PaperlayerView exact revision on arXiv

Source-linked AI summary

PocketVE: Stable and Property-Guided Structure-Based Drug Design with Variance-Exploding Diffusion

Peining Zhang, Jinbo Bi

arXiv:2609.08101v3q-bio.BMcs.LG

TL;DR

Protein-conditioned 3D molecule generation must balance pocket compatibility, molecular properties, and physical geometry, while strong guidance can destabilize coordinates. PocketVE combines VE-style denoising, classifier-free property guidance, and pocket perturbation; on CrossDocked2020 it improves Valid3D and strain energy relative to guided TAGMol.

  • Problem

    Strong conditioning in protein-conditioned 3D generation can drive coordinates toward non-physical conformations, while separate property classifiers make multi-objective steering inflexible and poorly scalable.

    Source: §1 · pp. 1–2

  • Method

    PocketVE couples VE/EDM-style coordinate parameterization with classifier-free multi-property guidance and training-time protein-pocket perturbation.

    Source: §1 · p. 2

  • Results

    Valid3D rises from 58.6 to 80.6 and strain energy falls from 457.4 to 127.9 versus guided TAGMol, with competitive docking and molecular-property scores under moderate guidance.

    Source: §1 · p. 2

  • Takeaways & Limitations

    The results suggest that geometric stability and inference-time conditional steering should be considered jointly in target-aware 3D molecular generation.

    Source: §1 · p. 2, §6 · p. 10

    The preferred guidance scale may be target-dependent, and stronger guidance can trade geometric and distributional fidelity for target-related objectives.

    Source: §5 · p. 10

Abstract

from arXiv · show

Protein-conditioned 3D molecule generation is a central challenge in structure-based drug design, requiring a balance between pocket compatibility, molecular properties, and physical geometry. We propose \textbf{PocketVE}, a protein-pocket-conditioned variance-exploding (VE) diffusion framework that couples stable coordinate denoising with inference-time property guidance. Specifically, PocketVE combines an EDM-style training and sampling setup for 3D denoising, classifier-free guidance for multi-property steering, and adaptive protein perturbation as a training-time pocket regularizer.On the CrossDocked2020 benchmark under GenBench3D, PocketVE improves Valid$_{3\text{D}}$ from 58.6 to 80.6 and reduces strain energy from 457.4 to 127.9 relative to its guided TAGMol architectural parent; relative to TargetDiff, it attains comparable Valid$_{3\text{D}}$ with lower strain energy (127.9 vs.\ 306.0), while retaining competitive docking and molecular-property scores under moderate guidance. A guidance-scale study shows that moderate guidance gives a favorable balance between target-related objectives and geometric quality, whereas stronger guidance can degrade geometry and distributional fidelity. Pocket-permutation and PoseCheck diagnostics further support pocket-specific spatial compatibility with reduced steric conflicts. Overall, the results suggest that geometric stability and inference-time property guidance should be considered as coupled design objectives.

1 Introduction

Protein-conditioned molecule generation must balance pocket compatibility and molecular properties with physical geometry, while strong guidance can destabilize coordinates. PocketVE couples a VE-based backbone with property guidance and pocket perturbation, improving geometric results on CrossDocked2020 under GenBench3D.

1 Introduction

  • Strong conditioning can push unconstrained ligand coordinates toward non-physical conformations, while separate property classifiers make multi-objective steering inflexible and poorly scalable.Atom clashes and high strain energy reflect the geometry risk; adjusting target or drug-like objectives generally requires a separate time-dependent classifier for each property.Source: §1 · pp. 1–2
  • PocketVE jointly combines VE/EDM-style coordinate parameterization, pocket-aware scale preservation, multi-property classifier-free guidance, and training-time pocket perturbation.The paper evaluates the framework as an integrated system and does not attribute its overall improvement to VE alone.Source: §1 · pp. 2–3
  • Valid3D rises from 58.6 to 80.6 and strain energy falls from 457.4 to 127.9 versus guided TAGMol, with competitive docking and molecular-property scores under moderate guidance.PocketVE has comparable Valid3D to TargetDiff with lower strain energy; guidance experiments show a trade-off between target-related properties and geometric fidelity.Source: §1 · p. 2
  • Vina Score, QED, and SA respond differently from Valid3D and strain energy as guidance scale changes, motivating a balance between target objectives and geometric quality.The paper characterizes this guidance-scale trade-off on CrossDocked2020 under GenBench3D.Source: §1 · pp. 2–3

2 Related Work

Prior work advances 3D generation, pocket-conditioned diffusion, and property guidance, but also motivates treating geometry and target objectives together. PocketVE brings VE-style geometric-stability design into the protein-conditioned setting.

