RENKIN — Retrosynthesis Engine for Knowledge-Informed Navigation

Computer-Aided Synthesis Planning (CASP) · Pure Rust · WebAssembly · Python
Named after 錬金 (れんきん, renkin) — Japanese for alchemy: just as alchemists transformed base metals into gold, RENKIN transforms target molecules back into cheap starting materials.

日本語版 README · 中文版 README · Documentation · Live Demo →

What is RENKIN?

RENKIN is an open-source retrosynthesis engine for computer-aided synthesis planning (CASP) that automatically discovers optimal chemical reaction routes from a target molecule back to cheap, commercially available starting materials.

Built entirely in Rust with the chematic cheminformatics crate — zero C/C++ dependencies, #![forbid(unsafe_code)] throughout. One codebase compiles to a native CLI, a Rust library, Python wheels (PyO3), and a WebAssembly module that runs entirely client-side in the browser.

Installation

pip install renkin          # Python
cargo add renkin            # Rust
npm install renkin          # JavaScript / Node.js

Live Playground

→ Try it now — runs entirely in WebAssembly: no installation, no server, no network calls.

Quick Start

import json
import renkin

result = json.loads(
    renkin.find_routes(
        target="CC(=O)Oc1ccccc1C(=O)O",  # Aspirin
        depth=5,
        max_routes=3,
    )
)

for route in result["routes"]:
    for step in route["steps"]:
        print(f"  {step['target']} → {' + '.join(step['precursors'])}  [{step['rule']}]")

import init, { find_routes } from './pkg/renkin.js';
await init();
const result = JSON.parse(find_routes("CC(=O)Oc1ccccc1C(=O)O", 5, 3, 0));

./target/release/renkin --target "CC(=O)Oc1ccccc1C(=O)O" --depth 5 \
    --templates data/templates_extracted_5000.smi --format tree

Target: CC(=O)Oc1ccccc1C(=O)O
Routes found: 3

Route 1  [score=1.02, depth=1]
OC(=O)c1ccccc1OC(=O)C
└── [extracted_169]
    ├── OC(=O)C  ✓ BB
    └── [OH]c1ccccc1C(=O)O  ✓ BB

Route 2  [score=1.02, depth=1]
OC(=O)c1ccccc1OC(=O)C
└── [extracted_145]
    ├── CC(=O)Cl  ✓ BB
    └── [OH]c1ccccc1C(=O)O  ✓ BB

Route 3  [score=1.03, depth=1]
OC(=O)c1ccccc1OC(=O)C
└── [extracted_238]
    ├── c1cccc(c1O)C(O)=O  ✓ BB
    └── C([OH])(=O)C  ✓ BB

Use --format mermaid for GitHub/Notion-compatible flowcharts.

Current Limitations

⚠️ Benchmark numbers are under active re-measurement after a validator-accuracy fix — historical 78.0%/95.9%/81.8%(ChEMBL) figures elsewhere in this repo predate that fix and are invalidated. RENKIN does not predict yields, calibrated experimental success probabilities, or side reactions, and does not search the literature automatically (success_probability is a template-frequency search-ranking score, not a calibrated prediction — see Benchmark for the current corrected numbers, full methodology, and known limitations).

Why RENKIN?

RENKIN is designed as a Rust-native synthesis planning stack:

Constraint-based Search

Restrict routes by the element composition of their building blocks.

Default search — all 5 routes for biphenyl:

renkin --target "c1ccc(-c2ccccc2)cc1" --templates data/templates_extracted_5000.smi --format tree

Routes found: 5
Route 1  [score=1.00, depth=1]  c1ccccc1Br + c1c(B(O)O)cccc1
Route 2  [score=1.03, depth=1]  c1ccccc1Br + c1c(B(O)O)cccc1
Route 3  [score=1.06, depth=1]  c1cc(Cl)ccc1 + c1c(B(O)O)cccc1
Route 4  [score=1.08, depth=1]  c1(I)ccccc1  + c1c(B(O)O)cccc1
Route 5  [score=1.08, depth=1]  c1ccccc1Br  + c1(B2OC(C(C)(C)O2)(C)C)ccccc1

