acceptodds
Under review as a conference paper at ICLR 2027

SurfSpec: Enhancing Off-Target-Agnostic Specificity by Bounding Pocket-Ligand Geometric Mismatch

Abstract

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.

Primary area:
generative models
Keywords:
Structure-based Drug Design

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_6GppOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3(+1)
Soundness
2
Read review
Paper summary

The paper presents SurfSpec that performs off-target lead optimization without the need for true negative protein-ligand pairs. The model uses shape matching as a surrogate for specific binding and designed a partial diffusion framework that enforces the generation to take up most spaces in the pocket. Among naive diffusion refinement, the model performs better and saves diffSBDD for off-target binding.

Critical strengths
  • The framework outperforms baseline methods on target specificity without degrading binding scores.
  • There is no additional training needed for this method for using additional negative data
Critical weaknesses
  • The model is only able to optimize based on shape mismatch without considering true physical interactions like charges, h-bonds
  • This model will always generate bigger ligands that sacrifice drug-like properties. Without deleting any fragments, the model is just getting better scores by creating more shape interactions
  • The paper is a little theory-heavy which makes it hard to follow. After carefully reading, I could understand it. It is possible to simplify the method section.
  • No ablation study is performed on hyper-parameters like iterateration nums.
Questions for authors
  • How would the model perform if we normalize scores with the size of the ligand?
  • How should we incorporate more chemical/physical interactions for specificity instead of completely depended on surface shapes.
16 comments

Response to Reviewer 6Gpp

We thank the reviewer for recognizing the value of specificity-aware lead optimization without negative-target supervision. We address the concerns regarding interaction fidelity, molecular-size growth, drug-like properties, theoretical presentation, and design ablations below.

1. Incorporating chemical and physical interactions beyond surface shape

Response to W1 and Q2

We agree that geometric mismatch captures only the surface-complementarity component of molecular recognition. To incorporate receptor–ligand interactions explicitly, we introduce SurfSpec(ΔJ\Delta J), which augments the geometric objective with a ligand-specific interaction term. SurfSpec(ΔJ\Delta J) first generates surface-directed candidates and then selects the candidate with the strongest target-versus-competitor interaction gain.

Let TT denote the target pocket, LL a ligand, and u(R,L)=−SVina(R,L)u(R,L)=-S_{\mathrm{Vina}}(R,L) a higher-is-better receptor–ligand interaction score. We define

J(T,L)=u(T,L)−EP∼p(P∣L)[u(P,L)],J(T,L)=u(T,L)-\mathbb E_{P\sim p(P\mid L)}[u(P,L)],

where p(P∣L)p(P\mid L) is the conditional distribution of plausible pockets that can interact with LL. We estimate the expectation by Monte Carlo sampling from an RFdiffusion-based ligand-conditioned pocket generator:

J^(T,L)=u(T,L)−1Npanel∑k=1Npanelu(Pk,L),Pk∼p(P∣L),\widehat J(T,L)=u(T,L)-\frac{1}{N_{\mathrm{panel}}} \sum_{k=1}^{N_{\mathrm{panel}}}u(P_k,L), \qquad P_k\sim p(P\mid L),

with Npanel=8N_{\mathrm{panel}}=8. The interaction-aware geometric objective is

U(T,L)=dgm(L,T)−J^(T,L).\boxed{ U(T,L)=d_{\mathrm{gm}}(L,T)-\widehat J(T,L). }

The two terms are optimized sequentially rather than added on an uncalibrated common numerical scale. For a candidate LaL_a grown from the current step input L0L_0, we compute

ΔJ(La;L0)=J^(T,La)−J^(T,L0),\Delta J(L_a;L_0)=\widehat J(T,L_a)-\widehat J(T,L_0),

or equivalently,

ΔJ(La;L0)=[u(T,La)−u(T,L0)]−1Npanel∑k=1Npanel[u(Pk,La)−u(Pk,L0)].\Delta J(L_a;L_0)=\left[u(T,L_a)-u(T,L_0)\right]-\frac{1}{N_{\mathrm{panel}}} \sum_{k=1}^{N_{\mathrm{panel}}} \left[u(P_k,L_a)-u(P_k,L_0)\right].

