IBEX

Three-dimensional generative models for structure-based drug design face a hard data ceiling: fewer than 200,000 experimentally determined protein-ligand complexes are public, three orders of magnitude short of the corpora behind comparable vision and language models. Under that scarcity, pocket-conditioned generators learn training-set biases rather than transferable geometric priors, and the failure shows up as poses insensitive to pocket mutagenesis or polarity inversion.

IBEX — Information-Bottleneck-EXplored — attacks the problem from the data side rather than the architecture side. Researchers at Shenzhen University, University of Nottingham Ningbo China, UC San Diego, and Changzhou University use PAC-Bayesian information-bottleneck theory to quantify the usable information each training sample carries under three masking regimes: side-chain completion, de novo generation, and scaffold hopping. The analysis yields a clear ordering — scaffold hopping is the most information-dense, with the lowest gradient-SNR variance, the earliest capacity-compression phase, and the smallest train-test gap — and predicts that training under the most geometrically constrained task gives the model greater effective capacity and better transfer.

Acting on that finding requires no new network. IBEX keeps the TargetDiff architecture and hyperparameters intact, changes only what the model is asked to reconstruct, and adds one deterministic post-hoc step: an L-BFGS optimization that refines each generated conformation against a physics-based energy in under a second.

#Key Features

• Information-bottleneck task selection: PAC-Bayesian analysis ranks masking strategies by information density and identifies scaffold hopping as the one that tightens the test-risk bound, replacing intuition about curriculum design with a measurable criterion.
• Unmodified generator backbone: The diffusion generator retains TargetDiff's architecture and hyperparameters, so the gains are attributable to the training task and the refinement step rather than to added capacity.
• Sub-second physics refinement: An L-BFGS search over six rigid-body degrees of freedom minimizes a differentiable surrogate of binding free energy built from five short-range contact terms, completing in under one second per conformation.
• Non-destructive acceptance rule: A refined pose is kept only when its energy improves on the generator's original placement, preventing noisy finite-difference gradients from degrading good poses.
• Zero-shot pocket generalization: One trained checkpoint covers 100 held-out receptors with no per-pocket retraining.

#Technical Details

The generator is an SE(3)-equivariant diffusion model over ligand atom coordinates and atom types, with Gaussian forward noise on coordinates and categorical noise on types, following the TargetDiff formulation. For each protein-ligand pair a spatial mask selects the atoms to regenerate while the remainder form the context, and the choice of mask defines the training task; protein coordinates are weakly perturbed with Gaussian noise of standard deviation 0.1 Å as a regularizer. Ancestral sampling then produces a coarse pose whose heavy atoms fall within a 10 Å sphere centered on the pocket.

Refinement optimizes a six-vector of translation and axis-angle rotation parameters against a weighted sum of five empirical contact terms — two Gaussian attraction terms, a hard-sphere repulsion penalty, a hydrophobic term, and a hydrogen-bond term — with weights fixed across all experiments. Gradients are approximated by finite differences with a step size of 1e-3 after scale normalization, and the L-BFGS search runs at a fixed learning rate of 0.1. Evaluation follows the CBGBench protocol on CrossDocked2020, which constructs its side-chain and scaffold tasks only after an independent train/test partition to prevent label leakage. Across 100 held-out pockets, IBEX raises zero-shot docking success from 53% to 64%, improves mean Vina score from -7.41 to -8.07 kcal/mol, and achieves the best median Vina energy on 57 of 100 pockets against 3 for the unmodified TargetDiff baseline. QED improves by 25%, and the pipeline leads the compared generators on validity, uniqueness, and diversity.

#Applications

The pipeline serves structure-based hit generation and scaffold hopping: given a resolved or modeled binding pocket, it proposes chemically valid three-dimensional ligands already posed inside the site, so downstream docking or free-energy work starts from a physically reasonable conformation. Scaffold hopping is doubly relevant here — it is both the training task and a standard medicinal-chemistry objective, where the goal is a new core that preserves binding while escaping intellectual-property or ADMET liabilities. Higher QED alongside improved Vina energies matters for teams that need candidates plausible enough to synthesize, not just poses that score well.

#Impact

IBEX's broader claim is methodological: in a data-limited regime, choosing what the model reconstructs can matter more than how large the model is, and information-bottleneck analysis gives a principled way to make that choice. Holding the TargetDiff backbone fixed makes the argument cleanly, since the gains cannot be attributed to added parameters, and the refinement step is architecture-agnostic enough to bolt onto other pocket-conditioned generators. Two caveats bound the result: all evaluation is in silico, resting on Vina scores and cheminformatics metrics rather than measured affinities, and the work is an arXiv preprint with no public code or checkpoint release.