Constrained search — boronic-acid coupling, no Br or I starting materials:

renkin --target "c1ccc(-c2ccccc2)cc1" --templates data/templates_extracted_5000.smi \
    --require-elements "B" --avoid-elements "Br,I" --format tree

Routes found: 1

Route 1  [score=1.06, depth=1]
c1ccccc1-c2ccccc2
└── [extracted_398]
    ├── c1cc(Cl)ccc1  ✓ BB
    └── c1c(B(O)O)cccc1  ✓ BB

Constraints compose freely and are enforced in two layers:

• --avoid-elements prunes expansions during search when a BB precursor contains a forbidden element (no dead-end nodes added to the heap).
• A final route-level post-filter is still applied for correctness.
• --require-elements is a route-level post-filter only.

Add --verbose to print search statistics (nodes expanded, elapsed time) to stderr. Performance counters are available in native builds only; disabled in WASM.

Template Evidence Metadata

Extracted templates only have a positional display name (extracted_{i}) that changes whenever the source .smi file is reordered or re-extracted, so external knowledge (a DOI, a reported yield, a known side reaction) can't be durably attached to one. Every template — hand-crafted and extracted — now has a stable template_id instead:

• Hand-crafted rules: rule:<rule_name> (e.g. rule:suzuki_retro).
• Extracted templates: smirks-sha256:<hex> — the SHA-256 hex digest of the trimmed SMIRKS string. Independent of file position, load order, and count; purely syntactic (no SMIRKS canonicalization — a semantically equivalent SMIRKS written differently gets a different ID).

Run renkin template ids <file.smi> to list every template's template_id, display name, SMIRKS, and weight (TSV by default, --format json for JSON) — use this to look up the IDs you need when authoring a sidecar file.

Attach curated evidence with --template-metadata sidecar.json (also available in Python as find_routes(..., template_metadata_path=...)), keyed by template_id:

{
  "schema_version": 1,
  "templates": {
    "smirks-sha256:ef8778a2888469d619c52cce7e74f6848e101049050dd1b765b78f32e3c94498": {
      "references": [
        { "id": "ref-1", "kind": "doi", "identifier": "10.xxxx/example" }
      ],
      "condition_candidates": [
        {
          "catalysts": ["Pd(PPh3)4"],
          "bases": ["K2CO3"],
          "solvents": ["EtOH", "water"],
          "temperature_c": { "min": 75.0, "max": 85.0 },
          "source": "literature",
          "scope": "template",
          "reference_ids": ["ref-1"]
        }
      ],
      "reported_yields": [
        {
          "percentage": { "min": 72.0, "max": 81.0 },
          "basis": "isolated",
          "source": "literature",
          "scope": "template",
          "reference_ids": ["ref-1"]
        }
      ],
      "warnings": [
        {
          "code": "possible_protodeboronation",
          "severity": "medium",
          "message": "Protodeboronation has been reported under prolonged aqueous heating.",
          "source": "literature",
          "scope": "template",
          "reference_ids": ["ref-1"]
        }
      ]
    }
  }
}

A matching step gets an evidence field with condition_candidates, reported_yields, references, and warnings; steps whose template has no sidecar entry get no evidence key at all. The sidecar is loaded and validated (schema version, duplicate/dangling reference IDs, yield range, range min <= max, non-empty DOI/patent identifiers) before search starts — malformed metadata is a hard error, and a template_id in the sidecar that matches no loaded rule prints a warning rather than failing silently.

What this is not:

• reported_yields is a curated record of what was reported externally — not a RENKIN prediction. step_confidence/success_probability are unaffected and keep meaning template-frequency-derived search-ranking scores, not experimental success rates.
• warnings reflects only what's explicitly present in the sidecar you supply — not automatic side-reaction detection.
• Templates without a matching sidecar entry get no fabricated evidence. Nothing is invented for missing data.

Yield/success prediction and automatic literature search are explicitly out of scope for this phase — tracked as future work in #41.

