Source-linked AI summary
SurfSpec: Enhancing Off-Target-Agnostic Specificity by Bounding Pocket-Ligand Geometric Mismatch
Minyeong Hwang, Yoorim Gang, Ziseok Lee, Wooyeol Lee, Young Bin Park, Jae-Mun Choi, Kyungsu Kim, Eunho Yang
TL;DR
Lead optimization must improve target binding without increasing unintended off-target interactions, yet affinity-driven methods do not explicitly control specificity and many specificity-aware methods require off-target information. The paper develops a target-only geometric mismatch analysis and SurfSpec, which grows ligands toward under-occupied target-pocket regions; on CrossDocked2020, it achieves higher empirical specificity than evaluated off-target-agnostic baselines while maintaining competitive target-pocket performance.
Problem
Affinity-driven lead optimization does not explicitly control specificity, while many specificity-aware approaches require prior off-target structures or undesired-target activity information.
Method
SurfSpec metricizes pocket–ligand surface mismatch, uses triangle-inequality bounds with geometry–affinity calibration, and alternates surface-directed pseudo-label generation with pocket-conditioned ligand-prior recovery.
Results
On CrossDocked2020, SurfSpec achieves higher empirical specificity than evaluated off-target-agnostic baselines while maintaining competitive target-pocket docking performance.
Takeaways & Limitations
Target-only geometric mismatch can support off-target-agnostic specificity optimization for geometrically separated off-target classes.
Takeaways & Limitations
The specificity lower-bound analysis assumes a quantile-calibrated geometry–affinity envelope over the empirical CrossDocked2020 ligand–pocket distribution.
Abstract
from arXiv · showhide
Lead optimization in structure-based drug design aims to improve target binding while avoiding unintended interactions with off-target pockets. However, existing affinity-driven methods do not explicitly control specificity, whereas current specificity-aware approaches commonly require prior knowledge of off-target structures. We address off-target-agnostic specificity-aware lead optimization by analyzing the geometric mismatch between a ligand and the target pocket. We provide a conservative specificity lower bound for geometrically separated off-targets without requiring access to off-target structures. By metricizing pocket--ligand mismatch, the triangle inequality shows that reducing target--ligand mismatch improves a conservative lower bound on mismatch to a separated off-target class, which can be translated into a specificity lower bound through an empirical geometry--affinity calibration. Motivated by this analysis, we introduce SurfSpec, an off-target-agnostic lead optimization framework that iteratively grows ligands toward under-occupied regions of the target pocket surface. SurfSpec alternates between linker generation toward selected target-surface patches, which provides geometric pseudo-labels, and refinement under a pocket-conditioned ligand prior, which restores these pseudo-labels into valid ligands. On the CrossDocked2020 test set, SurfSpec reduces geometric mismatch and outperforms evaluated off-target-agnostic lead optimization baselines in empirical specificity, while maintaining competitive target-affinity improvement.
1 Introduction
The paper frames specificity-aware lead optimization without off-target access as a target-only geometric problem and introduces SurfSpec to grow ligands toward under-occupied target-pocket regions. It reports higher empirical specificity than evaluated off-target-agnostic baselines while maintaining competitive target-pocket performance.
- Motivation: High target affinity alone may not ensure specificity because optimized ligands can also bind unintended pockets.The paper motivates specificity-aware optimization because off-target interactions can reduce selectivity and cause adverse effects.
- Motivation: The framework is motivated as an alternative to specificity-aware methods that rely on off-target pocket information or undesired-target activity labels.Those approaches incorporate explicit negative supervision but depend on prior off-target information.
- Geometric principle: A target-only geometric principle uses Jensen–Shannon surface mismatch, triangle-inequality bounds, and geometry–affinity calibration to estimate specificity against separated off-targets.This analysis avoids accessing off-target structures during optimization.
- SurfSpec: SurfSpec grows ligands toward under-occupied target-pocket surface regions by alternating surface-directed pseudo-label generation with pocket-conditioned ligand-prior recovery.The recovery step restores geometric modifications into valid ligands while maintaining consistency with the learned molecular distribution.
- Evaluation: On CrossDocked2020, SurfSpec achieves higher empirical specificity than evaluated off-target-agnostic baselines while maintaining competitive target-pocket docking performance.Further analyses report improved pocket coverage and geometric fit, with strong pseudo-label validity and fidelity.
2 Related Work
Related work spans target-conditioned lead optimization, geometric complementarity methods, and training-free sample refinement. SurfSpec differs by explicitly metricizing pocket–ligand surface complementarity and using a simpler refinement strategy for its ligand-growth pipeline.
