ChiralAI

AI-guided discovery and validation engine for biocatalytic chiral molecule targets.

---

The Problem

Enantiopure compounds are essential in pharmaceutical synthesis, agrochemistry, and advanced materials. ~80% of chiral active pharmaceuticals now have to be single enantiomers. Biocatalysis is the preferred route: enzymes are inherently chiral, operate under mild conditions, and can achieve >99% ee (enantiomeric excess). But identifying the right enzyme, for the right substrate, with the right stereochemical outcome, in a host that can actually produce the compound, that requires manually stitching together five different tools and databases that were never designed work compatibly.

No existing software does this end-to-end.

---

What ChiralAI Does

Given a natural-language query, ChiralAI runs a grounded discovery and validation pipeline:

User query (e.g., "enantiopure amine building block for beta-lactam synthesis")
    ↓
GPT-4.1 — suggests 5 candidate chiral molecules with defined R/S stereochemistry
    ↓
RDKit — validates chirality; identifies and assigns R/S stereocenters
    ↓
KEGG — maps compounds to known metabolic pathways and enzyme classes
    ↓
Route predictor — backward search through KEGG reactions from target to central metabolites
    ↓
BRENDA — retrieves known ee values for ECs across all route steps (deduplicated)
    ↓
COBRApy FBA — checks metabolic feasibility in E. coli iJO1366; flags cofactor requirements
    ↓
Composite scorer — ranks candidates by ee source, Tanimoto substrate similarity, and feasibility
    ↓
Timestamped CSV + JSON output with full provenance and confidence tiers

The LLM is the orchestration and reasoning layer, not the scientific ground truth. Every suggestion is grounded in a database call or computational result. LLM-claimed ee values are labeled "llm_claim" and discounted; BRENDA-verified ee is labeled "brenda_verified" and weighted higher.

---

Why This Is a Real Gap

Tool	What It Does	What It Misses
RetroBioCat	Biocatalytic route planning	No stereochemistry or enantioselectivity awareness
ASKCOS	Organic retrosynthesis	Not built for enzymatic pathways
ChemCrow	GPT-4 + chemistry tools	Organic synthesis only; no biocatalysis
COBRApy	Genome-scale metabolic FBA	Ignores stereochemistry entirely
BRENDA	Gold-standard enzyme database with ee values	A database, not a discovery tool

ChiralAI's contribution is integration — connecting chiral validation, pathway context, enantioselectivity data, and metabolic feasibility in a single workflow accessible via natural language.

---

Quickstart

# 1. Install dependencies
pip install -r requirements.txt

# 2. Set credentials
cp env.example .env
# Add: OPENAI_API_KEY, BRENDA_EMAIL, BRENDA_PASSWORD

# 3. Run
python3 main.py

Enter a query like:

• "suggest a chiral amino acid precursor for asymmetric synthesis"
• "enantiopure lactone building blocks for biodegradable polymers"
• "(R)-selective secondary alcohol for pharmaceutical synthesis via E. coli fermentation"

Results are saved to a timestamped CSV and JSON sidecar in the project directory.

---

Architecture

ChiralAI/
├── main.py                            # Orchestrator — runs the full pipeline
├── ChiraLLM/
│   ├── query_handler.py               # GPT-4.1 — molecule suggestion (5 ranked candidates)
│   ├── chirality_checker.py           # RDKit — stereocenter detection and R/S assignment
│   ├── database_validator.py          # KEGG REST API — pathway and enzyme lookup
│   ├── brenda_client.py               # BRENDA SOAP — ee values from enzyme substrate data
│   ├── feasibility_checker.py         # COBRApy FBA — metabolic feasibility in iJO1366
│   ├── enantioselectivity_scorer.py   # Composite scorer — ranks by ee, Tanimoto, feasibility
│   └── route_predictor.py            # Tier 1 biosynthesis route predictor — A* over KEGG graph
└── utils/
    └── file_saver.py                  # Timestamped CSV + JSON export

---

Output Columns (CSV)

Column	Source	Notes
scoring_composite_score	Scorer	0–1; weighted ee + Tanimoto + feasibility
scoring_confidence	Scorer	high / medium / low
scoring_top_enzyme_ec	BRENDA / KEGG	Best-ranked EC number
scoring_top_enzyme_ee	BRENDA / LLM	ee% value
scoring_top_enzyme_source	Scorer	brenda_verified or llm_claim
scoring_stereo_confirmed	RDKit	True only if all stereocenters are R/S assigned
scoring_feasibility_flux	COBRApy	mmol/gDW/h in iJO1366; None if not in model
scoring_notes	Scorer	Human-readable provenance and caveats

---

Scientific Grounding

• RDKit — stereocenter detection, CIP R/S assignment, Morgan fingerprints for Tanimoto
• KEGG — compound, pathway, and enzyme commission data
• BRENDA — 112k enzymes, 5.8M data points; ee% extracted from substrate commentary fields
• PubChem — substrate SMILES resolution for Tanimoto computation
• COBRApy + iJO1366 — genome-scale E. coli K-12 metabolic model for flux analysis

---

Known Limitations

• BRENDA credentials required for verified ee. Without BRENDA_EMAIL / BRENDA_PASSWORD in .env, all ee values fall back to LLM claims labeled llm_claim. BRENDA registration is free.
• E. coli only. The FBA layer uses iJO1366 (E. coli K-12). Secondary metabolites and many pharmaceutical targets return not_in_model. Other host models (S. cerevisiae, P. putida) are not yet supported.
• Tier 1 only. Route prediction works for compounds KEGG already covers (~12k reactions). Novel targets require Tier 2 (RetroRules SMARTS retrobiosynthesis), planned for the next sprint.

---

Roadmap

• Tier 2 — novel-target retrobiosynthesis via RetroRules + RDKit RunReactants
• Non-E. coli host models — S. cerevisiae (iMM904), P. putida support in FBA layer
• Engineered variant data — wire BRENDA getEngineering for directed evolution candidates
• Name↔SMILES stereo consistency check — programmatic CIP verification against molecule name

---

Questions or ideas? Connect on LinkedIn: https://www.linkedin.com/in/alexeimanuel/