Target-Agnostic 3D Molecular Generation

  • VE-style parameterizations have improved geometric stability in target-agnostic 3D generation, and PocketVE applies this design principle where ligand coordinates must remain consistent with explicit pocket geometry.Earlier approaches progressed from equivariant coordinate denoising to latent or joint structure–geometry modeling and unified atom-type, coordinate, and bond generation.Source: §2 · p. 3

Structure-Based Drug Design

  • Pocket-conditioned generation evolved from spatial and autoregressive models to joint non-autoregressive diffusion, interaction-prior methods, adaptive subcomplexes, and flow-based approaches.GenBench3D evaluates conformation quality, while PoseCheck examines pose violations and missing interactions that redocking may obscure.Source: §2 · p. 3

Conditional Diffusion

  • Property-control methods include classifier and classifier-free guidance, alignment, and reinforcement-learning approaches, while stronger conditional pressure can degrade 3D geometry.TAGMol uses external property predictors and scale calibration; BADGER studies affinity control and joint affinity, QED, and SA optimization.Source: §2 · p. 3

3 Method

PocketVE models pocket-conditioned ligands in a hybrid coordinate and atom-type state space. Its coupled denoising, guidance, and pocket-perturbation components support either unguided generation or inference-time property steering.

3.1 Overview

  • The model learns a conditional ligand distribution given a protein pocket, representing ligand coordinates and atom types alongside pocket coordinates and features.The framework couples EDM-style coordinate denoising, inference-time classifier-free guidance, and training-time protein perturbation.Source: §3.1 · p. 4

3.2 EDM-style generation backbone

  • PocketVE uses VE coordinate diffusion with EDM preconditioning and scale-preserving ligand inputs so ligand–pocket distances remain coherent in the shared coordinate frame.The continuous coordinate process samples a noise level, and the coordinate prediction includes a skip term anchored to the noisy input.Source: §3.2 · p. 4

3.3 Sampling procedure

  • At inference, ligand coordinates start around the pocket center and follow a decreasing VE noise schedule with Euler updates, while atom types use a discrete flow-matching sampler.The number of ligand atoms is sampled from a pocket-size-conditioned prior.Source: §3.3 · p. 5

3.4 Classifier-free guidance

  • Classifier-free guidance steers Vina, QED, and normalized SA targets encoded as training-set percentile bins, avoiding a separate property predictor.The method applies guidance to coordinate predictions and atom-type logits; stronger guidance follows the auxiliary condition more strongly.Source: §3.4 · p. 5

3.5 Adaptive protein perturbation

  • Gaussian pocket perturbation regularizes training against over-reliance on one crystallographic structure and encourages a neighborhood-aware conditional distribution.The main-results schedule is σp = min(0.1σ, 0.5), which discourages sharp prediction changes under small pocket-coordinate variations.Source: §3.5 · p. 6

3.6 Full objective

  • The final objective uses coordinate and atom-type predictions from perturbed pockets, and the trained model can run unguided or with classifier-free guidance at test time.Guidance provides a steerable trade-off between fidelity and target-specific preference.Source: §3.6 · p. 6

4 Experiments

Experiments evaluate PocketVE on filtered CrossDocked2020 under GenBench3D, comparing target-aware generation methods across binding, molecular-property, and geometry metrics. Main results and ablations examine the property–geometry trade-off and the effects of guidance and pocket perturbation.

4.1 Experimental setup

  • The evaluation uses filtered CrossDocked2020 poses, 100 test proteins with 100 samples each, and GenBench3D geometry metrics alongside Vina, QED, normalized SA, and diversity.Baselines include guided TAGMol as the architectural parent, TargetDiff as the primary geometry comparison, and PAFlow as the strong-affinity comparison.Source: §4.1 · pp. 6–7

4.2 Main Results

  • PocketVE improves geometric quality while retaining competitive docking and molecular-property scores, rather than uniformly dominating every metric.Table 1 summarizes binding affinity, molecular properties, geometry, and diversity for PocketVE at s = 5 and other baselines.Source: Table 1 · p. 7, §4.2 · p. 8
  • Moderate PocketVE CFG improves predicted binding while staying near the high-validity, low-strain regime; PAFlow ranks first on Vina-based metrics but pays a large geometric cost.Figure 2 plots −Vina Scoremed against Valid3D and log-scale strain energy; AR and Pocket2Mol are off-scale because of weaker median Vina scores.Source: Figure 2 · p. 8, §4.2 · p. 8
  • Unguided PocketVE most closely matches reference distance and atom-type distributions, while PocketVE has lower rigid-fragment relaxation RMSD than several baselines across most fragment sizes.In the representative 5D7N example, PocketVE places the ligand closer to the reference pose while maintaining a compact structure inside the pocket.Source: §4.2 · p. 8, Figure 3 · p. 8, Figure 4 · p. 9