- Lead Optimization and Molecular Specificity: Existing SBDD methods use autoregressive generation, equivariant diffusion, or controllable optimization for pocket-conditioned molecular generation and lead optimization.Examples include PMDM, DecompOpt, and Delete.
- Geometric Complementarity: Classical docking and modern SBDD models use pocket geometry or surface complementarity to identify poses and guide molecular generation, elaboration, or optimization.These approaches establish geometric conditioning as a recurring strategy in structure-based design.
- Geometric Complementarity: Other geometry-conditioned methods impose reference shapes, multimodal ligand profiles, learned geometric representations, or user-specified geometric conditions.The cited methods include Diff-Shape, ShEPhERD, GeoRCG, and UniGuide.
- Geometric Complementarity: Prior work does not explicitly formulate pocket–ligand surface complementarity as a metricized mismatch objective linked to molecular specificity.SurfSpec is oriented by this mismatch formulation.
- Training-Free Sample Refinement: Training-free refinement methods include reverse-path, variational-posterior, and consistency-oriented optimization or correction paradigms.These methods provide general mechanisms across inverse problems and continuous data domains.
3 Preliminaries
The paper defines specificity as target affinity relative to the strongest off-target and seeks a target-only alternative because the true off-target set is unavailable. It represents pocket–ligand geometry through surface-induced probability measures compared by Jensen–Shannon distance.
- Representations: A ligand is represented by atom coordinates and atom-type features, while a pocket is defined analogously; the methodology identifies both up to separate rigid transformations.The rigid-body abstraction is used for the methodology derivation.
- Problem Setup: Specificity is defined as the affinity gap between the target pocket and the strongest off-target pocket.Affinity is larger for stronger binding, and a specific ligand combines strong target affinity with weaker competing-pocket affinity.
- Problem Setup: Direct optimization of specificity is generally impractical because the true off-target set is unavailable during lead optimization.Target-only methods can improve off-target affinity, while predefined off-target sets may miss challenging pockets outside the collection.
- Geometric mismatch: Geometric mismatch compares the ligand surface with the ligand-oriented pocket surface through probability measures induced by signed-distance-based Boltzmann densities.The two surfaces and measures are illustrated in Figure 2.
- Geometric mismatch: The Jensen–Shannon distance measures discrepancy between the induced probability distributions, and its square-root form is a metric satisfying the triangle inequality.A smaller geometric mismatch indicates stronger agreement between the ligand and pocket surface fields.
4 Methodology
The methodology defines target-only geometric mismatch and uses it to guide iterative ligand growth toward under-occupied pocket regions, followed by prior-guided recovery into valid ligands.
- Geometric mismatch: Geometric mismatch is defined as the Jensen–Shannon distance between ligand- and pocket-induced probability measures on a shared ligand-local domain.The measures are constructed from signed-distance-based Boltzmann densities.
- Geometric mismatch: SurfSpec uses geometric mismatch as a target-only design principle for improving target-pocket surface fit without accessing off-target pockets during optimization.
- Pipeline: SurfSpec alternates geometric expansion and low-noise prior-guided recovery until the growth budget is reached or target-pocket Vina performance worsens.The low-noise reverse process approximately samples from the anchored clean distribution.
- Pseudo-label generation: The pipeline selects nearby under-occupied target-surface patches and generates linker-based pseudo-labels directed toward them.Pocket-clashing atoms within the 2 Å threshold are truncated from the generated linker structure.
- Pseudo-label refinement: A pocket-conditioned ligand prior refines each pseudo-label into a valid ligand while encouraging closeness to the geometric anchor.The anchored clean density combines the pretrained ligand prior with an anchor-fidelity term.
5 Experiments
Experiments on CrossDocked2020 evaluate SurfSpec against lead-optimization and refinement baselines, finding stronger empirical specificity alongside favorable target-pocket behavior and refinement quality.
- Benchmark setup: SurfSpec is evaluated on 100 CrossDocked2020 test complexes, starting from each reference ligand and generating one optimized molecule per method.The main implementation permits at most K = 3 ligand-growth iterations.
- Lead optimization: SurfSpec achieves the best empirical specificity among evaluated methods, including the highest average score and thresholded success rates.
- Lead optimization: SurfSpec achieves the best pocket occupancy, low target-pocket geometric mismatch, strong target-pocket Vina docking, and zero clash rate.These results indicate improved target-surface coverage while preserving favorable docking behavior.
