PaperlayerView exact revision on arXiv

Source-linked AI summary

RIDE: Reference-Anchored Inference-Time Diffusion Editing for Scaffold Hopping

Ruoxi Gao, Frazier N. Baker, Trieu Nguyen, Xia Ning

arXiv:2609.35623v1cs.AIcs.LG

TL;DR

Scaffold hopping seeks structurally distinct molecules that retain functional groups and resemble a reference ligand in 3D, but existing diffusion approaches lack a principled way to jointly promote novelty and preserve shape. RIDE edits a recovered reference diffusion trajectory through segment selection and value-guided sampling. Across benchmarks, it consistently produces lower 2D and higher 3D similarity than baselines.

  • Problem

    Existing diffusion scaffold-hopping methods condition generation on functional groups but lack a principled mechanism to jointly enforce 2D novelty and preserve the reference ligand’s 3D shape.

    Source: §1 · p. 1

  • Method

    RIDE recovers the reference diffusion noise trajectory, selects a segment for editing via noise perturbation, and performs value-guided scaffold sampling.

    Source: §1 · pp. 1–2

  • Results

    11.7% lower 2D similarity and 7.3% higher 3D similarity on average: RIDE consistently improves both measures over baselines.

    Source: §1 · p. 2

  • Takeaways & Limitations

    RIDE accommodates different reward functions and can preserve 3D similarity even when it is not explicitly included in the reward.

    Source: §1 · p. 2

Abstract

from arXiv · show

Scaffold hopping is a critical task in drug discovery, which seeks to discover new, structurally distinct molecules that share key functional groups and similar 3D shape with a reference binding ligand. Existing diffusion-based scaffold hopping methods formulate the problem as conditional generation of scaffolds given the functional groups. However, they lack a principled mechanism to jointly enforce 2D structural novelty and preserve the 3D shape of the reference ligand. Here, we introduce RIDE, a Reference-anchored Inference-time Diffusion Editing framework for scaffold hopping. RIDE recovers the reference diffusion noise trajectory conditioned on the binding pocket and functional groups, selects an optimal trajectory segment for editing via noise perturbation, and conducts a value-guided scaffold sampling to generate new scaffolds. Extensive experimental results demonstrate that, compared to baselines, RIDE consistently generates scaffolds with lower 2D similarity and higher 3D similarity to the reference, with an average improvements of 11.7% and 7.3%, respectively. Further analysis reveals that RIDE can accommodate various reward functions, and can preserve 3D similarity even when this is not explicitly included in the reward. Two case studies illustrate RIDE's ability to generate distinct scaffolds with different structures and properties, and its ability to introduce substantial 2D variation while maintaining very high 3D similarity. RIDE is publicly available at https://anonymous.4open.science/r/RIDE-C8A0.

1 INTRODUCTION

Scaffold hopping seeks new molecules with distinct structures but shared functional groups and similar 3D shape; existing diffusion approaches lack a principled way to jointly achieve novelty and preserve shape. RIDE edits a recovered reference trajectory and, across 60 tasks, improves the 2D- and 3D-similarity profile over baselines.

1 INTRODUCTION

  • Scaffold hopping seeks structurally distinct molecules that retain key functional groups and a reference ligand’s similar 3D shape.It is motivated by exploring varied structures with similar binding but different properties, synthesizability, and patentability.Source: §1 · p. 1
  • Diffusion scaffold-hopping methods condition scaffold generation on functional groups but leave structural divergence largely to chance, while SBDD and LBDD approaches each lack one of the desired controls.SBDD approaches lack explicit guarantees of substantial 2D novelty; LBDD approaches lack explicit protein-ligand binding guidance.Source: §1 · p. 1
  • RIDE recovers the reference diffusion noise trajectory, selects a segment for editing via noise perturbation, then uses value-guided sampling to generate new scaffolds.Its reward can jointly measure 2D novelty and 3D shape similarity to guide editing and sampling.Source: §1 · pp. 1–2
  • 11.7% lower 2D similarity and 7.3% higher 3D similarity on average: RIDE consistently improves both measures over baselines across 60 scaffold hopping tasks.The tasks involve protein-ligand complexes related to major human diseases.Source: §1 · p. 2

2 RELATED WORK

Scaffold-hopping methods use structure-based or ligand-based generation, but these approaches emphasize different objectives. RIDE incorporates a reference scaffold into structure-based generation to target low 2D similarity, high 3D similarity, and strong binding affinity.