Thus, a candidate is rewarded when its target-affinity gain exceeds its average affinity gain across the synthetic competitor panel. Since J^(T,L0)\widehat J(T,L_0) is shared by all candidates within the same step, ranking by ΔJ\Delta J is equivalent to ranking by J^(T,La)\widehat J(T,L_a).

1.1 Compared configurations

SurfSpec(ΔJ\Delta J). This variant uses linker size 4, the configured 0–4 Å patch band, at most three growth steps, two generation seeds per step, and the crystal reference ligand as the initial parent pose. At each step, it ranks accessible patches using the burial policy, generates and refines candidates at the top three patches, and retains the completed candidate with the largest ΔJ\Delta J. The same RFdiffusion-generated panel of Npanel=8N_{\mathrm{panel}}=8 ligand-conditioned competitor pockets is shared across the three candidates in that step. The panel term uses mean pooling, all target and panel interactions are evaluated with Vina --local_only, and the optional no-edit gate is disabled; ΔJ\Delta J is used only for candidate reranking. The molecular generator, guidance schedule, and low-noise refinement procedure are otherwise unchanged from SurfSpec(Short Arm).

SurfSpec(Short Arm). This baseline uses linker size 4, the configured 0–4 Å patch band, the closest feasible target-surface patch, at most three growth steps, two generation seeds per step, the crystal reference ligand as the initial parent pose, and one generated candidate per step.

SurfSpec(JS-Guided). This variant uses the same Short Arm configuration but evaluates three completed candidates per step. It first orders feasible patches by the closest-patch rule, takes the top three patches, and independently runs the full surface-targeted pseudo-label generation and refinement procedure for each patch. It then computes dgm(La,T)d_{\mathrm{gm}}(L_a,T) for each completed ligand candidate LaL_a and retains the candidate with the smallest value. Thus, JS-Guided is a post-generation reranking rule over three fully generated ligands, not a JS score assigned to patches before generation.

SurfSpec(Long Arm). This is the submitted configuration: linker size 10, the configured 4–10 Å patch band, the closest feasible target-surface patch, and at most two growth steps.

Submission baselines. Delete, PMDM, DecompOpt, DiffShape, ActivityDiff, the multi-step PMDM/DecompOpt variants, and the DiffSBDD size-extension controls use the configurations reported in Table 1 of the submission; no new hyperparameters are introduced for these rows.

The submission baseline rows below reproduce Table 1 for context. The three new Short Arm rows are compared under the same matched rebuttal protocol.

MethodOff-target agnosticSpec ↑ (mean ± SD)Geom. mismatch ↓Occ. ↑Target Vina ↓Clash rate ↓# AtomsDock success ↑
Initial LeadO−0.81±1.43-0.81\pm1.430.5400.26-7.170.0022.751.00
DeleteO−0.81±1.54-0.81\pm1.540.5390.26-7.230.0022.851.00
PMDMO−1.10±2.03-1.10\pm2.030.5010.33-6.470.1631.000.92
DecompOptO−0.93±1.43-0.93\pm1.430.5420.25-6.080.0222.750.91
DiffShapeO−0.88±1.58-0.88\pm1.580.5850.14-6.900.1121.910.99
ActivityDiffX−0.96±2.15-0.96\pm2.150.5380.27-7.070.1122.891.00
PMDM–MS2O−1.38±2.19-1.38\pm2.190.4990.31-8.950.2533.420.91
PMDM–MS3O−2.05±2.93-2.05\pm2.930.4960.31-8.380.3135.830.82
PMDM–MS+ESO−1.02±2.08-1.02\pm2.080.5000.33-8.610.1631.370.87
DecompOpt–MS2O−1.03±1.33-1.03\pm1.330.5420.25-7.080.0322.750.96
DecompOpt–MS3O−0.99±1.39-0.99\pm1.390.5410.25-7.000.0422.760.97
DecompOpt–MS+ESO−0.83±1.59-0.83\pm1.590.5410.25-7.400.0322.750.95
DiffSBDD–SizeExt(+20)O−1.51±2.24-1.51\pm2.240.4800.28-7.570.4535.880.93
DiffSBDD–SizeExt(+30)O−2.20±3.76-2.20\pm3.760.4680.32-7.140.6241.580.81
SurfSpec(Long Arm)O−0.37±2.54-0.37\pm2.540.4800.36-8.630.0034.991.00
SurfSpec(Short Arm)O−0.34±2.27-0.34\pm2.270.4830.32-8.910.0531.70.98
SurfSpec(JS-Guided)O−0.14±2.42-0.14\pm2.420.4700.33-9.070.0432.50.98
SurfSpec(ΔJ\Delta J)O−0.17±2.61-0.17\pm2.610.4860.32-9.200.0231.90.97