Substrate-specific examples (schema_version: 2)

Everything above is template-level: it applies to every step using that template, regardless of the actual molecule. schema_version: 2 adds examples — a per-template array where each entry is one curated record of this exact reaction, keyed by target_smiles/precursor_smiles:

{
  "schema_version": 2,
  "templates": {
    "smirks-sha256:...": {
      "references": [{ "id": "ref-1", "kind": "doi", "identifier": "10.xxxx/example" }],
      "examples": [{
        "id": "ex-1",
        "target_smiles": "c1ccc(-c2ccccc2)cc1",
        "precursor_smiles": ["Brc1ccccc1", "c1ccccc1"],
        "conditions": { "catalysts": ["Pd(PPh3)4"], "solvents": ["EtOH"], "source": "literature", "scope": "substrate_specific", "reference_ids": ["ref-1"] },
        "reported_yield": { "percentage": 78.0, "basis": "isolated", "source": "literature", "scope": "substrate_specific", "reference_ids": ["ref-1"] },
        "reference_ids": ["ref-1"]
      }]
    }
  }
}

examples requires schema_version: 2 (a hard error under 1); under schema_version: 2, reported yields must live under examples[].reported_yield too — a non-empty template-level reported_yields is a hard error there (it stays allowed under schema_version: 1), so a substrate-specific number can't leak onto every step using that template. Every condition/yield/warning nested inside an example must be scoped substrate_specific.

A route step's evidence.examples are resolved, not just copied from the sidecar: matched against that step by canonical target SMILES plus the canonical, order-independent precursor set (reordering precursor_smiles in the sidecar changes nothing), with every exact-substrate match kept and same-template-different-substrate precedents capped at 3. Each resolved entry carries a match_kind (exact_substrate/template_only) in the JSON itself, plus a template_examples_total count — so JSON/Python consumers, not just --format explain, can tell "evidence for this exact reaction" apart from "literature precedent for a different substrate." --format explain shows exact-substrate matches first, each labeled either Exact substrate example: or "different substrate; not a prediction", with conditions/ reported_yield/warnings each showing their own cited references directly underneath (deduplicated when the same reference backs more than one part of an example). See Reaction Evidence guide for full matching/validation semantics.

Importing evidence from ORD. renkin evidence match (exact-set batch template matching, no fuzzy/similarity matching) and scripts/ord_evidence_audit.py (offline, network-free) turn a locally-downloaded Open Reaction Database corpus into a schema_version: 2 sidecar — every accepted record is independently re-validated by RENKIN's own loader, and anything not uniquely matched, unambiguous, and provenanced is excluded and counted in an audit report rather than guessed at. RENKIN itself never fetches or searches the literature; reported yields are citations, not predictions. ORD's reaction data is CC-BY-SA-4.0, a different license from RENKIN's own MIT code — see Reaction Evidence guide for the full acceptance criteria and licensing split.

Key Features

step_confidence/success_probability are not yields or measured success rates. They're template-frequency-derived search-ranking scores (rule_weight / max_rule_weight, multiplied across a route's steps) used to order candidate disconnections during search — not a calibrated probability of experimental success, and not an expected isolated yield. Route-level experimental yield/success-rate reporting is not implemented.

Pipeline Examples

# Route cost scoring with commercial prices
renkin -t "Cc1ccc(-c2ccccc2)cc1" --bb-prices data/prices.csv --format json

# Standalone forward prediction — no route search involved
renkin-forward predict --reactants "Oc1ccccc1C(=O)O" "CCO" --report --max-results 5

# Forward validation — pipe find_routes output directly
renkin -t "CC(=O)Oc1ccccc1C(=O)O" --format json | renkin-forward validate

# Faster template retrieval with bond-center index (~24% speedup)
renkin -t "c1ccc(NC(=O)c2ccccc2)cc1" --templates data/templates_extracted_5000.smi --bond-index

Benchmark

USPTO-50k test set (4,907 molecules, full evaluation):

Evaluation definition: A molecule is solved if find_routes returns at least one route whose leaf precursors are all in the building block set, within depth=5 and beam=100. Ground-truth reactants from USPTO-50k are not checked — any commercially accessible route counts.