2 RELATED WORK

  • Structure-based and ligand-based scaffold hopping emphasize binding affinity and reference 3D similarity, respectively, while current formulations do not explicitly optimize 2D similarity to the reference scaffold.RIDE incorporates the reference scaffold into structure-based generation to target low 2D similarity and high 3D similarity while maintaining strong binding affinity.Source: §2 · p. 2
  • Structure-based drug-design methods generate molecules for protein pockets without a reference ligand, whereas ligand-based methods use known binding ligands to guide 3D shape similarity.Examples include IPDiff, which uses pocket-ligand interaction embeddings, and ShEPhERD, which jointly learns structure, shape, electrostatics, and pharmacophore.Source: §2 · p. 2
  • Scaffold hopping can be formulated under structure-based or ligand-based design through inpainting or conditioning on functional groups.The passage identifies both strategies as ways generative models formulate scaffold hopping.Source: §2 · p. 2

3 PRELIMINARIES

The preliminaries describe recovering reference states and their corresponding noise trajectories through diffusion inversion, including extensions to categorical states. They also introduce value-based inference-time optimization, which estimates expected final-state reward for intermediate states and can be reformulated using sampled outputs.

3.1 INVERSION-BASED DIFFUSION METHODS

  • DDIM inversion maps a reference sample back to intermediate states, providing a recovered state trajectory that can be represented by a corresponding noise sequence.The trajectory spans diffusion timesteps; the passage also describes inversion extensions to categorical diffusion.Source: §3.1 · p. 3
  • For categorical diffusion, inversion uses a categorical forward transition and a predefined noise schedule; the positive-part operator and a small margin support strict category selection.The recovered categorical reference trajectory corresponds to a noise sequence.Source: §3.1 · p. 3

3.2 VALUE-BASED METHODS FOR INFERENCE-TIME OPTIMIZATION IN DIFFUSION

  • Value-based inference-time optimization estimates an intermediate state’s expected future reward from final states sampled under the diffusion model.The reward is defined on the final state.Source: §3.2 · p. 3
  • A reformulated value estimate uses noise-prior lookahead samples and connects each intermediate state to those outputs through the forward transition.The two steps are to sample final states and rewards independently of the intermediate state, then compute its value in closed form.Source: §3.2 · p. 3

4 METHODS

RIDE reframes scaffold hopping as editing a reference scaffold, addressing the limitation that conditional generation does not explicitly use reference structure to balance 2D novelty and 3D similarity. It recovers and perturbs the reference noise trajectory, then uses value-guided sampling to generate candidate scaffolds.

4.1 PROBLEM DEFINITION AND NOTATIONS

  • Existing conditional scaffold generation does not explicitly use the reference scaffold to tailor the desired 2D and 3D similarity profile.A ligand is represented by scaffold S and functional groups F; the functional groups correspond to atoms involved in key pocket interactions.Source: §4.1 · pp. 3–4

4.2 REFERENCE-ANCHORED SCAFFOLD GENERATION VIA NOISE EDITING

  • RIDE formulates scaffold hopping as reference-anchored generation, optimizing a distribution that retains reference 3D shape while encouraging 2D deviation.The method modifies inference-time sampling of the base scaffold-generation model.Source: §4.1 · p. 4, §4.2 · p. 4

4.2.1 REFERENCE NOISE TRAJECTORY RECOVERY

  • RIDE recovers the state and noise trajectory that reproduce the reference scaffold under the base diffusion model.The model uses Gaussian diffusion for atom positions and categorical diffusion for atom types.Source: §4.2.1 · p. 4

4.2.2 REFERENCE TRAJECTORY EDITING VIA NOISE PERTURBATION

  • RIDE perturbs noise only within a selected segment of the reference trajectory, retaining the remaining noise to help preserve 3D shape and binding pose.Within the segment, it resamples position and type noise; the segment length L is a hyperparameter.Source: §4.2.2 · pp. 4–5
  • For each candidate starting timestep, RIDE scores K Monte Carlo samples and selects a segment start prioritizing reduced 2D similarity.The recovered trajectory already supports 3D-shape preservation, and the selected segment provides a starting point for subsequent optimization.Source: §4.2.2 · p. 5

4.2.3 VALUE-GUIDED SCAFFOLD SAMPLING

  • Within the selected segment, RIDE estimates intermediate scaffold-state values and greedily selects noise and next states to guide sampling.Iteratively selecting states along edited trajectories yields a new transition distribution for scaffold hopping.Source: §4.2.3 · pp. 5–6

5 EXPERIMENTS

Across 60 human-disease-related scaffold-hopping tasks, RIDE improves reference similarity profiles over its base models and baselines. Its results also show that guided sampling and iterative hopping affect novelty, shape similarity, and binding-related properties.