- Baseline comparison: Repeated optimization and early stopping do not match SurfSpec, while naive size-extension baselines combine low geometric mismatch with high clash rates and weak empirical specificity.
- Specificity certificate: The certificate margin is positive for 65.96% of randomly sampled target–off-target pairs, making the specificity certificate non-vacuous for a substantial fraction of pairs.The certificate remains conservative.
- Refinement quality: Under shared pseudo-labels, SurfSpec provides strong prior consistency, low clash and repulsion measures, favorable target-affinity diagnostics, perfect valence validity, and reasonable pseudo-label faithfulness.
6 Conclusion
The conclusion presents SurfSpec as a target-only geometric approach to off-target-agnostic specificity-aware optimization, with favorable CrossDocked2020 specificity and docking results.
- Conclusion: The paper studies specificity-aware lead optimization when only the target pocket is available during optimization.
- Conclusion: Its analysis uses metricized pocket–ligand surface mismatch to show that reducing target mismatch weakly improves a conservative specificity lower bound for geometrically separated off-targets.
- Conclusion: SurfSpec expands ligands toward under-occupied target-pocket regions through surface-directed ligand growth.
- Conclusion: On CrossDocked2020, SurfSpec achieves higher empirical specificity, improves target-pocket occupancy, and maintains competitive docking performance with low geometric mismatch and zero clash rate.
- Conclusion: The related-method comparison distinguishes SurfSpec by jointly supporting inference-time geometric control, surface complementarity, target-only specificity, and OOD pseudo-label refinement without objective-specific retraining.
B.2 Proof of the Specificity Certificate
The specificity certificate combines metric geometry with a quantile-calibrated geometry–affinity envelope to provide a high-probability lower bound for separated off-targets.
- Calibration assumption: The geometry–affinity assumption treats mismatch as constraining the attainable affinity range with high probability while allowing non-geometric affinity variation.Residual variation includes electrostatics, hydrogen bonding, and desolvation.
- Geometric separation: A δ-separated off-target class remains at least δ away from the target pocket under the ligand-oriented surface metric throughout optimization.
- Geometric bound: The triangle inequality yields a lower bound on ligand–off-target mismatch from target mismatch and target–off-target separation.
- Specificity certificate: Theorem 1 converts this geometric lower bound into a high-probability specificity lower bound using the percentile-calibrated geometry–affinity envelope.
- Specificity certificate: Reducing target mismatch ϵ weakly improves the certified lower bound because it increases the separated off-target mismatch margin [δ −ϵ]+.
B.3 Low-Noise Analysis of Anchored Refinement
The anchored-refinement analysis models valid, pseudo-label-consistent ligands as a locally concentrated distribution and studies low-noise score-flow recovery under explicit assumptions.
- Anchored concentration: The anchored clean distribution is assumed to concentrate inside a small ball around its mean.This formalizes the joint constraints of pocket-conditioned validity and similarity to the pseudo-label.
- Mode recovery: Proposition 1 establishes mode recovery under anchored concentration, with the convergence point bounded near the forward mean.The bound is expressed as a distance of at most |ατ|r0 from the forward mean ατµanc.
- Approximate score flow: The low-noise analysis assumes the implemented DPS-style anchored score is locally close to the exact anchored score on the visited region.This local approximation supports analysis of the approximate score flow rather than only the exact flow.
- Gaussian surrogate: Proposition 3 bounds the Gaussian surrogate’s divergence from the anchored distribution using KL, Wasserstein, and total-variation controls.The total-variation result follows from Pinsker’s inequality, while the Wasserstein result uses a shared Gaussian-noise coupling.
- Re-noising: Under the stated assumptions, re-noising around the recovered point yields a controlled approximation, with the flow-error term vanishing at the exact mode.The re-noised point is defined from the approximate flow output with Gaussian perturbation.
C.1 Implementation Details for Ligand Growth and Refinement
The implementation uses a shared iterative ligand-growth loop, surface-directed linker proposals, and pocket-conditioned refinement, with controls for baseline, surrogate, and refinement comparisons.
- Comparison baselines: Baseline variants test repeated application and early stopping, while DiffSBDD-SizeExt tests naive atom addition without controlled surface selection.ActivityDiff supplies approximate negative guidance from randomly selected targets as a surrogate off-target-aware baseline.
- SurfSpec refinement: SurfSpec removes atoms within 2 Å of pocket atoms, performs low-noise recovery at τ = 0.15 for 100 steps, and then applies 50 reverse-SDE steps.The refined ligand is locally optimized with AutoDock Vina before the next outer iteration.