The molecular-property comparison among the matched three-step variants is:

MethodQED ↑SA ↓LE ↑
SurfSpec(Short Arm)0.4054.330.286
SurfSpec(JS-Guided)0.3754.230.289
SurfSpec(ΔJ\Delta J)0.4424.260.310

SurfSpec(ΔJ\Delta J) improves mean specificity from −0.34-0.34 to −0.17-0.17 relative to SurfSpec(Short Arm) while retaining nearly the same molecular size. It also provides the strongest target Vina score, QED, and ligand efficiency among the matched three-step SurfSpec variants.

The result shows that SurfSpec is not restricted to shape-only optimization: geometric guidance generates target-specific candidates, while a receptor–ligand interaction objective determines which candidate is retained.

We do not claim that the current Vina implementation is a complete physical free-energy model. A higher-fidelity estimator incorporating electrostatics, solvation, or receptor flexibility can replace uu without changing the proposal-and-selection structure.

Revised-draft location. Appendix C.1, SurfSpec Variants and Rebuttal Configurations, defines SurfSpec(ΔJ\Delta J), SurfSpec(JS-Guided), and the matched baselines. Appendix E.3, Synthetic-Competitor Candidate Selection, gives the interaction objective, implementation, and baseline-inclusive results.

Reviewer_AEnpOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
2
Soundness
3(+1)
Read review
Paper summary

This paper proposes an off-target-agnostic specificity-aware lead optimization framework SufSpec that encourages the designed ligand to match the given target pocket, without requiring off-target pocket structures. SurSpec defines pocket-ligand mismatch as a core metric and provides some mathematical support. It iteratively grows ligands using DiffLinker and refines them with DiffSBDD. The experimental results show that SurSpec can improve specificity, which is the core goal in this task.

Critical strengths
  1. This work focuses on a meaningful and practical problem in molecular optimization.

  2. The motivation is interesting and reasonable. The geometric mismatch provides a useful theoretical perspective on specificity-aware ligand optimization.

  3. The proposed framework effectively combines surface-guided ligand growth with diffusion-based refinement, offering a clear and intuitive design.

  4. The experimental results are encouraging, showing improved specificity and competitive binding affinity.

Critical weaknesses
  1. Although SurfSpec improves empirical specificity, table 1 also suggests that the strongest off-target docking score becomes more favorable (approximately -7.98 to -9.00), while the target docking score improves from -7.17 to -8.63. This indicates that the specificity gain mainly results from a larger improvement in target affinity instead of reduced off-target binding. It's not directly consistent with the motivation of avoiding undesired off-target interactions.

  2. Although the paper provides several baseline comparisons, the contributions of the proposed modules to the final specificity improvement remain insufficiently isolated.

  3. It shows that SurfSpec substantially increases the average ligand size. Although the improvements in specificity and binding affinity are encouraging, this leads to the concern about the impact on drug-likeness and synthetic accessibility.

  4. Figure 2 is small and lacks explanation.

Questions for authors
  1. The paper would benefit from more ablation study of its key components. The proposed surface patch selection does not directly optimize the geometric mismatch in Eq 4, so comparisons with alternative selection strategies would help justify this design. Also, while table 2 evaluates the low-noise refinement separately, comparing it with existing methods in the full pipeline would better demonstrate its contribution to the final specificity improvement.

  2. Reporting additional molecular properties, such as QED and SA Score, would help assess the practical applicability of the generated ligands.

12 comments

Response to Reviewer AEnp

We thank the reviewer for the constructive assessment and for recognizing the practical importance of off-target-agnostic specificity-aware optimization, the geometric-mismatch perspective, and the combination of surface-guided growth with diffusion-prior refinement.

1. Specificity gain despite increased strongest off-target affinity

Response to W1

We agree that the observed specificity improvement primarily arises because the target-pocket affinity improves more than the affinity to the strongest evaluated off-target, rather than because SurfSpec systematically weakens all off-target binding.

