De novo molecular generation with optical property preconditioning at the token level

2026-06-06 · Source: Machine Learning · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Computational Chemistry & Materials Informatics · Depth: Expert, quick

Summary

A study benchmarks a token-conditioned autoregressive language model, specifically a GPT2 variant, for de novo OLED molecular generation in a low-data environment. The model is pretrained on extensive chemical corpora, enhanced with discrete property tokens, and fine-tuned using multi-task optimization. It targets vertical absorption energy and oscillator strength, incorporating the HOMO-LUMO gap as an auxiliary descriptor. Evaluation via TDDFT reveals the generated library reproduces the training distribution's optical-property support, trending towards lower molecular weight. While token-level control is directional, it shows local irregularities and isn't fully orthogonal. Controllability strongly depends on local electronic environments; for instance, conjugated aromatic-carbon motifs improve target satisfaction, whereas aryl nitriles exhibit systematic red-shifting and reduced control. This work establishes a quantitative benchmark and emphasizes assessing model reliability within chemically meaningful subspaces.

Key takeaway

For research scientists developing generative models for novel material design, this work underscores the critical need to assess model reliability beyond aggregate property distributions. You should implement chemotype-resolved analyses to understand how local electronic environments influence conditional control, especially when generating OLED molecules. This approach will help you identify and address specific chemical motifs that exhibit reduced controllability or systematic biases, ultimately improving the precision and utility of your molecular generation efforts.

Key insights

Token-conditioned autoregressive models can generate OLED molecules with targeted optical properties, but control varies by chemical motif.

Principles

• Generative model reliability requires chemically meaningful subspace assessment.
• Low-data regimes benefit from pretraining on large chemical corpora.
• Multi-task optimization enhances fine-tuning for molecular generation.

Method

Pretrain a GPT2 model on chemical corpora, augment with discrete property tokens, then fine-tune using multi-task optimization to condition on specific optical properties like absorption energy and oscillator strength.

In practice

• Use TDDFT for evaluating generated molecule fidelity.
• Incorporate HOMO-LUMO gap as an auxiliary electronic descriptor.
• Analyze controllability based on local electronic environments.

Topics

• De novo molecular generation
• OLED design
• Optical properties
• GPT2 models
• Chemical motifs
• TDDFT evaluation

Best for: AI Scientist, Research Scientist

Related on AIssential

• Sample Efficient Generative Molecular Optimization with Joint Self-Improvement — Joint Self-Improvement optimizes molecular design by integrating generative and predictive models to overcome sample scarcity and distribution shift.
• LEAP: A closed-loop framework for perovskite precursor additive discovery — Coupling a domain-specialized LLM with active learning and expert review accelerates perovskite additive discovery under low-data conditions.
• Agentic Molecular Recovery via Molecule-Aware Exploration — Invalid molecular drafts require identity-preserving recovery, not just validity repair, to maintain structural cues and chemical identity.
• De novo design of peptides localizing at the interface of biomolecular condensates — A computational pipeline enables de novo design of peptides targeting biomolecular condensate interfaces.
• Order Matters in Retrosynthesis: Structure-aware Generation via Reaction-Center-Guided Discrete Flow Matching — Atom ordering in neural representations can encode chemical knowledge, improving retrosynthesis model efficiency and accuracy.
• MōLe-Λ: Learning the Coupled-Cluster Response State for Energies, Gradients, and Properties — MōLe-Λ efficiently predicts full CCSD response states by learning both right- and left-hand amplitudes, expanding accessible quantum chemistry properties.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Machine Learning.