- Refinement comparison: Refinement comparisons keep pseudo-labels, priors, linker size, patch selection, and stopping criteria fixed to isolate the recovery procedure.This protocol compares refinement methods under a shared one-step pseudo-label recovery setting.
- Surface patch selection: At each iteration, the method selects the closest feasible visible surface patch within 4–10 Å and grows a linker of size 10 toward it.Candidate patches are surface-filtered, visibility-filtered, and selected by minimum ligand–fragment distance.
C.3 Evaluation Metrics
The evaluation measures off-target agnosticism, docking-based specificity, geometric complementarity, pocket occupancy, chemical validity, and pseudo-label preservation.
- Method classification: A method is off-target agnostic when it does not use off-target structures, scores, or activity labels during ligand growth.Off-target pockets are used only for evaluation unless explicitly stated otherwise.
- Specificity: Empirical specificity compares target-pocket Vina score with the strongest predicted binder among evaluated off-target pockets.The evaluation also reports thresholded success rates for Spec > 0.2, Spec > 0.4, and Specavg > 0.6.
- Geometric fit: Geometric mismatch measures ligand–pocket surface complementarity with Jensen–Shannon distance, where lower values indicate better geometric fit.The target-pocket surface is restricted to points within 8 Å of ligand heavy atoms.
- Pocket coverage: Pocket occupancy is the fraction of target-pocket grid points covered by the generated ligand, with higher values indicating broader coverage.Average, first-quartile, and third-quartile occupancy are reported.
- Refinement quality: Refinement quality is assessed through clash rate, valence validity, bond-length MMD, RMSD, and topological similarity.Lower RMSD and bond-length MMD indicate closer geometric preservation, while higher valence validity and topological similarity are favorable.
- Geometric certificate: The geometric certificate is activated when M_L(O) > 0, yielding a nonzero triangle-inequality lower bound on ligand–off-target mismatch.The certificate has positive margin for 65.96% of filtered usable sampled pairs.
D.2 Additional Validation with Resampled Off-Targets
Additional evaluations resample off-target pockets, relax docking timeouts, vary the random seed, and inspect per-sample specificity distributions to test robustness.
- Resampled evaluation: The resampled evaluation uses a different set of 10 off-target pockets and increases the Vina timeout from 300 to 1800 seconds.It also reports docking success as the fraction of attempted dockings returning finite scores.
- Specificity results: SurfSpec achieves the largest empirical specificity improvement, best average score, and best threshold success rates under the relaxed resampled evaluation.It also has best average pocket occupancy, zero clash rate, full docking success, and competitive target-pocket docking performance.
- Baseline comparison: DiffSBDD-SizeExt attains the lowest geometric mismatch but has higher clash rates, lower docking success, and weaker empirical specificity.The comparison distinguishes geometric mismatch reduction from reliable specificity and docking behavior.
- Seed robustness: With a new random seed for generation and off-target sampling, SurfSpec remains best across average empirical specificity and all specificity success rates.This supports robustness to sampled ligands and evaluated off-target sets.
- Distributional analysis: The per-sample specificity distribution for SurfSpec shifts upward relative to the initial lead and target-only baselines.The reported shift indicates that the improvement is not driven only by a small number of outlier samples.
D.4 Qualitative Results
SurfSpec grows ligands toward under-filled pocket regions while maintaining plausible conformations and low target-pocket geometric mismatch. Its empirical specificity distribution is shifted upward, and a second-seed comparison reports the best average and thresholded specificity rates.
- Qualitative comparisons: SurfSpec expands ligands toward under-filled pocket regions while maintaining plausible pocket-bound conformations and low geometric mismatch.Lower dgm indicates better surface complementarity.
- Qualitative comparisons: Under a second random seed, SurfSpec achieves the best empirical specificity across the average score and all thresholded success rates.The comparison uses relaxed 1800 s docking timeouts and evaluates ligand generation and off-target sampling.
- Qualitative comparisons: SurfSpec shifts the empirical specificity distribution upward relative to target-only baselines.The distribution measures target-pocket preference over the strongest evaluated off-target pocket.
- Ligand growth pipeline: Each growth iteration selects the closest feasible unoccupied surface patch within 4–10 Å of the current ligand.The pipeline stops when no feasible patch exists.
- Ligand growth pipeline: SurfSpec samples and attaches a linker toward the selected patch, removes pocket-clashing linker atoms, and obtains a geometric pseudo-label.The stated clash threshold is 2 Å.