Database Sources

Databases NOT yet included

DBAASP activity / hemolytic data

Updating the databases

Audit Methodology

Algorithm

Novelty Thresholds

Patent Risk Classification

Running the Audit

Standard usage

Interpreting results

Before ordering synthesis

What Audit Level Do We Have?

De Novo Campaign Results (2026-06-29)

Agents / Automated Use