4.3 Ablations and Analysis

  • The cumulative ablation attributes a large geometric improvement over TAGMol to the integrated PocketVE backbone and sampler, with moderate CFG shifting affinity-related objectives and protein perturbation improving the final geometry-oriented operating point.Removing the σdata factor changes mean Vina from −7.589 to +92.237 and clash-free rate from 94.23% to 3.39%, while Valid3D is 78.72% versus 82.39%.Source: §4.3 · p. 9, Table 2 · p. 9
  • With extra CFG passes included, PocketVE remains substantially faster than most diffusion and autoregressive baselines and slightly faster than PAFlow in the same runtime regime.The 100-step sampler is a deliberate operating point supported by the sampling-step ablation, not a naive truncation of a longer trajectory.Source: §4.3 · p. 9
  • Increasing CFG first improves target-related objectives while largely preserving geometry, but overly large scales reduce Valid3D and increase distribution drift; s = 5 is representative, while s = 10 favors affinity at a geometric cost.Property distributions and conflicting-condition analyses also show trade-offs in geometry, distributional fidelity, and diversity at stronger guidance.Source: §4.3 · p. 10
  • Among evaluated protein perturbation settings, σp = min(0.1σ, 0.5) gives the most favorable trade-off across local JSD, atom-type JSD, Vina metrics, and centroid distance.Lighter perturbation further reduces strain, whereas stronger schedules improve broader distributional statistics and SA at some cost in target-specific precision.Source: §4.3 · p. 10

5 Limitations

PocketVE’s practical deployment is constrained by inference cost, target-dependent guidance, and the scope of its evaluation and modeling assumptions.

5 Limitations

  • Classifier-free guidance doubles function evaluations per sampling step because it requires conditional and unconditional forward passes.The preferred guidance scale may vary by target, and stronger guidance can trade geometric and distributional fidelity for target-related objectives.Source: §5 · p. 10
  • Evaluation uses a filtered 100-pocket CrossDocked2020 split, docking-based affinity proxies, and benchmark-specific protocols; broader generalization and experimental validation remain future work.Vina proxies may reward over-optimization of binding geometry rather than true binding affinity.Source: §5 · p. 10
  • Gaussian pocket perturbation is a simple regularizer, not a realistic model of protein conformational flexibility.The TAGMol-to-PocketVE transition bundles backbone and sampler choices, so the experiments do not establish VE-only causal attribution; cross-method comparisons are also confounded by generated atom count.Source: §5 · p. 10
  • All results use first-order Euler sampling, which is not claimed to be optimal, and the guidance scale s = 5 was selected descriptively on the test-benchmark sweep.The fixed σdata = 10 choice was robust across the evaluated sensitivity range, but s = 5 was not selected using an independent validation set.Source: §5 · p. 10

6 Conclusion

PocketVE combines target-aware coordinate denoising and inference-time property control to improve geometric stability while supporting conditional steering. Its accompanying appendices describe the guidance rules, coordinate scaling, and sampling setup used to implement that framework.

6 Conclusion

  • PocketVE improves geometric stability in unguided generation and offers a more favorable geometry–target-objective trade-off under guidance on CrossDocked2020.The conclusion identifies broader target families, matched component controls, and stronger links between computational diagnostics and experimental validation as future work.Source: §6 · p. 10

A.1 Noise-Level Dependence of Explicit Classifier Guidance

  • Under the VE parameterization, explicit classifier guidance enters the denoiser through a σ^2-scaled property gradient and the sampler update through a σ-scaled term.The passage relates the external classifier gradient to the VE Tweedie identity and the coordinate denoiser prediction.Source: §A.1 · p. 15

A Classifier-Free Guidance Details

  • Classifier-free guidance extrapolates between conditional and dropped-condition predictions to steer ligand coordinates and atom-type logits toward requested properties.The coordinate rule uses the conditional prediction plus a scaled difference from the dropped-condition prediction; the same extrapolation is applied directly to categorical logits.Source: §A.2 · p. 16

A.2 Recovering the Classifier-Free Guidance Rule

  • The classifier-free coordinate rule is the dropped-condition prediction plus s times the conditional-minus-dropped-condition prediction.The categorical branch uses the analogous extrapolation on logits.Source: §A.2 · p. 16