Corrected baseline (commit e20dc8c, 2026-07-22)

402 building blocks (unique compounds actually loaded from data/building_blocks.smi — see below), 5,000 extracted templates, 28 handcrafted rules, depth=5, beam=100. These three rates are a nested series over the same 4,907 targets, not independent numbers, and none is an experimentally-verified synthesis success rate or a human-chemist-reviewed route-accuracy figure. provenance_validated_solved_rate is not a measured chemical-accuracy rate and not a proven lower bound on correctness — it only counts routes the current validator can positively confirm, and an unknown fraction of "invalid" verdicts may be validator false negatives rather than real chemistry or route errors (the split is unmeasured). Full methodology, per-rule breakdown, and reproduction command: tasks/phase31_final_remeasurement_run.md · Full benchmark details →

Historical progression (pre-fix, invalidated — see notice above)

⚠️ The figures in this subsection (78.0% single-pass, 95.9% cascade, 81.8% ChEMBL OOD) predate the 31.11/31.12 fixes, are invalidated, and have not been re-measured. Kept for continuity only — do not cite as current performance.

Evaluation note: All numbers use the standard USPTO-50k train/test split (same corpus). Templates are extracted from the training set and evaluated on the test set. Numbers reflect performance within the USPTO-50k domain; out-of-distribution generalization was separately evaluated via ChEMBL approved drugs (81.8%, 409/500, also not re-measured).

* 29/50 chunks, previous binary
† 50/50 chunks — 72.1% (3,540/4,907) confirmed
BB counts in this historical table (463/480/509) are as originally documented at each point in time — legacy documentation values, not re-verified against ChemEnv::bb_count(). The corrected-baseline section above uses the actually-loaded count (402) for the current data/building_blocks.smi.

Note: LocalRetro (53.4%) and GLG (58.0%) report single-step top-1 prediction accuracy — a different metric, not directly comparable.

Benchmark scope note: USPTO-50k is used here as a standardized sanity benchmark, not as proof of broad real-world synthesis performance. The corpus covers a narrow slice of reaction space (primarily C–C and C–N bond formations common in pharmaceutical synthesis), and reaction types with sparse USPTO representation are systematically underserved. Out-of-distribution performance on ChEMBL approved drugs (81.8%, 409/500, pre-fix, not re-measured) suggested the rule set generalizes beyond the test corpus, but neither historical number should be interpreted as a guarantee of route quality on arbitrary targets.

PaRoutes compatibility

RENKIN is compatible with the PaRoutes multi-step benchmark. Download their stock compounds and target molecules, then pass them directly:

renkin-bench \
  --input paroutes_n1_targets.smi \
  --building-blocks paroutes_stock.smi \
  --templates data/templates_extracted_5000.smi \
  --depth 5 --beam-width 100

The JSON output includes avg_nodes_expanded, avg_confidence, avg_convergency, and avg_success_prob (Retro-prob style) alongside the standard solved/success_rate metrics.

Competitive Landscape

⚠️ RENKIN's row below uses the corrected raw_solved_rate (20.09%, see notice near the top of this README) — the 95.9% cascade figure some earlier versions of this table cited is invalidated and not re-measured; it is not included here.

raw_solved_rate is the closest available RENKIN metric to the published route-finding success rates of the other planners above, but the figures are not directly comparable — stock size, template library, target set, search budget, and route-quality checks all differ across systems, and this table does not establish RENKIN as better or worse than the alternatives.

RENKIN's goal: match state-of-the-art accuracy using only curated rules and auto-extracted SMIRKS templates — no GPU, no training data, no black boxes. Under RENKIN's benchmark setting (corrected baseline, commit e20dc8c, 2026-07-22), it reaches 20.09% raw_solved_rate (986/4,907) single-pass — see the Benchmark section above for the full nested-metric series and why the stricter provenance_validated_solved_rate (0.88%) is not RENKIN's measured or bounded correctness rate. RENKIN runs anywhere: browser, CLI, Python — single cargo build.

⚠️ The table above lists tools under different evaluation conditions. No matched-condition experiment against other tools has been performed.

MCP Server

renkin-mcp exposes retrosynthesis as an MCP tool so AI agents (Claude, etc.) can call it directly.

