# Layered-Lattice Molecular AI: Beatable Targets

Generated: 2026-05-31

## Thesis

SCBE should not try to beat the best molecular AI systems everywhere. The beatable target is narrower:

> Robust compound reasoning under topology loss, activity cliffs, scaffold leakage, and explanation requirements.

This matches the SCBE layered-lattice idea: keep multiple graphs/fields underneath the main molecular graph, with weights for atom identity, proton/electron counts, valence, fragments, descriptors, geometry, spectra, literature, and claim boundaries.

## What The Best Systems Are Good At

| System family | Strength | Why not compete head-on yet |
| --- | --- | --- |
| Schrödinger / enterprise modeling suites | Physics-based molecular modeling, enterprise drug/materials workflows. | Huge expert-built stack; beat by bridging/provenance, not by reimplementing. |
| RDKit/OpenEye/Open Babel/CDK | Cheminformatics parsing, descriptors, fingerprints, file formats, substructures. | Use them as engines; SCBE adds orchestration and explanation. |
| ChemCrow/Coscientist-style agents | LLM + chemistry tools + planning; Coscientist connects to lab automation. | SCBE should stay computational-only unless expert-reviewed lab integrations exist. |
| GNN/graph-transformer molecular ML | Learned property prediction and molecular representation. | Strong on some benchmarks, but vulnerable to data splits, activity cliffs, and interpretability issues. |
| TDC/MoleculeNet/GuacaMol/MOSES | Public benchmark ecosystems. | Use these to measure claims; do not call local fixtures official scores. |

## Current Weak Spots In The Field

1. **Activity cliffs**
   - Small structural changes can cause large activity/property changes.
   - Smooth learned embeddings often struggle because they assume nearby structures have nearby behavior.
   - This is a direct fit for a multi-field card system that tracks small local differences separately from global similarity.

2. **Topology loss**
   - Atom counts/formulas alone do not preserve molecular identity.
   - Example: ethanol and dimethyl ether both reduce to C2H6O.
   - SCBE already has a benchmark showing recomposition after an atom-mud step requires descriptors/fragments, not atom bags alone.

3. **Scaffold leakage and split weakness**
   - Scaffold splits can overestimate real generalization because "different" scaffolds may still be chemically similar.
   - SCBE can add explicit split receipts and similarity checks across multiple fields.

4. **Fingerprints remain hard to beat**
   - Classical fingerprints/descriptors are still competitive with deep models on many molecular prediction problems.
   - SCBE should treat fingerprints as one lattice, not a primitive baseline to ignore.

5. **Explainability**
   - Molecular GNN explanations are still an active research problem.
   - SCBE can win by producing transparent move histories: which cards preserved identity, which were lost, which decided recomposition.

## SCBE Layered-Lattice Model

Represent each compound as several connected fields:

```text
main graph: atoms + bonds
  -> atom bag lattice: element counts, isotope/proton/neutron/electron state when available
  -> descriptor lattice: MW, LogP, TPSA, HBD/HBA, ring counts, fragment counts
  -> fingerprint lattice: Morgan/ECFP, MACCS, topology fingerprints
  -> fragment lattice: SMARTS/substructures, functional groups, pharmacophore cards
  -> electron field: partial charges, HOMO/LUMO/density proxy when engines exist
  -> geometry field: conformers, distances, angles, shape/electrostatics
  -> literature field: PubChem/ChEMBL/BindingDB/papers/patents
  -> governance/proof field: safety decision, source, hash, claim boundary
```

The value is not one graph. The value is a reversible and auditable weaving of fields.

## First Things We Can Try To Beat

### 1. Atom-Mud Recomposition

Already implemented:

```bash
node packages/cli/bin/scbe.js bench compound-decompose --json
```

Goal: expand from 3 cases to 100 known pairs/classes where formulas are ambiguous and identity requires fragments/descriptors.

Win condition:

- atom-only baseline fails or is ambiguous,
- SCBE layered-lattice recomposition recovers the known compound,
- receipt explains which fields mattered.

