- 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
1.1 Compared configurations
SurfSpec(). 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 . The same RFdiffusion-generated panel of 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; 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 for each completed ligand candidate 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.
| Method | Off-target agnostic | Spec ↑ (mean ± SD) | Geom. mismatch ↓ | Occ. ↑ | Target Vina ↓ | Clash rate ↓ | # Atoms | Dock success ↑ |
|---|---|---|---|---|---|---|---|---|
| Initial Lead | O | 0.540 | 0.26 | -7.17 | 0.00 | 22.75 | 1.00 | |
| Delete | O | 0.539 | 0.26 | -7.23 | 0.00 | 22.85 | 1.00 | |
| PMDM | O | 0.501 | 0.33 | -6.47 | 0.16 | 31.00 | 0.92 | |
| DecompOpt | O | 0.542 | 0.25 | -6.08 | 0.02 | 22.75 | 0.91 | |
| DiffShape | O | 0.585 | 0.14 | -6.90 | 0.11 | 21.91 | 0.99 | |
| ActivityDiff | X | 0.538 | 0.27 | -7.07 | 0.11 | 22.89 | 1.00 | |
| PMDM–MS2 | O | 0.499 | 0.31 | -8.95 | 0.25 | 33.42 | 0.91 | |
| PMDM–MS3 | O | 0.496 | 0.31 | -8.38 | 0.31 | 35.83 | 0.82 | |
| PMDM–MS+ES | O | 0.500 | 0.33 | -8.61 | 0.16 | 31.37 | 0.87 | |
| DecompOpt–MS2 | O | 0.542 | 0.25 | -7.08 | 0.03 | 22.75 | 0.96 | |
| DecompOpt–MS3 | O | 0.541 | 0.25 | -7.00 | 0.04 | 22.76 | 0.97 | |
| DecompOpt–MS+ES | O | 0.541 | 0.25 | -7.40 | 0.03 | 22.75 | 0.95 | |
| DiffSBDD–SizeExt(+20) | O | 0.480 | 0.28 | -7.57 | 0.45 | 35.88 | 0.93 | |
| DiffSBDD–SizeExt(+30) | O | 0.468 | 0.32 | -7.14 | 0.62 | 41.58 | 0.81 | |
| SurfSpec(Long Arm) | O | 0.480 | 0.36 | -8.63 | 0.00 | 34.99 | 1.00 | |
| SurfSpec(Short Arm) | O | 0.483 | 0.32 | -8.91 | 0.05 | 31.7 | 0.98 | |
| SurfSpec(JS-Guided) | O | 0.470 | 0.33 | -9.07 | 0.04 | 32.5 | 0.98 | |
| SurfSpec() | O | 0.486 | 0.32 | -9.20 | 0.02 | 31.9 | 0.97 |
The molecular-property comparison among the matched three-step variants is:
| Method | QED ↑ | SA ↓ | LE ↑ |
|---|---|---|---|
| SurfSpec(Short Arm) | 0.405 | 4.33 | 0.286 |
| SurfSpec(JS-Guided) | 0.375 | 4.23 | 0.289 |
| SurfSpec() | 0.442 | 4.26 | 0.310 |
SurfSpec() improves mean specificity from to 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 without changing the proposal-and-selection structure.
Revised-draft location. Appendix C.1, SurfSpec Variants and Rebuttal Configurations, defines SurfSpec(), SurfSpec(JS-Guided), and the matched baselines. Appendix E.3, Synthetic-Competitor Candidate Selection, gives the interaction objective, implementation, and baseline-inclusive results.
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), which augments the geometric objective with a ligand-specific interaction term. SurfSpec(ΔJ) first generates surface-directed candidates and then selects the candidate with the strongest target-versus-competitor interaction gain.
Let T denote the target pocket, L a ligand, and u(R,L)=−SVina(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)],
where p(P∣L) is the conditional distribution of plausible pockets that can interact with L. We estimate the expectation by Monte Carlo sampling from an RFdiffusion-based ligand-conditioned pocket generator:
J(T,L)=u(T,L)−Npanel1∑k=1Npanelu(Pk,L),Pk∼p(P∣L),
with Npanel=8. The interaction-aware geometric objective is
U(T,L)=dgm(L,T)−J(T,L).
The two terms are optimized sequentially rather than added on an uncalibrated common numerical scale. For a candidate La grown from the current step input L0, we compute
ΔJ(La;L0)=J(T,La)−J(T,L0),
or equivalently,
ΔJ(La;L0)=[u(T,La)−u(T,L0)]−Npanel1∑k=1Npanel[u(Pk,La)−u(Pk,L0)].
Thus, a candidate is rewarded when its target-affinity gain exceeds its average affinity gain across the synthetic competitor panel. Since J(T,L0) is shared by all candidates within the same step, ranking by ΔJ is equivalent to ranking by J(T,La).