Research map / Probabilistic and scientific modeling / Molecules and proteins

Beyond Drug Discovery: The Nanotechnology Molecular Optimization (NMO) Benchmark

NeurIPS 2026 · Probabilistic and scientific modeling › Molecules and proteins · approach: Structure-based 3D molecule generation · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces the NMO benchmark of three quantum-simulation-scored nanotechnology molecular design tasks (phonon transport, thermoelectrics, optomechanics) with a strict protocol, and a baseline extending Genetic GFN with Graph Group SELFIES and synthetic pretraining.

Problem

Generative molecular design benchmarks rely on simple proxy oracles and pharmaceutical pretraining, which do not transfer to nanotechnology objectives involving molecules bound to gold electrodes under hard structural constraints.

Approach

Replace proxy oracles with xTB-based quantum transport and optomechanics simulations under a strict protocol (single configuration, fixed 10k evaluation budget, fixed fragment library), and propose Graph Group SELFIES (a valid-by-construction fragment DAG encoding with native anchor modeling) plus synthetic random-graph pretraining to remove pharmaceutical bias, applied within an extended Genetic

Key results

• Full method achieves RI 5/5 on TE, 4/5 on PH and 5/5 on MO, with mean top-10 TE fitness 1.19 and PH fitness 0.75 across five seeds.
• f-RAG and GenMol score near zero on all three NMO tasks, while molGA finds relevant MO candidates in 5/5 seeds.
• The baseline finds a candidate exceeding the ZT>3 threshold for thermoelectrics and surpasses previous optomechanics best P=7.88.

Surprising findings

• Simple molGA outperforms advanced SOTA methods f-RAG and GenMol on the NMO optomechanics task.
• Auxiliary descriptors in the full method degrade MO performance, while ZINC pharmaceutical pretraining helps on MO.

Limitations

• xTB calculations may overestimate upconversion intensities in the high-fitness tail; DFT checks give smaller values.
• PH task has a degenerate regime where long molecules yield artificially low conductance, requiring length bounds.
• Only five seeds due to computational cost, limiting statistical power.

Experimental setup

Datasets: synthetic random graph dataset (300,000 graphs) ZINC (for REINVENT and Genetic GFN ablations)

Benchmarks: NMO TE task NMO PH task NMO MO task PMO

Metrics: Top-10 AUC mean top-10 fitness mean top-10 SA score relevance indicator

Baselines: f-RAG GenMol molGA REINVENT Genetic GFN

Builds on: Genetic GFN Group SELFIES PMO benchmark REINVENT molGA

Closest papers by meaning

• $\nabla^2$DFT: A Universal Quantum Chemistry Dataset of Drug-Like Molecules and a Benchmark for Neural Network Potentials (NeurIPS 2024 · similarity 0.86)
• Efficient Evolutionary Search Over Chemical Space with Large Language Models (ICLR 2025 · similarity 0.86)
• Molecular Optimization Model with Patentability Constraint (AAAI 2024 · similarity 0.86)
• Enhancing Ligand Validity and Affinity in Structure-Based Drug Design with Multi-Reward Optimization (ICML 2025 · similarity 0.85)
• From Artificially Real to Real: Leveraging Pseudo Data from Large Language Models for Low-Resource Molecule Discovery (AAAI 2024 · similarity 0.85)
• A Sober Look at LLMs for Material Discovery: Are They Actually Good for Bayesian Optimization Over Molecules? (ICML 2024 · similarity 0.85)