5.1 EXPERIMENTAL SETTING

  • The evaluation uses 60 complexes with 58 unique ligands, and RIDE’s atom-level trajectory recovery constrains edited scaffolds to the reference atom count.The test set excludes complexes with no or all ligand atoms identified as functional-group atoms; baselines use the same scaffold size for fair comparison.Source: §5.1 · p. 6

5.2.1 OVERALL COMPARISON

  • 11.7% lower Sim2D and 7.3% higher Sim3D on average versus corresponding base models; these gains exclude ShEPhERD.Table 1 compares RIDE with baselines on the test set using L = 100.Source: §5.2.1 · pp. 6–7, Table 1 · p. 7
  • RIDE(1) improves both Sim2D and Sim3D across conDitar-a, IPDiff-a, and DiffHopp, while its Sim3D exceeds ShEPhERD’s.On DiffHopp, RIDE(1) also improves Vina S from −2.223 to −4.593; across base models, Conn.(%) improves by 5.0% on average, while QED and SA dip slightly.Source: §5.2.1 · p. 7
  • RIDE(1) with QED rewards raises QED by 7.0% for conDitar-a and 7.1% for IPDiff-a; two-hop gains are 10.3% and 16.9%, respectively.Without explicitly optimizing Sim3D, it remains above the corresponding baselines by 8.1% for conDitar-a(QED) and 4.8% for IPDiff-a(QED).Source: §5.2.1 · pp. 7–8
  • RIDE(2) generally lowers Sim2D further while keeping Sim3D above 0.890 on conDitar-a and IPDiff-a, but average Vina S shifts from −7.484 to −7.257.The authors suggest accumulated deviations across hops may degrade binding affinity, highlighting a trade-off between the similarity profile and affinity.Source: §5.2.1 · p. 8

5.2.2 COMPARISON OF RANDOM PERTURBATION AND VALUE-GUIDED SAMPLING

  • 8.3% lower average Sim2D and 1.6% higher Sim3D with value-guided sampling than random perturbation across four settings.The comparison uses conDitar-a and pairs T/10 versus T with RIDE(1) versus RIDE(2); the reported gains are larger for T than T/10.Source: Table 2 · p. 8, §5.2.2 · p. 8

1 AND FIXED t1

The reference-optimal perturbation start provides a strong starting point for value-guided sampling, while perturbation timing shifts the Sim2D–Sim3D balance. Case studies show generated scaffolds with varied structures and properties while retaining high 3D similarity.

5.2.3 COMPARISON OF REFERENCE-OPTIMAL t∗

  • The reference-optimal t*1 gives the best Sim2D with competitive Sim3D against fixed t1 values; later perturbation improves Sim3D at the cost of Sim2D.RIDE(2) shows lower Sim2D and Sim3D than RIDE(1), with smaller variations across candidate t1 values, indicating less sensitivity to segment choice.Source: §5.2.3 · pp. 8–9, Figure 2 · p. 9

5.2.4 CASE STUDIES

  • AK1BA examples span linker flexibility: 6 rotatable bonds in one generated linker, versus 5 in the reference and 4 in another generated scaffold.The more flexible example has slightly lower Vina S; the examples have low Sim2D and high Sim3D relative to the reference.Source: §5.2.4 · p. 9, Figure 3 · p. 9
  • TNKS2 examples retain very high Sim3D while changing atom types, bonds, ring size, and linker structure.Both replace one ketone with an alcohol and use carbon to bind the two cyclic functional groups; one also replaces a benzene with cyclopentane and adds a rotatable bond.Source: §5.2.4 · p. 9, Figure 4 · p. 9

6 CONCLUSION

RIDE generates scaffolds with low 2D similarity and high 3D similarity to the reference across multiple settings. The authors identify additional molecular-property optimization and adaptive perturbation as directions for future work.

