hiwosy-bio 0.1.0

The bioisostere trust layer — score a molecular swap with full evidence.

Hiwosy BIO — the bioisostere trust layer

Paste a molecule and a proposed swap. Get a defensible confidence score and the full evidence trail in five seconds.

Hiwosy BIO is the trust layer for bioisosteric replacements. Not a discovery engine. Not a recommender. Not a generator. A scorer.

A medicinal chemist who is about to commit six weeks of bench time to synthesizing an analog wants one question answered, fast: "Is this swap supported by the public data — and how strongly?"

That is the only question Hiwosy BIO answers. It does so with:

• A 0.0–1.0 confidence score decomposed into 9 transparent factors.
• The full evidence trail: every matched molecular pair from ChEMBL that supports or refutes the swap, broken down by target class.
• Off-target divergence detection (the Vioxx-style safety flag).
• Liability annotations for known-problematic functional groups.
• Synthetic accessibility scoring on the proposed analog.
• Novelty classification (KNOWN / SEMI-NOVEL / NOVEL).
• Patent landscape — stubbed in v1, hookable to SureChEMBL.

---

Positioning vs. the rest of the world

Tool	What it does	Why a chemist still needs Hiwosy BIO
SwissBioisostere (free DB)	Lookup table	No confidence, no breakdown, no context
BIOSTER (paid catalogue)	Curated list	Static, no quantitative confidence
Schrödinger Bioisostere Search	Embedded in Maestro	Black box, no factor decomposition
MMPDB (open-source)	Raw MMP extraction	Low-level, no scoring
Generative ML (REINVENT, MMPT-RAG)	Generate analogs	No traceability — can't tell you why
SureChEMBL	Patent search	No scientific support score

Hiwosy BIO is the single tool you run before committing to synthesis. Show it to your PI. Show it to a regulator. Show it to a patent attorney. Every confidence number traces back to the underlying compound pairs. No black boxes. No vibes. Always show your work.

---

Quick start

pip install -r requirements.txt

python -m hiwosy_bio.cli score \
  --mol "CC(C)Cc1ccc(C(C)C(=O)O)cc1" \
  --swap "carboxylic_acid:tetrazole" \
  --target-class "cyclooxygenase"

Or in Python:

from hiwosy_bio import ConfidenceEngine

engine = ConfidenceEngine.default()
report = engine.score(
    smiles="CC(C)Cc1ccc(C(C)C(=O)O)cc1",     # ibuprofen
    swap=("carboxylic_acid", "tetrazole"),
    target_class="cyclooxygenase",
)
print(report.render())

---

What the output looks like

─────────────────────────────────────────────────────
INPUT
  Molecule       : CC(C)Cc1ccc(C(C)C(=O)O)cc1
  Proposed swap  : carboxylic_acid → tetrazole
  Context        : target_class = cyclooxygenase
─────────────────────────────────────────────────────
CONFIDENCE     : 0.79  (HIGH — would synthesise)
CLASSIFICATION : KNOWN
─────────────────────────────────────────────────────
Evidence: 142 matched pairs
  ├── 38   from same target class (cyclooxygenase)
  ├── 87   from related targets
  └── 17   from unrelated targets

9-Factor breakdown:
  F1 Observations         0.91   142 pairs ≥ threshold
  F2 Target diversity     0.84   54 distinct targets
  F3 Activity quality     0.72   geomean ratio 1.3x
  F4 Scaffold diversity   0.78   29 distinct Murcko scaffolds
  F5 Consistency          0.81   variance 0.18
  F6 Fragment Tanimoto    0.65
  F7 Δ Property dist      0.84   ΔPSA +12, ΔlogP -0.3
  F8 Δ LLE                0.71   avg -0.2
  F9 Size match           0.92   ΔMW +12 Da

Liability flags:
  ⚠ tetrazole adds an ionizable acidic moiety (pKa ~4.9)
  ⚠ tetrazole may alter HSA binding profile

Off-target divergence : uniform ✓ (CV log-ratio 0.18)
Synthetic accessibility: 3.2 (moderate)
Novelty               : KNOWN

Patent landscape (SureChEMBL stub):
  ⓘ patent backend not configured

Top supporting MMP examples:
  CHEMBL3801 ↔ CHEMBL3802  COX-1  ratio=1.10x
  CHEMBL5621 ↔ CHEMBL5622  PTGS2  ratio=0.92x
  ... 140 more (use --show-evidence to dump all)
─────────────────────────────────────────────────────

---

Architecture

hiwosy_bio/
  core.py          ConfidenceEngine — main orchestrator
  evidence.py      Loads V3 candidates_learned.json, filters by target class
  factors.py       9-factor confidence (wraps molecular-discovery V3 validator)
  divergence.py    Off-target divergence detector (CV of log activity ratios)
  liabilities.py   Per-group liability lookup
  synth_score.py   RDKit SA-Score on the proposed analog
  target_classes.py ChEMBL target → class taxonomy
  molecule.py      SMILES parsing, group localization, analog construction
  report.py        Renders the formatted output
  cli.py           Command-line entry point
data/
  liabilities.json    ~30 known group liability annotations
  target_classes.json ChEMBL target → class mapping (curated subset)

The engine reuses the 9-factor ConfidenceValidator from molecular-discovery and reads candidate evidence from any candidates_learned.json produced by the V3 cascade. Hiwosy BIO is the consumer surface; molecular-discovery remains the data factory.

---

What Hiwosy BIO does NOT do

This is the discipline. Each is a real temptation:

• ❌ Generate novel molecules — leave to ML
• ❌ Predict ADMET — leave to ADMETlab
• ❌ Retrosynthesis — leave to AiZynth
• ❌ Active learning loops — workflow product, locks out API distribution
• ❌ 3D shape matching — niche; opt-in factor only
• ❌ Multi-target selectivity bioisosteres — niche; filter inside engine

Every feature is a way to make the confidence number more defensible, faster, or more actionable. If a feature does not directly improve "score one swap in 5 seconds with full evidence," it does not ship.

---

License

TBD. Open-source candidate (MIT or Apache-2.0).