B EDM Preconditioning and Target-Aware Coordinate Scaling

  • PocketVE adapts EDM input scaling so ligand and unscaled pocket coordinates remain aligned in their shared geometric graph.Matching output recentering avoids distorting protein–ligand distances and keeps the ligand in the pocket’s physical frame near the clean limit.Source: B EDM Preconditioning and Target-Aware… · pp. 16–17

C Sampling Details

  • The coordinate branch uses σdata = 10.0 to cover both ligand extent and pocket-relative displacement, while atom types use a separate discrete flow-matching sampler.The reported VE range is σmax = 800 and σmin = 10^-3; the noise schedule is generalized arcsin, and atom types are updated with a DFM transition-rate sampler.Source: C Sampling Details · pp. 17–18

D Sampling-Step Ablation

The ablations examine sampling-step count, pocket perturbation, guidance strength, and a RampUp guidance schedule. Across these tests, operating choices balance target-related objectives against geometric quality and sampling cost.

D Sampling-Step Ablation

  • Moving from 25 to 100 sampling steps improves predicted binding strength and 3D validity across representative CFG scales, while 100 to 200 steps yields smaller gains relative to added cost.N = 100 is the default; Figure 5 plots −Vina Scoremed and Valid3D, with the dashed line marking that setting.Source: D Sampling-Step Ablation · p. 18, Figure 5 · p. 18

E Pocket Perturbation Details

  • A bounded, noise-dependent pocket perturbation schedule balances local smoothing against preservation of target-specific conditioning information.The main schedule is σp = min(0.1σ, 0.5); its ablation improves local distance fidelity, atom-type JS, Vina metrics, and centroid distance, while stronger perturbation favors broader distributional and SA-oriented statistics at some cost to target-specific precision.Source: E Pocket Perturbation Details · p. 19

F Guidance Scale Sensitivity Details

G RampUp Guidance Schedule

  • At similar median docking strength, RampUp gives a comparable moderate-guidance trade-off, but its advantage is not consistent at stronger docking targets.For example, constant CFG s = 2 gives median Vina Score −7.33 and median Vina Min −7.78, versus −7.49 and −7.81 for RampUp with final scale s = 5; the matched RampUp setting has JSD-All 0.0488 versus 0.0490, completion 0.960 versus 0.954, and molecular stability 34.7% versus 34.4%.Source: G RampUp Guidance Schedule · p. 21

H Component-Wise Ablation Summary

The component-wise comparisons attribute geometric gains to a bundled backbone-and-sampler transition, while separate tables summarize run variability and reference-normalized multi-objective hits.

H Component-Wise Ablation Summary

  • Relative to guided TAGMol, unguided PocketVE gains about 20 Valid3D points and reduces strain energy by about 64%, reflecting a bundled backbone-and-sampler change rather than an isolated VE effect.Moderate CFG from s = 0 to s = 5 improves Vina Score, QED, and SA while preserving high 3D validity; the main Muon–AdamW setting gives the best overall geometry and strain, and replacing the generalized arcsin schedule with log-uniform sampling weakens several metrics.Source: H Component-Wise Ablation Summary · pp. 21–22, Table 6 · p. 21

I Repeated-Run Variability

  • Table 7 reports final PocketVE metrics as medians across three independent runs with standard deviations.The summary includes representative property, affinity, geometry, and distribution metrics.Source: Table 7 · p. 22, I Repeated-Run Variability · p. 22

J Reference-Normalized Multi-Objective Hits

  • Among the filtered methods, PocketVE has the highest Joint-Ref average and Joint Success rate for samples jointly no worse than the reference ligand in docking affinity, QED, and SA.Joint Success is the percentage of pockets with at least one such molecule.Source: J Reference-Normalized Multi-Objective… · p. 22, Table 8 · p. 22

K Additional Property–Validity Trade-Offs

The appendix extends the guidance-scale analysis to drug-likeness, synthesizability, and intraligand distance distributions. Moderate guidance retains high 3D validity, while stronger guidance and distance-profile diagnostics reveal limits to property steering and distributional fidelity.