### 2. Activity-Cliff Pair Explanation

Build a fixture from known matched molecular pairs:

```text
pair A/B -> high structural similarity -> large property/activity delta
```

Win condition:

- SCBE identifies the local changed fragment,
- explains why global similarity is misleading,
- labels the result as cliff-risk instead of smoothing over it.

### 3. Scaffold-Leakage Detector

Given train/test molecule sets:

- compute scaffold split,
- compute fingerprint similarity across split,
- flag pairs that are technically split but chemically near.

Win condition:

- SCBE catches leakage-like pairs that a simple scaffold report misses.

### 4. Multi-Field Candidate Ranking

Given a compound description and candidate list:

- rank by formula, fragments, descriptors, fingerprints, and literature aliases.

Win condition:

- SCBE ranks the known target above formula-matched or graph-near decoys,
- produces field-by-field explanation.

### 5. Fresh-Agent Reproduction

Give a fresh weak model only:

```bash
scbe bench compound-decompose --json
```

Win condition:

- it can explain the atom-mud ambiguity and recover the evidence without hidden context.

### 6. Bijective Reaction Round Trips

SCBE should treat bijective reactions as a proof surface:

```text
compound graph -> decomposition cards -> transform cards -> recomposition -> same canonical identity or declared new identity
```

A reaction is "bijective" only inside a declared representation. For example:

- canonical SMILES -> RDKit molecule -> canonical SMILES can be bijective enough for identity if stereochemistry/isotopes are preserved;
- molecule -> atom bag is not bijective, because topology is lost;
- molecule -> fingerprint is not bijective, because many molecules can collide or share features;
- reaction path -> products can be non-bijective if leaving groups, solvent, conditions, or stereochemistry are omitted.

Win condition:

- SCBE can state which transformation is reversible, which is lossy, and which extra cards restore identity.
- The benchmark must include both passing reversible paths and intentionally non-bijective paths.

## Next Implementation Branches

1. `feat(chem): add activity cliff fixture`
   - Use RDKit fingerprints/descriptors.
   - Synthetic or public known pairs first.
   - Produce local benchmark, not public leaderboard claim.

2. `feat(chem): expand atom-mud recomposition corpus`
   - Add formula-isomer groups from RDKit-valid molecules.
   - Compare atom-only vs layered-lattice recovery.

3. `feat(chem): add scaffold leakage detector`
   - Input two molecule sets.
   - Output scaffold split report plus cross-field similarity warnings.

4. `feat(chem): pubchem evidence bridge`
   - Resolve name -> CID/SMILES/InChIKey.
   - Attach source URLs and input/output hashes.

5. `feat(chem): bijective reaction harness`
   - Round-trip canonical SMILES, formula, fragments, fingerprints, and atom bags.
   - Score each transformation as `BIJECTIVE`, `LOSSY_RECOVERABLE`, or `LOSSY_AMBIGUOUS`.
   - Add reaction-style transforms after identity round trips are proven.

## Sources

- Schrödinger platform: https://www.schrodinger.com/platform/
- RDKit: https://rdkit.org/
- OpenEye OEChem TK: https://www.eyesopen.com/medchem-tk
- ChemCrow / LLMs with chemistry tools: https://www.nature.com/articles/s42256-024-00832-8
- Coscientist autonomous chemical research: https://www.nature.com/articles/s41586-023-06792-0
- MoleculeNet: https://moleculenet.org/
- TDC: https://tdcommons.ai/start
- MoleculeNet paper: https://pmc.ncbi.nlm.nih.gov/articles/PMC5868307/
- Activity cliffs limitation benchmark: https://pmc.ncbi.nlm.nih.gov/articles/PMC9749029/
- Scaffold validation limitations: https://www.emergentmind.com/topics/scaffold-based-validation
- FunQG quotient graph limitations summary: https://pure.unileoben.ac.at/en/publications/funqg-molecular-representation-learning-via-quotient-graphs