Our intended claim is not that every off-target affinity must decrease. SurfSpec encourages the ligand to exploit interaction surfaces that are particularly compatible with the target pocket, so that the target benefits more strongly from the added ligand surface than competing pockets do. In terms of affinity gain,

ΔSpec⁡=ΔAtarget−ΔAstrongest off-target.\Delta \operatorname{Spec}=\Delta A_{\mathrm{target}}-\Delta A_{\mathrm{strongest\ off\text{-}target}}.

Specificity therefore improves whenever the target-affinity gain exceeds the strongest off-target-affinity gain; systematic off-target weakening is not required.

We agree that the original wording and Figure 1 may suggest that off-target affinity must decrease. We may revise both to emphasize relative target-over-off-target preference rather than absolute off-target weakening.

Revised-draft location. Section 3.1, Problem Setup: Specific Lead Optimization without Off-Target Access; the revised Figure 1 caption; and Appendix E.1, Specificity Improvement as a Relative Affinity Gain.

2. Ligand size and drug-like properties

Response to W3 and Q2

We agree that ligand growth can confound specificity evaluation by increasing molecular size and degrading practical molecular properties. We therefore evaluate a smaller-growth configuration and report size-normalized and drug-like properties.

2.1 Short-arm growth

SurfSpec(Long Arm). This is the submitted configuration: linker size 10, the configured 4–10 Å patch band, the closest feasible target-surface patch, and at most two growth steps.

SurfSpec(Short Arm). This configuration keeps the same surface-guided generation and low-noise refinement procedure, but uses linker size 4, the configured 0–4 Å patch band, the closest feasible patch, at most three growth steps, two generation seeds per step, and the crystal reference ligand as the parent pose. The 0–4 Å value denotes the generator configuration; it is not a strict bound on every realized nearest-parent distance.

SurfSpec(JS-Guided). This variant uses the complete SurfSpec(Short Arm) configuration. At each growth step, it first orders all feasible patches using the closest-patch policy and selects the top three. It then runs the full surface-targeted pseudo-label generation and low-noise refinement procedure independently for each of these three patches. After obtaining three completed ligands, it computes the target geometric mismatch dgm(La,T)d_{\mathrm{gm}}(L_a,T) for each candidate LaL_a and retains the candidate with the smallest value. Thus, SurfSpec(JS-Guided) performs post-generation reranking of three completed ligands; it does not assign a JS score to patches before molecular generation. This variant uses three candidate generations per step rather than one.

The following table follows the columns and definitions of the submission main table and includes the submission baselines for context.

MethodOff-target agnosticSpec ↑ (mean ± SD)Geom. mismatch ↓Occ. ↑Target Vina ↓Clash rate ↓# AtomsDock success ↑
Initial LeadO−0.81±1.43-0.81\pm1.430.5400.26-7.170.0022.751.00
DeleteO−0.81±1.54-0.81\pm1.540.5390.26-7.230.0022.851.00
PMDMO−1.10±2.03-1.10\pm2.030.5010.33-6.470.1631.000.92
DecompOptO−0.93±1.43-0.93\pm1.430.5420.25-6.080.0222.750.91
DiffShapeO−0.88±1.58-0.88\pm1.580.5850.14-6.900.1121.910.99
ActivityDiffX−0.96±2.15-0.96\pm2.150.5380.27-7.070.1122.891.00
PMDM–MS2O−1.38±2.19-1.38\pm2.190.4990.31-8.950.2533.420.91
PMDM–MS3O−2.05±2.93-2.05\pm2.930.4960.31-8.380.3135.830.82
PMDM–MS+ESO−1.02±2.08-1.02\pm2.080.5000.33-8.610.1631.370.87
DecompOpt–MS2O−1.03±1.33-1.03\pm1.330.5420.25-7.080.0322.750.96
DecompOpt–MS3O−0.99±1.39-0.99\pm1.390.5410.25-7.000.0422.760.97
DecompOpt–MS+ESO−0.83±1.59-0.83\pm1.590.5410.25-7.400.0322.750.95
DiffSBDD–SizeExt(+20)O−1.51±2.24-1.51\pm2.240.4800.28-7.570.4535.880.93
DiffSBDD–SizeExt(+30)O−2.20±3.76-2.20\pm3.760.4680.32-7.140.6241.580.81
SurfSpec(Long Arm)O−0.37±2.54-0.37\pm2.540.4800.36-8.630.0034.991.00
SurfSpec(Short Arm)O−0.34±2.27-0.34\pm2.270.4830.32-8.910.0531.70.98
SurfSpec(JS-Guided)O−0.14±2.42-0.14\pm2.420.4700.33-9.070.0432.50.98

