Reward Transport Controls Molecular Properties in Flow Matching

This paper introduces "Reward Transport," a novel technique for controlling specific properties of generated molecules within flow matching models. Traditionally, the coupling mechanism in flow matching, which pairs noise vectors with data points, is viewed primarily as a computational choice. However, this research re-frames it as an alignment interface. By matching noise and data according to a desired molecular property, the method embeds controllable structure directly into the learned flow field. Reward Transport leverages optimal transport coupling during the training phase to align a scalar coordinate in the noise space with molecular rewards. This alignment enables a powerful capability during inference: by simply varying this noise-space coordinate, the generated molecular distribution can be steered to exhibit desired properties, such as logP or QED. Crucially, this control is achieved without needing an external oracle, a separate reward model, gradient guidance, or additional computational overhead during generation. The approach offers a principled, continuously adjustable "distribution-level control knob." Empirical results on standard datasets like ZINC-250K and GuacaMol demonstrate monotone control over logP and consistent control over QED. Interestingly, the same control knob produces opposite structural responses for different targets (e.g., growing molecules for logP but shrinking for QED), confirming it's not merely a generic size bias. This method is also complementary to existing techniques like classifier-free guidance.