Scaffold extraction from molecular structures

scaffold-extraction-from-molecular-structures

License: restricted — no clear open-source license detected for the underlying tool; verify licensing before commercial use or redistribution. <!-- asb-license-banner -->

Summary

Extract chemical scaffolds from natural product molecular structures to isolate conserved core frameworks that encode evolutionary lineage and biosynthetic ancestry. This isolates scaffold-derived patterns for contrastive learning in pre-training natural product foundation models.

When to use

When pre-training or fine-tuning a molecular representation model on natural products and you need to encode scaffold-derived evolutionary patterns as a distinct learning signal. Trigger: input is SMILES strings or molecular graphs from natural product databases (e.g., LOTUS, custom biosynthetic collections) and your downstream task includes taxonomy classification, gene-level or microbial-level evolutionary inference, or virtual screening where scaffold conservation reflects biosynthetic ancestry.

When NOT to use

• Input molecules are purely synthetic (non-natural products); conventional scaffolding lacks discriminative power for evolutionary lineage.
• Downstream task is chemical property prediction (logP, solubility) rather than evolutionary or biosynthetic inference; scaffold features are not optimized for these targets.
• Molecular graph already includes explicit side-chain encoding and you are bypassing explicit scaffold extraction; dual objectives may over-regularize the model.

Inputs

• SMILES strings (CSV or pickle format, as in pretrain_smiles.pkl)
• Molecular graphs (PyG Data objects or rdkit mol objects)
• Natural product molecular structure database (e.g., LOTUS, ONTOLOGY, custom BGC-derived collections)

Outputs

• Scaffold representations (Murcko scaffolds as SMILES or graph embeddings)
• Side-chain information (atomic/bond features isolated from scaffold)
• Scaffold-annotated molecular graph dataset (input to contrastive learning objective)
• Scaffold similarity matrix or contrastive pairs for pre-training

How to apply

Parse input SMILES strings into molecular graphs, compute Murcko scaffolds (core ring systems and linkers with all side chains removed) to isolate evolutionary-conserved backbones. Extract side-chain information separately to preserve biosynthetic diversity. Integrate scaffold representations into the contrastive learning objective alongside full-molecule representations, such that molecules sharing scaffolds are pulled into proximity in embedding space. The rationale is that natural products exhibit strong scaffold clustering by organism and biosynthetic pathway; explicitly encoding scaffold similarity improves the model's capacity to learn evolutionary information at both gene and microbial levels (as validated in NaFM's taxonomy classification benchmarks). Scaffold and side-chain information are complementary: contrastive learning on scaffolds captures evolutionary commonality while masked graph modeling on full structures preserves biosynthetic diversity.

Related tools

• RDKit (Compute Murcko scaffolds from SMILES and convert molecules to graph representations)
• PyTorch Geometric (PyG) (Represent scaffold and side-chain information as graph nodes/edges; construct batches for contrastive learning)
• NaFM pre-training framework (Integrate scaffold extraction and side-chain encoding into dual contrastive + masked modeling pipeline) — https://github.com/TomAIDD/NaFM-Official

Examples

python scripts/setup_data.py; python train.py --conf examples/Pretrain.yml

Evaluation signals

• Scaffold extraction preserves ring systems and linkers while removing all side chains; verify by comparing Murcko scaffold SMILES to original SMILES (scaffold should be a substructure).
• Contrastive pairs: molecules with identical scaffolds but different side chains should have high cosine similarity in scaffold embedding space; validate via k-nearest neighbor analysis on scaffold embeddings.
• Downstream taxonomy classification accuracy on held-out natural products improves when NaFM-pretrained scaffold representations are used vs. models pre-trained on synthetic molecules (as reported in paper benchmarks).
• Scaffold clustering in learned embedding space correlates with known biosynthetic pathways or organism taxonomy (gene/microbial-level analysis signal).
• Ablation: remove scaffold extraction and re-train; contrastive loss should be higher and downstream taxonomy accuracy should drop, confirming scaffold information is a material learning signal.

Limitations

• Murcko scaffold extraction assumes ring systems encode evolutionary lineage; linear natural products (no rings) yield trivial or empty scaffolds and may not benefit from this approach.
• Scaffold conservation is organism/pathway-dependent; scaffolds extracted from unrelated natural product families may not cluster meaningfully.
• Scaffold extraction is computationally expensive for large pre-training datasets; preprocessing (SMILES standardization, scaffold generation) must be run once and cached (as in NaFM's filter.py step).
• Side-chain diversity is lost in scaffold representation; models must use both scaffold and full-molecule objectives to retain biosynthetic details.

Evidence

• [intro] Our method integrates contrastive learning with masked graph modeling, effectively encoding scaffold-derived evolutionary patterns alongside diverse side-chain information: "Our method integrates contrastive learning with masked graph modeling, effectively encoding scaffold-derived evolutionary patterns alongside diverse side-chain information"
• [other] Identify and document the scaffold extraction and side-chain encoding mechanisms used to represent evolutionary and structural diversity.: "Identify and document the scaffold extraction and side-chain encoding mechanisms used to represent evolutionary and structural diversity."
• [intro] Through detailed analysis at both gene and microbial levels, NaFM reveals a strong capacity for learning evolutionary information: "Through detailed analysis at both gene and microbial levels, NaFM reveals a strong capacity for learning evolutionary information"
• [intro] conventional molecular representation techniques are not well-suited to the unique structural and evolutionary features of natural products: "conventional molecular representation techniques are not well-suited to the unique structural and evolutionary features of natural products"