DriftingMol: Decoder-Coupled Drift for One-Pass Property-Conditional Molecular Generation

2026-05-23

AI Summary

This work addresses the high cost and limited diversity in continuous-property-conditioned molecular generation by proposing DriftingMol, a two-stage framework. It first constructs a beta-VAE in the SELFIES latent space and then leverages a frozen decoder’s feature maps to guide a DiT generator, enabling single-pass generation of molecules satisfying target properties. The key innovation lies in the novel decoder-coupled drifting mechanism, which employs decoder feature gradients to construct a pullback metric aligned with molecular decoding, substantially enhancing property control. Experiments on ZINC250K demonstrate a Spearman correlation coefficient of 0.510 for QED (with 94.7% uniqueness) and an average correlation of 0.598 under four simultaneous property constraints, significantly outperforming existing drifting-based methods.

Abstract

Property-conditional molecular generation should produce valid, diverse molecules while responding to continuous target values at low sampling cost. We introduce DriftingMol, a two-stage framework that adapts drifting models to a SELFIES latent molecular space. A frozen SELFIES beta-VAE provides the latent space, and the hidden representation of its decoder serves as the drift feature map. In decoder-coupled drift, decoder weights remain fixed, but drift gradients are backpropagated through the decoder feature map to a DiT generator, inducing a pullback metric aligned with molecular decoding. On ZINC250K, the default setting achieves QED Spearman correlation 0.493 with 94.7% uniqueness, while the strongest decoder-coupled condition reaches 0.510. Under protocol-matched four-property conditioning, decoder-coupled drift reaches mean Spearman correlation up to 0.598. Across 15 controlled variants, models that preserve the gradient path through decoder features achieve higher correlations than the tested latent-space, random-feature, and external-feature drift variants, while detached or stop-gradient decoder controls yield near-zero QED correlation and very low uniqueness. These results indicate that decoder-coupled drift is a useful low-cost mechanism for property-biased molecular generation, requiring one generator evaluation and one frozen decoder pass.

Problem

• property-conditional molecular generation
• molecular generation
• continuous target values
• low sampling cost
• valid and diverse molecules

Innovation

• decoder-coupled drift
• property-conditional generation
• SELFIES latent space
• pullback metric
• molecular generation