SynOmega

SynOmega covers a range of prediction capabilities for organic small-molecule reactions, including single-step forward reaction prediction, single-step retrosynthesis, multi-step route planning, and synthesizability scoring.

Why three layers

The core of SynOmega is three decoupled layers, each committing to a single narrow interface, so that each can be replaced and evaluated on its own:

flowchart TD
    A["synthesizability scoring<br/>synthesizability"] --> B["multi-step search<br/>search (Retro* / MCTS / best-first)"]
    B --> C["single-step model<br/>single-step: product SMILES → candidate reactants"]
    A -.asks.-> A1["can the target be reached within N steps<br/>from purchasable materials?"]
    B -.on.-> B1["search for a route over the AND-OR graph"]
    C -.answers.-> C1["disconnect one step at a time"]

The layers connect only through a deliberately narrowed interface — a single-step backend only needs to implement predict(smiles, top_k) -> [Prediction] — so the planner and the scorer do not care whether the candidates come from a graph neural network, a Transformer, or pure template matching.

The single-step retrosynthesis, multi-step search, and synthesizability scoring above form the backbone of SynOmega; the other two capabilities share the same reaction templates and graph models:

• Single-step forward reaction prediction (forward): reactants → product, the "mirror image" of the retrosynthetic single-step model, reusing the same reaction template library.
• Reaction plausibility scoring (plausibility): assigns a reaction a score for "how chemically plausible it is".

Technology stack at a glance

Capability	Method core
Single-step retrosynthesis	D-MPNN neural template classifier (can also fall back to a pure template-rule backend, no torch required)
Single-step forward prediction	Mirror of the same D-MPNN template classifier + RDKit forward template application
Multi-step route planning	Retro* by default, also includes MCTS and best-first, all searching over the AND-OR graph
Synthesizability scoring	Continuous SynScore based on the "number of non-purchasable starting materials" in the optimal route
Reaction plausibility	Dual-tower D-MPNN, no atom mapping, comparing the reactant graph vs the product graph
Molecular identity	InChIKey used uniformly throughout (consistent deduplication, stock lookup, and caching across tools)

Documentation guide

• Feature Guide — one chapter per feature, covering installation and command-line / Python-API usage (matching the research-report chapters one-to-one).
• Research Report — chapter-by-chapter model/algorithm architecture, pseudocode, training sets, evaluation metrics, and figures for each capability module. The evaluation protocol, terminology, and data conventions are defined centrally in that chapter's Overview and referenced by each capability chapter.

Driving an AI agent (Claude Code / OpenClaw / …)?

Install the SynOmega skill and the agent can call all six capabilities directly: clawhub install synomega. Source and manual install: github.com/zbc0315/synomega-skill.

On the training-data conventions

All models described here are trained on a large-scale atom-mapped commercial reaction corpus. The raw reactions are not distributed with the software; what is released publicly are the trained model weights. See the Overview for the sampling conventions of each evaluation set (the ZINC purchasable building-block set, the ChEMBL target set).