IonBind: RL-aligned Generative Cascade for Selective Ion Binder Design

Selective complexation of a target ion underlies applications from sensing, ion separation, and recycling to biomedicine, where performance depends on relative rather than absolute affinity among ions that differ only subtly in size and charge. Selectivity is relative, non-differentiable and multi-objective, with no per-structure label, rendering supervised generative approaches inaccessible, while scoring and ranking are often left as post-hoc steps, disconnected from generation. IonBind closes that loop as a modular two-step generative cascade for de novo ligand design, coupling an autoregressive transformer for discrete composition with a diffusion model for continuous geometry. To steer the generation toward the target, we post-train the two policies jointly via reinforcement learning. We propose an RL environment that scores the stability constants ($\log K$) of each generated ligand candidate against the target ion (Li$^{+}$) and its competitors (Ca$^{2+}$, Na$^{+}$, K$^{+}$), resulting in the generation of ligands with high selective affinity for Li$^{+}$. Experiments show that joint post-training of both policies outperforms single-policy post-training. This modular generative cascade resolves contributions to selectivity and extends inverse molecular design to relative targets with no first-principles rule to narrow a design space too vast for heuristic search or trial-and-error experimentation.