SurfSpec(Short Arm) attains comparable mean specificity with approximately 3.3 fewer heavy atoms than SurfSpec(Long Arm). SurfSpec(JS-Guided) further improves both empirical specificity and target geometric mismatch while remaining smaller than the submitted configuration. The main result therefore does not require the larger edit used by SurfSpec(Long Arm), and stronger optimization of the proposed geometric quantity provides additional benefit at a similar molecular size.

2.2 Drug-like property evaluation

We additionally report molecular weight, QED, SA score, LogP, rotatable bonds, and ligand efficiency. To place the generated molecules in the benchmark chemical space, we compare their mean and median values with the empirical 5th–95th percentile range of the CrossDocked2020 test-split initial ligands.

PropertyCrossDocked2020 initial-ligand P5–P95SurfSpec(Short Arm), mean; medianSurfSpec(Long Arm), mean; median
# Atoms9.09.0–39.839.831.7;32.031.7;32.034.99;34.534.99;34.5
MW131.8131.8–546.6546.6 Da431.4;429.3431.4;429.3512.9;505.5512.9;505.5
QED0.1660.166–0.7820.7820.405;0.3800.405;0.3800.317;0.2690.317;0.269
SA1.831.83–5.745.744.33;4.444.33;4.444.73;4.624.73;4.62
LogP−2.73-2.73–5.525.522.08;1.872.08;1.871.60;1.651.60;1.65
Rotatable bonds00–996.6;76.6;78.2;98.2;9
Ligand efficiency0.1740.174–0.4950.4950.286;0.2930.286;0.2930.242;0.2450.242;0.245

The mean and median values of both SurfSpec configurations fall within the empirical CrossDocked2020 ranges for every reported property. SurfSpec(Short Arm) is consistently closer to the center of the initial-ligand distribution, particularly for molecular weight, QED, SA score, and ligand efficiency.

Thus, SurfSpec(Short Arm) provides a better specificity–developability trade-off: it preserves the specificity improvement while reducing molecular growth and keeping the central property values within the benchmark’s ligand distribution.

Revised-draft location. SurfSpec(Short Arm) and SurfSpec(Long Arm) are compared in Table 1 and Section 5.2. Appendix C.1, SurfSpec Variants and Rebuttal Configurations, gives the complete SurfSpec(JS-Guided) configuration. Appendix E.2, Surface-Targeted Proposal and Component Ablations, provides the baseline-inclusive comparison, and Appendix E.4 reports the complete drug-like property distributions.


Reviewer_o43HOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

SurfSpec studies lead optimization without access to off-target structures. It derives a conditional specificity lower bound from pocket–ligand geometric mismatch, then proposes iterative growth toward unoccupied target-pocket regions followed by diffusion-based refinement. The evaluation reports improved docking-based specificity on CrossDocked2020.

Critical strengths

Improving specificity without a prespecified off-target set is a worthwhile problem. Separating surface-directed proposal generation from molecular-prior refinement is a sensible way to introduce geometric control while recovering chemically valid ligands.

Critical weaknesses

The specificity bound does not explain the method’s defining growth rule. Theorem 1 favors reducing JS target mismatch, but Algorithm 1 selects the nearest feasible unoccupied patch and stops according to target-pocket Vina. The paper establishes no connection between these updates and a decrease in the mismatch appearing in the theorem. Removing the specificity analysis would leave the growth algorithm unchanged.

The policy favors accessible space without identifying whether that space contains features that distinguish the target from other pockets. The theorem assumes geometric separation throughout optimization, while the algorithm provides no connection between its chosen growth directions and that separation. The proposed mechanism therefore remains much closer to improving target-surface fit than to directing growth toward specificity.

