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

Technical Summary: De Novo Molecular Generation with Optical Property Preconditioning at the Token Level

Problem Statement
The discovery of organic light-emitting diode (OLED) molecules with targeted optical properties is hindered by the scarcity of high-quality experimental data and the limited reliability of conditional control in generative models across diverse chemical motifs. While optical informatics has advanced in predicting properties from molecular structures, the inverse design problem—generating new molecules with specific absorption energies and oscillator strengths—remains constrained. Existing approaches often rely on large datasets, lack high-fidelity electronic-structure validation (e.g., Time-Dependent Density Functional Theory, TD-DFT), or assess success only through aggregate distributions rather than chemical subspaces. This paper addresses the challenge of conditional molecular generation in a realistic low-data regime, focusing on whether token-level conditioning can reliably steer the generation of OLED candidates toward specific optical targets evaluated at the TD-DFT level.

Methodology
The authors propose a token-conditioned autoregressive language model pipeline based on GPT-2, structured in three distinct training stages:

1. Chemical-Language Pretraining: The model is pretrained on the ChEMBL library to learn SMILES grammar and valid molecular structures without property labels.
2. Property Pretraining: The model is trained on a large computational OLED dataset (460,205 molecules) where continuous optical properties are discretized into tokens. Specifically, vertical absorption energy ($\Delta E_{S0 \to S1}$), oscillator strength ($f_{S0 \to S1}$), and the HOMO–LUMO gap are binned and prepended to the SMILES strings as special tokens. Positional embeddings are zeroed at these token positions to treat them as global conditioning signals rather than sequential elements.
3. Multi-Task Fine-Tuning: The model is fine-tuned on a curated, smaller dataset (~41,500 molecules) with consistent, high-quality TD-DFT labels (computed at the $\omega$B97X-D3/def2-SVPD level). This stage employs Nash Multi-Task Learning to balance the optimization of multiple property objectives and prevent catastrophic forgetting.

Generation and Validation
Candidates are generated by sampling conditioned SMILES sequences across a grid of absorption and oscillator strength tokens. The generated library undergoes canonicalization and filtering for validity and uniqueness. Crucially, the final evaluation is performed at the TD-DFT level (using the $\omega$B97X-D3 functional and def2-TZVP basis set) to assess electronic-structure observables, rather than relying solely on token compliance or proximity to the training distribution.

To analyze controllability, the authors employ the Overlapping Fragment Molecular Representation (OFraMP) to resolve local electronic environments. This allows for a chemotype-resolved analysis of how specific local motifs influence the model's ability to satisfy joint target constraints.

Key Results

• Global Distributional Fidelity: The generated library reproduces the dominant optical property support of the training distribution (spanning near-IR to UV) but exhibits a structural shift toward lower molecular weights and fewer heavy atoms. The model remains on the optical-property manifold while exploring structurally more compact scaffolds.
• Token-Level Controllability: Conditioning is directionally effective; median absorption energy and oscillator strength generally increase with higher token indices. However, control is not fully orthogonal. Conditioning on one property influences the other, leading to calibration errors. The generated set shows a stronger anti-correlation between absorption energy and oscillator strength than the training set, consistent with the energy gap law, particularly in the low-energy (red-shifted) regime.
• Chemotype-Dependent Reliability: Controllability is highly dependent on local electronic environments:
  • Enriched Motifs: Moderately conjugated aromatic-carbon motifs (e.g., radius-2 aromatic carbon environments) show improved joint target satisfaction (2.65x lift over baseline). These environments appear to occupy a smoothly tunable electronic regime.
  • Suppressed Motifs: Electron-withdrawing motifs, particularly aryl nitriles, exhibit systematic red-shifting and reduced controllability. Molecules containing these motifs frequently fail to meet joint targets, showing a mean absorption bin shift of +1.75 relative to the target. This is attributed to strong local electronic perturbations that exceed the resolution of the discrete bin conditioning.

Significance and Claims
The paper establishes a quantitative benchmark for conditional OLED molecular generation, demonstrating that while token-level optical preconditioning is effective at a coarse level, its reliability is not uniform. The authors claim that model reliability must be assessed in chemically meaningful subspaces rather than through aggregate property distributions alone.

The study highlights that the success of discrete property conditioning depends on whether local electronic perturbations remain within the effective resolution of the conditioning scheme. Consequently, the paper reframes conditional molecular generation as a problem requiring chemotype-resolved analysis to identify where control is robust and where it degrades, rather than a purely generative task. The work does not claim to have solved the inverse design problem universally but provides a rigorous framework for evaluating and understanding the limitations of current generative models in low-data, high-stakes chemical domains.