6 CONCLUSION

  • RIDE consistently generates scaffolds with low 2D similarity and high 3D similarity to the reference scaffold.Source: §6 · p. 10
  • RIDE is effective under different base models, reward functions, perturbation segment lengths, and numbers of scaffold-hopping iterations.Source: §6 · p. 10
  • Future work may extend RIDE to optimize additional molecular properties and improve molecule quality, and investigate more adaptive perturbation strategies.Source: §6 · p. 10

AI USE STATEMENT

The authors used AI tools during preparation to assist with language editing and literature identification. They reviewed and verified all AI-assisted content.

AI USE STATEMENT

ETHICS STATEMENT

The authors note that RIDE may generate harmful molecules and call for responsible human expert supervision. They specify safety review before laboratory synthesis or testing and adherence to applicable rules and practices.

ETHICS STATEMENT

  • The authors acknowledge that RIDE could generate harmful content, including unsafe scaffolds and molecules.Source: ETHICS STATEMENT · p. 10
  • Human expert chemists should verify generated scaffolds and molecules for safety before laboratory synthesis or testing.Source: ETHICS STATEMENT · p. 10
  • Users are urged to follow applicable safety regulations, ethical guidelines, laws, and professional best practices.Source: ETHICS STATEMENT · p. 10

REPRODUCIBILITY STATEMENT

The authors make the source code and datasets public. They report experimental protocol in Section 5.1 and hyperparameters in Appendix A.6.

REPRODUCIBILITY STATEMENT

A.1 ADAPTING DIFFUSION-BASED SBDD MODELS TO SCAFFOLD HOPPING

The supplementary sections describe how pretrained SBDD models are adapted for scaffold hopping, outline RIDE’s sampling and evaluation procedures, and report additional experiments on inversion, reward choice, and implementation.

A.1 ADAPTING DIFFUSION-BASED SBDD MODELS TO SCAFFOLD HOPPING

  • Adaptation fixes the reference functional groups, restricts diffusion loss to scaffold atoms, and updates diffusion parameters; the authors note that inference-time inpainting may cause distribution mismatch.The mismatch may lead to invalid structures or unfavorable conformations and make trajectory modifications harder to control.Source: §A.1 · p. 14

A.2.4 VALUE-GUIDED SAMPLING

  • Within the selected perturbation segment, value-guided sampling proposes noise candidates at selected timesteps and uses estimated value to choose scaffold-state candidates.The process uses lookahead samples and proceeds through reverse transitions.Source: §A.2.4 · p. 15

A.4 BASELINES FOR EXPERIMENTS

  • Experiments compare adapted conDitar and IPDiff, DiffHopp, and ShEPhERD; the first three also serve as RIDE base models, while ShEPhERD is not used as one because it relies on inpainting.The models are evaluated directly as scaffold-hopping baselines.Source: §A.4 · p. 18

A.5 EVALUATION METRICS

  • Evaluation measures reference similarity with Sim2D and Sim3D, drug-development properties with binding-affinity and drug-likeness metrics, and scaffold connection success.For each reference, 100 samples are generated, and only valid, connected molecules are evaluated.Source: §A.5 · p. 19

A.6 ANCHOR MOLECULE SELECTION FOR RIDE(2)

  • For RIDE(2), the anchor is selected by retaining the top 30% of one-hop molecules by 3D similarity, then choosing the lowest-Sim2D candidate; IPDiff’s interaction embeddings are fixed to those from the anchor ligand across both hops.RIDE(1) and RIDE(2) use different reward weights and diffusion-step settings across base models.Source: §A.6 · p. 19, §A.7 · p. 19, §A.8 · p. 19

A.9 ADDITIONAL EXPERIMENTAL RESULTS

  • Across t1 values, lower Sim2D accompanies changes in QED, and QED and Sim3D respond similarly; dynamic timestep selection identifies a starting point balancing all three.The figure compares Sim2D with QED and with Sim3D across t1 values.Source: §A.9 · p. 19, Figure 5 · p. 20
  • Trajectory inversion improves both Sim2D and Sim3D; with RIDE(1), Sim3D rises from 0.833 to 0.866 at t∗ 1 and from 0.852 to 0.882 after search.The comparison uses conDitar-a; the table caption identifies generation results with and without inversion.Source: §A.9 · pp. 19–20, Table 3 · p. 20
  • Across tested λ values, larger λ is associated with lower Sim2D and gradually decreasing Sim3D, while QED and SA remain relatively stable; λ = 0.5 offers a favorable tested trade-off.Vina S varies moderately, with better values generally observed at higher Sim3D.Source: §A.9 · p. 20
Loading 2609.35623v1…