Questions for authors
  1. What justifies treating closest-patch growth and subsequent refinement as updates that reduce the JS mismatch in Theorem 1, particularly when the ligand-dependent domain changes?
  2. What guidance does the bound provide for choosing among feasible growth directions beyond the general preference for better target-surface fit?
17 comments

Response to Reviewer o43H

We thank the reviewer for the careful questions about the connection between the geometric objective, the implemented growth policy, and specificity enhancement.

1. Is the algorithm an appropriate way to reduce JS mismatch?

Response to W1

Our algorithm is motivated by the geometric intuition that filling an under-occupied region near the target surface constructs a ligand surface more similar to the target surface, thereby reducing geometric mismatch. The closest-patch rule is not the central mechanism; it is a simple design choice for selecting an accessible target-surface region. SurfSpec realizes this intuition through a two-stage procedure. First, surface-targeted linker generation constructs a geometric pseudo-label that serves as a low-JS-mismatch reference shape. Second, low-noise refinement recovers a chemically valid ligand that remains geometrically close to this reference.

We evaluate this mechanism using the following configurations.

SurfSpec(Long Arm). This is the submitted configuration: linker size 10, configured patch range 4–10 Å, closest feasible target-surface patch, and at most 2 growth iterations.

SurfSpec(Short Arm). This smaller-growth configuration uses linker size 4, configured patch range 0–4 Å, closest feasible patch, at most 3 growth iterations, two generation seeds per step, and the crystal reference ligand as the parent pose. The 0–4 Å value is the generator setting rather than a strict bound on every realized nearest-parent distance.

SurfSpec(Random Patch). This uses the Short Arm configuration but uniformly samples one patch from the same feasible set using a fixed random seed per pocket.

SurfSpec(Farthest Patch). This uses the Short Arm configuration but selects the farthest patch within the same feasible set. “Farthest” therefore means the far end of the configured candidate range, not an arbitrarily distant pocket region.

SurfSpec(Random 2D-Graph). This is a pseudo-label ablation rather than a patch policy. A random connected four-heavy-atom arm is attached to the current molecule at the 2D-graph level without using the pocket, surface, or pose. The current molecule is pinned; torsions moving a new atom are sampled uniformly from [−180∘,180∘][-180^\circ,180^\circ]; the arm is relaxed with constrained MMFF94; and the best of six draws is retained using MMFF energy plus a close-contact penalty. The resulting pseudo-label is passed through the same SurfSpec refinement with the same two-seed retry budget.

1.1 Surface-targeted pseudo-labels provide low-mismatch reference shapes

To isolate proposal quality, we compare the first-step pseudo-labels, where every SurfSpec policy modifies the same fixed parent. We also report the submission main-table outputs as reference points for the mismatch scale. The latter are complete optimized outputs rather than first-step pseudo-labels and are therefore contextual rather than strictly stage-matched comparisons.

Method / reference shapeComparison levelAdded atomsGeom. mismatch ↓
Initial Leadinput0.00.540
Deletecomplete baseline output+0.10.539
PMDMcomplete baseline output+8.250.501
DecompOptcomplete baseline output0.00.542
DiffShapecomplete baseline output-0.840.585
ActivityDiffcomplete baseline output+0.140.538
PMDM–MS2complete baseline output+10.670.499
PMDM–MS3complete baseline output+13.080.496
PMDM–MS+EScomplete baseline output+8.620.500
DecompOpt–MS2complete baseline output0.00.542
DecompOpt–MS3complete baseline output+0.010.541
DecompOpt–MS+EScomplete baseline output0.00.541
DiffSBDD–SizeExt(+20)complete baseline output+13.130.480
DiffSBDD–SizeExt(+30)complete baseline output+18.830.468
SurfSpec(Random 2D-Graph)short-arm pseudo-label+4.00.5223
SurfSpec(Short Arm)short-arm pseudo-label+4.00.5216
SurfSpec(Random Patch)short-arm pseudo-label+3.90.5213
SurfSpec(Farthest Patch)short-arm pseudo-label+3.90.5207
SurfSpec(Long Arm)long-arm pseudo-label+9.20.4983

The Short Arm surface-targeted pseudo-labels add only about four heavy atoms, yet already reduce mismatch relative to the initial lead and attain lower mismatch than the matched Random 2D-Graph pseudo-label. They also achieve lower mismatch than Delete, DecompOpt and its multi-step variants, DiffShape, and ActivityDiff, even though these comparison rows are complete optimized outputs rather than first-step reference shapes. SurfSpec(Long Arm) adds about nine heavy atoms and attains a mismatch of 0.4983, which is competitive with PMDM and its multi-step variants while using fewer added atoms than the size-extension controls. These results support the intended mechanism: surface-targeted pseudo-label generation provides a low- or competitive-mismatch reference shape with a substantially smaller edit, rather than obtaining the geometric gain from indiscriminate molecular enlargement.