Setup — add to claude_desktop_config.json:

{
  "mcpServers": {
    "renkin": { "command": "/path/to/renkin-mcp" }
  }
}

Tools (6):

The server auto-detects data/building_blocks.smi and data/templates_extracted_5000.smi in the working directory. Falls back to the embedded DEFAULT_BUILDING_BLOCKS / default_rules() defaults if not found (152 unique building blocks per ChemEnv::bb_count(), 28 handcrafted rules — verified 2026-07-22; a "509-BB / 20-rule" figure was previously documented here without verification).

cargo build --release
# binary: target/release/renkin-mcp

Architecture

Workspace scope

┌──────────────────────────────────────────────────────────────────┐
│ renkin workspace (this repository)                               │
│                                                                  │
│  renkin  (retrosynthesis)         renkin-forward                  │
│  ──────────────────────           ─────────────────────────────  │
│  target → precursors              reactants → products           │
│  A* / AND-OR search               template-based forward         │
│  route scoring & constraints      (validates retro routes)       │
│        │                                    │                    │
│        └──────────────────┬─────────────────┘                    │
│                           ▼                                      │
│               chematic  (molecular representation,               │
│               SMILES, substructure matching, reaction SMARTS)    │
└──────────────────────────────────────────────────────────────────┘

Internal data flow (renkin crate)

Target SMILES
     │
     ▼
┌─────────────────────────┐
│     chem_env.rs         │  ← chematic wrapper
│  - SMILES parse         │     canonical-SMILES FxHashSet BB lookup (O(1))
│  - 20 built-in + up to 50k via --templates  │     fragment sanitization + ring-leak filter
│  - Building block check │     apply_retro memoization cache
└────────────┬────────────┘
             │  par_iter (rayon / sequential on WASM)
             ▼
┌─────────────────────────┐
│      search.rs          │  ← A* / AND-OR Tree Search
│  - Priority queue       │     SA Score heuristic + memoization
│  - Closed list          │     beam search (SmallVec frontier)
│  - Arc<PathNode> paths  │     O(1) path sharing per child
└────────────┬────────────┘
             │
             ▼
┌─────────────────────────┐
│      score.rs           │  ← Heuristic / Cost Function
│  - SA Score (chematic)  │     h = Σ(1 + 0.5·(sa−1)/9)
│  - MW step cost         │     g = Σ(1 + total_mw/2000)
└────────────┬────────────┘
             │
             ▼
┌─────────────────────────┐   (optional)
│      scorer.rs          │  ← Phase B: NN Template Scorer
│  - tract-onnx           │     Pure Rust ONNX inference
│  - --scorer flag        │     molecule-specific template ranking
└────────────┬────────────┘
             │
             ▼
  JSON  ←  CLI / Python / WASM

Project Structure

renkin/                          ← Cargo workspace root
├── Cargo.toml
├── src/                         ← renkin crate (retrosynthesis)
│   ├── lib.rs                   # public library
│   ├── main.rs                  # CLI binary (--templates, --template-metadata, --scorer, --constraints, --objectives flags)
│   ├── bin/benchmark.rs         # renkin-bench binary (--plausibility flag)
│   ├── bin/doctor.rs            # renkin-doctor diagnostic binary
│   ├── bin/fp.rs                # renkin-fp ECFP4 fingerprint (nn-scoring feature)
│   ├── bin/mcp.rs               # renkin-mcp MCP server (6 tools)
│   ├── chem_env.rs              # retro rules + BB lookup + template loader
│   ├── score.rs                 # SA Score heuristic + step cost
│   ├── search.rs                # A* / AND-OR tree engine + beam pruning
│   ├── scorer.rs                # Phase B: tract-onnx NN template scorer
│   ├── candidate.rs             # one-step candidate proposal (offline reranking foundation, not wired into search)
│   ├── pool_export.rs           # candidate-pool JSONL + reproducibility-manifest export
│   ├── python.rs                # PyO3 bindings (--features python)
│   └── wasm.rs                  # wasm-bindgen bindings (cfg = wasm32)
├── crates/                      ← sibling crates
│   ├── renkin-forward/          # forward reaction prediction (reactants → products)
│   └── renkin-kg/               # reaction knowledge graph builder (GraphML / Cypher export)
├── data/
│   ├── building_blocks.smi              # 402 curated commercial starting materials (loaded/deduplicated count)
│   ├── templates_extracted_5000.smi     # 5,000 auto-extracted SMIRKS templates
│   ├── benchmark_targets.smi            # internal benchmark set
│   └── bench_chunks/                    # USPTO-50k per-chunk results
├── scripts/
│   ├── extract_templates.py         # rdchiral template extraction pipeline
│   ├── run_benchmark_chunks.sh      # resumable chunked benchmark runner
│   ├── train_reranker.py            # candidate reranker training/evaluation (dev tool, offline only — see docs/guides/reranker-candidate-pools.md)
│   └── tests/                       # unittest suite for train_reranker.py
├── docs/                # MkDocs source → kent-tokyo.github.io/renkin/
└── mkdocs.yml

