README.md

metadata

title: Meta-LoRA Molecular Generator
emoji: ⚗️
colorFrom: yellow
colorTo: purple
sdk: gradio
sdk_version: 5.38.0
app_file: app.py
python_version: '3.11'
pinned: false
license: mit
short_description: Few-shot molecular generation via context-conditioned LoRA

Meta-LoRA Molecular Generator

Scaffold-Episodic Meta-Learning with Context-Conditioned LoRA for Few-Shot Molecular Generation.

Given a small support set of SMILES strings from the same scaffold family, the model generates novel, valid, drug-like molecules — zero gradient steps at inference.

Architecture

• BaseGrammarTransformer — 3.2M frozen params, pre-trained on ZINC250k
• EnhancedContextEncoder — GRU + Morgan FP fusion → constraint vector z
• ContextConditionedLoRA — rank-16 LoRA on Q and V projections, conditioned on z
• Training — Scaffold-episodic meta-training (Bemis-Murcko families)

Metrics (10-trial mean ± std, ZINC250k, 5-shot)

Metric	Value
Validity	96.8 ± 2.1%
Uniqueness	99.1 ± 0.8%
Novelty	98.3 ± 1.4%
Avg Tanimoto	0.4231 ± 0.038

Usage

Paste 3–10 SMILES from the same scaffold family into the support set box and click Generate.