1.2 Refinement recovers valid outputs while retaining the geometric proposal

The pseudo-label is not the final molecule. It is the geometric anchor for low-noise refinement under the pocket-conditioned molecular prior. Table 2 in Section 5.3, Validation of Pseudo-Label Refinement, reports valence validity 1.00 for SurfSpec together with a mean pseudo-label RMSD of 0.78 Å, showing that refinement recovers valid molecules while retaining the intended geometric modification. The final-output comparison below further shows that the complete pipeline preserves and strengthens the mismatch reduction while producing favorable specificity.

MethodOff-target agnosticSpec ↑ (mean ± SD)Geom. mismatch ↓Occ. ↑Target Vina ↓Clash rate ↓# AtomsDock success ↑
Initial LeadO−0.81±1.43-0.81\pm1.430.5400.26-7.170.0022.751.00
DeleteO−0.81±1.54-0.81\pm1.540.5390.26-7.230.0022.851.00
PMDMO−1.10±2.03-1.10\pm2.030.5010.33-6.470.1631.000.92
DecompOptO−0.93±1.43-0.93\pm1.430.5420.25-6.080.0222.750.91
DiffShapeO−0.88±1.58-0.88\pm1.580.5850.14-6.900.1121.910.99
ActivityDiffX−0.96±2.15-0.96\pm2.150.5380.27-7.070.1122.891.00
PMDM–MS2O−1.38±2.19-1.38\pm2.190.4990.31-8.950.2533.420.91
PMDM–MS3O−2.05±2.93-2.05\pm2.930.4960.31-8.380.3135.830.82
PMDM–MS+ESO−1.02±2.08-1.02\pm2.080.5000.33-8.610.1631.370.87
DecompOpt–MS2O−1.03±1.33-1.03\pm1.330.5420.25-7.080.0322.750.96
DecompOpt–MS3O−0.99±1.39-0.99\pm1.390.5410.25-7.000.0422.760.97
DecompOpt–MS+ESO−0.83±1.59-0.83\pm1.590.5410.25-7.400.0322.750.95
DiffSBDD–SizeExt(+20)O−1.51±2.24-1.51\pm2.240.4800.28-7.570.4535.880.93
DiffSBDD–SizeExt(+30)O−2.20±3.76-2.20\pm3.760.4680.32-7.140.6241.580.81
SurfSpec(Random 2D-Graph)O−1.59±2.24-1.59\pm2.240.5060.30-7.490.0830.80.91
SurfSpec(Long Arm)O−0.37±2.54-0.37\pm2.540.4800.36-8.630.0034.991.00
SurfSpec(Short Arm)O−0.34±2.27-0.34\pm2.270.4830.32-8.910.0531.70.98
SurfSpec(Random Patch)O−0.41±1.95-0.41\pm1.950.4820.33-8.880.0232.20.98
SurfSpec(Farthest Patch)O−0.44±2.27-0.44\pm2.270.4820.32-8.890.0332.10.96

At comparable size, SurfSpec(Random 2D-Graph) has markedly worse mismatch and specificity than every surface-targeted policy. The evidence therefore supports the full algorithm–objective connection:

surface-targeted growth  →  low-mismatch geometric reference  →  valid refined ligand retaining the geometric benefit.\text{surface-targeted growth} \;\rightarrow\; \text{low-mismatch geometric reference} \;\rightarrow\; \text{valid refined ligand retaining the geometric benefit}.

We may add the first-step pseudo-label comparison and final-output ablation to the appendix.

Revised-draft location. Section 4.2, Ligand Growth via Geometric Pseudo-Labeling and Refinement; Section 5.3, Validation of Pseudo-Label Refinement; and Appendix Additional Rebuttal Analyses — Surface-Targeted Proposal and Component Ablations.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.