K Additional Property–Validity Trade-Offs

  • Weak-to-moderate guidance improves QED and SA while retaining high Valid3D, whereas stronger guidance eventually lowers Valid3D.Figure 8 shows the QED–Valid3D and SA–Valid3D trade-offs across baselines and PocketVE guidance scales.Source: Figure 8 · p. 23, K Additional Property–Validity… · p. 23

L Additional Distance-Distribution Diagnostics

  • PocketVE’s guided-sampling distance distributions remain comparatively close to reference, while larger JSDs coincide with broader short- and mid-range distance distortions.The distributions are a fidelity diagnostic, not an explicit measure of physical validity.Source: L Additional Distance-Distribution… · p. 23, Figure 9 · p. 23

M Test-Time Protein Perturbation

A controlled stress test adds Gaussian noise to pocket coordinates at inference while holding ligand initialization, sampling schedule, and guidance scale fixed. Pocket perturbation training generally preserves distance-distribution fidelity, but severe test-time corruption still weakens scores and clash-free rates.

M Test-Time Protein Perturbation

  • The stress test adds Gaussian noise only to pocket coordinates, keeping ligand initialization, sampling schedule, and CFG scale fixed at s = 5.It probes tolerance to small conditioning-geometry errors rather than serving as a standard evaluation setting.Source: M Test-Time Protein Perturbation · p. 24
  • Perturbation-trained variants generally retain lower distance-distribution JSDs, but severe test-time pocket corruption weakens Vina-based scores and clash-free rates for all models.The stronger 0.3σ training perturbation improves SA and strain at several noise levels; the milder min(0.1σ, 0.5) perturbation retains the strongest Vina-based scores at low test-time noise.Source: M Test-Time Protein Perturbation · p. 24, Table 9 · p. 24

N Pocket-Diagnostic Details

These diagnostics examine pocket specificity, property steering, and global diversity under guidance. They show cross-property competition and indicate that pooled diversity measures may miss pocket-level behavior.

N Pocket-Diagnostic Details

  • Pocket shuffling is a post-hoc full-pipeline specificity control, not ligand regeneration under a mismatched pocket or an isolated causal test.The fixed cyclic permutation has no self-matches across 100 pockets; the same 9,471 ligands are compared after translation to the target pocket center.Source: N Pocket-Diagnostic Details · p. 24
  • PoseCheck changes are mixed, and total interaction counts should not be read as interaction quality: reference ligands average 11.79 interactions, close to PocketVE’s 11.99.The analysis uses original poses without redocking and averages pockets equally; clash-free rates and contact categories are reported separately.Source: N Pocket-Diagnostic Details · p. 24, Table 10 · p. 25, Table 11 · p. 25

O Property-Steering Diagnostics

  • Vina score shifts are reported directionally, not as strict Vina-bin attainment, because evaluation uses vina_score while training boundaries use vina_dock.Table 14 also reports favorable-bin attainment for QED and SA.Source: O Property-Steering Diagnostics · p. 25, Table 14 · p. 26

P CFG Diversity and Reference Coverage

  • Higher CFG mildly reduces molecule-level uniqueness and increases fingerprint similarity, while reference-scaffold coverage decreases; scaffold entropy and top-scaffold concentration do not decline monotonically.These pooled summaries can mask within-pocket convergence or pocket-specific reference recovery.Source: P CFG Diversity and Reference Coverage · p. 25

O Property-Steering Diagnostics

  • Ordered condition requests move all three aggregate medians in the requested directions, while conflicting requests expose competition among Vina, QED, and SA.For example, (4, 0, 2) improves Vina but lowers QED relative to (2, 4, 2); outcomes do not establish calibrated independent control.Source: O Property-Steering Diagnostics · p. 25, Table 13 · p. 26

Q Coordinate-Scale Robustness and Diagnostics

The appendix tests coordinate-scale sensitivity and provides additional qualitative cases. No evaluated σdata setting dominates all metrics, and the selected value is presented as a fixed operating choice within a robust range.

Q Coordinate-Scale Robustness and Diagnostics

  • No σdata setting dominates all metrics: 10 has the highest evaluation success, 5 the most favorable mean Vina, 7.5 the highest SA and lowest median strain, and 15 the lowest evaluation success.Across the sweep, JSD-All-12Å ranges from 0.0592 to 0.0665, JSD-CC-2Å from 0.2533 to 0.2819, and atom-type JS from 0.1215 to 0.1458; σdata = 10 is a fixed choice, not a sharply optimized constant.Source: Q Coordinate-Scale Robustness and… · p. 26, Table 16 · p. 27
  • Removing the σdata factor in Eq. (32) leaves reconstruction and ligand-only geometry similar but collapses pocket-aware spatial compatibility, while removing sampling-time noise sharply lowers evaluation success and geometric quality.The VEDA-style comparison bundles the TAGMol architecture with σdata = 1 and s = 0, so it does not isolate σdata’s effect.Source: Q Coordinate-Scale Robustness and… · p. 26, Table 17 · p. 27

R Additional Case Studies

Loading 2609.08101v3…