Roadmap

Recently shipped

• apply_retro/run_reactants performance regression resolved — chematic moved from a narrow git-pinned fix to the published 0.8.0 release (upstream automorphism-orbit-pruned canonicalization, chematic#193); on a fixed 30-target gate, measured in one session against current master: total elapsed 34.7% faster, p95 33.8% faster, and the single worst-case target 42.2% faster (confirmed via repeated isolated measurement, not a one-off run). Zero correctness change (apply_retro call counts identical across versions)
• renkin-forward CLI hardening — versioned ForwardPredictionReport, deterministic candidate IDs/merge/provenance, reactant-order-independent matching (up to 3 reactants), strict CLI/route-JSON validation
• RETROSPECT-inspired offline candidate-reranking foundation — candidate proposal/selection separation, feature schema v1, manifest v2, leakage-safe train/val/test splitting, 7 deterministic baseline arms + trained-ranker arm, paired bootstrap + offline gate tooling (#59; foundation only — no trained model or accuracy result yet, not wired into route search)
• Stable template_id (rule:<name> / smirks-sha256:<hex>) + --template-metadata evidence sidecar + renkin template ids (#41 phase 1)
• Substrate-specific examples (schema_version: 2) — per-step exact-substrate vs. same-template-different-substrate resolution, surfaced in --format explain and as match_kind in JSON (#41 phase 2)
• Deterministic ORD (Open Reaction Database) evidence import — offline renkin evidence match exact-set batch template matcher + scripts/ord_evidence_audit.py audit/converter into schema_version: 2 sidecars; no network access, no fuzzy matching, ambiguous/unprovenanced records excluded and counted in an audit report rather than guessed at (#41 phase 3A)
• renkin-bench cascade — multi-stage search (fast defaults → hard cases re-run deeper); only unsolved targets propagate to later stages. 78.0% → 95.9% on USPTO-50k
• renkin-bench --failure-taxonomy — classify unsolved targets by cause (beam limit / depth limit / template gap / stock near-miss)
• Graph-based ester cleavage — BFS-leakage-free R-C(=O)-O-R' → RCOOH + R'OH
• --top-templates N — frequency-rank filter: use the top-N most frequent templates for speed / less noise
• raw / validated / practical solved-rate metrics (--plausibility --practical-max-steps N)
• Retro cache hit-rate in SearchStats + --verbose

In progress

• Template retrieval index (element bitmask + bond-center prefilter) for the 50k template set
• Calibrated route confidence (map success_probability to empirical solve rate)

Next

• Graph rule expansion — sulfonamide / carbamate / urea cleavage (one PR per family, each with benchmark delta)
• Stock-aware planning (price / hazard / availability re-ranking)

Citation

If you use RENKIN in academic work, please cite it — see CITATION.cff for the canonical, version-tracked citation record. GitHub's "Cite this repository" button (top of the repo page) reads it directly and can export APA or BibTeX on demand.

Security

Report vulnerabilities via GitHub Private vulnerability reporting. See SECURITY.md.

License

MIT

GitHub Topics: retrosynthesis cheminformatics wasm rust drug-discovery casp synthesis-planning computational-chemistry

If RENKIN saves you time, a GitHub star helps others discover it.