acceptodds
Under review as a conference paper at ICLR 2027

Blending experimental electron density information to protein-ligand interaction

Abstract

Machine learning models for protein-ligand interactions commonly rely on atomic coordinates, which are constructed by fitting molecular models to experimental electron-density maps. However, most models represent each complex as a single fitted structure, which can omit information about conformational heterogeneity, positional uncertainty, and local density support for individual atoms at binding interfaces. To fill this gap, we present CoDE, a "Co"ordinate and "D"ensity "E"ncoder, pre-trained on experimental density maps. We represent each binding site on a shared three-dimensional voxel grid combining atomic occupancy, experimental electron density, and its gradient magnitude. Using approximately 100K curated complexes, we pre-train a channel-wise vision transformer to jointly reconstruct masked atomic and density channels, learning a representation that integrates structural and experimental information. We evaluate the frozen encoder on two downstream tasks: (i) binding-affinity regression under leakage-controlled splits and (ii) pocket-conditioned ligand generation. Across these settings, CoDE improves over an identically pre-trained coordinate-only encoder and remains competitive with task-specific baselines, showing that density complements fitted atomic coordinates for protein–ligand interaction modeling and that the density-blended representation is useful for structure-based ligand generation.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Experimental density map, binding affinity, drug design

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 4 reviews, above 85% of the others.
Reviewer_RiUROpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces CoDE, a Coordinate and Density Encoder that combines atomic occupancy, experimental electron density, and density gradient magnitude on a shared voxel grid. A Channel-ViT is pretrained through joint masked reconstruction on PLINDER complexes. The frozen representations are then used for binding-affinity regression and pocket-conditioned ligand generation with VoxBind. The results show improvements over the matched coordinate-only encoder.

Critical strengths
  1. The motivation is convincing. Experimental electron-density maps contain information that a single fitted structure does not fully capture, and the paper investigates whether this information can improve protein–ligand modeling.

  2. The CoE baseline and channel ablations support the contribution of experimental electron density.

Critical weaknesses
  1. My main concern is how far these results generalize beyond the PLINDER pretraining data. The authors exclude downstream validation and test entries by exact PDB ID, but complexes with different IDs may still have similar pockets, or ligands. These similarities are not reported. It remains unclear whether density also helps on complexes with few close neighbors in the pretraining set.

  2. The chemical interpretation of the improvements remains limited. It is unclear which binding-site features are better represented with electron density and why this improves affinity predictions. A few representative examples linking density features, binding interactions, and prediction changes would make the explanation more convincing.

Questions for authors
  1. Could the authors report nearest-neighbor similarities between the test and pretraining sets for pockets and ligands separately? With the existing model, does CoDE still improve over CoE and other baseline on low-similarity subsets?

  2. For binding-affinity regression, could the authors analyze a few representative complexes where electron density clearly helps, linking the relevant electron density features and specific binding interactions to changes in model predictions? This would help explain the chemical basis of the improvement beyond the aggregate metrics.

3 comments

We thank the reviewer for the comments on generalization and chemical interpretation, and address both below with new analyses. We highlight additional content and modifications in the revised paper in dark green.

[W1, Q1] Generalization under pocket or ligand similarity

In Appendix F.2 we report nearest-neighbor similarity to the pre-training corpus for pockets and ligands separately: sequence identity with BLAST+ [1] following ATOM3D [2], and ligand ECFP4 Tanimoto. While the ligand similarity is low, many test proteins have a close sequence neighbor. Stratifying the test set leaves too few complexes per bin, so we instead removed every protein at or above 30% identity to a CL3 test protein from the corpus (nearly 12%) and the downstream train split.

We additionally checked the new pre-training corpus with PLINDER’s pocket-lDDT clusters [3, 4], which come from Foldseek alignment rather than sequence. As a result, the fraction of test complexes whose nearest pre-training neighbor has pocket-lDDT ≥\geq 95 decreased from 67.9% to 0.5%.

MethodPre-trainingPearson r↑r \uparrowSpearman ρ↑\rho \uparrowRMSE ↓\downarrow
DeepDTA✗0.481 ± 0.0120.448 ± 0.0131.536 ± 0.022
MolTrans✗0.363 ± 0.1760.344 ± 0.1421.629 ± 0.106
HonestAffinity$^{\ddagger}$✓0.331 ± 0.0400.309 ± 0.0441.731 ± 0.075
Nesso-1$^{\ddagger}$✓0.6730.6591.367
HBGSA✗0.522 ± 0.0160.496 ± 0.0211.547 ± 0.043
EGNN✗0.561 ± 0.0210.532 ± 0.0211.486 ± 0.058
EGNN$_{\text{TargetDiff}}$$^{\ddagger}$✓0.557 ± 0.0220.527 ± 0.0221.483 ± 0.029
GET✗0.585 ± 0.0050.553 ± 0.0051.421 ± 0.021
CheapNet✗0.567 ± 0.0240.529 ± 0.0221.496 ± 0.061
IPNet✗0.540 ± 0.0280.515 ± 0.0151.476 ± 0.038
IPNet$^{\ddagger}$✓0.487 ± 0.0070.432 ± 0.0051.609 ± 0.017
AEV-PLIG✗0.452 ± 0.0310.408 ± 0.0311.637 ± 0.033
DSMBind$^{\ddagger}$✓0.4640.480N/A
BindNet\textcolor{darkgreen}{BindNet}\textcolor{darkgreen}{\text{✓}}0.435 ± 0.0320.439 ± 0.0151.655 ± 0.068
ProFSA$^{\ddagger}$✓0.616 ± 0.0060.582 ± 0.0061.461 ± 0.015
CoE✓0.617 ± 0.0020.581 ± 0.0021.488 ± 0.054
CoDE (Ours)✓0.630 ± 0.0010.610 ± 0.0011.452 ± 0.060

We report the sequence similarity-filtered results in Table 13, where most models drop in performance. Nonetheless, the ordering is preserved, and CoDE stays competitive and improves over CoE. Models marked with ‡\ddagger are pre-trained on a corpus we cannot filter and contain proteins sequence-similar to CL3 test proteins, reported for context.

[1] BLAST+: architecture and applications, BMC Bioinformatics 2009

[2] ATOM3D: Tasks On Molecules in Three Dimensions, NeurIPS 2021

[3] PLINDER: The protein-ligand interactions dataset and evaluation resource, bioRxiv 2024

[4] lDDT: a local superposition-free score for comparing protein structures and models using distance difference tests, Bioinformatics 2013

[W2, Q2] Representative examples linking density, interaction, and prediction changes

We have added representative examples in Appendix F.3, linking density features and binding interactions to changes in prediction for binding-affinity regression. We selected complexes with the largest reduction in the absolute error from CoE to CoDE, computed the protein-ligand interactions with ProLIF [1], and examined what each model receives as input at those interaction sites. At the interaction sites identified by ProLIF, the experimental map carries density in the region between the interacting groups, where the atom-blob channels fall off. CoE receives only the atom-centered channels and therefore has little input in that region.

[1] ProLIF: a library to encode molecular interactions as fingerprints, Journal of Cheminformatics 2021

Reviewer_RiUROpenReview ↗

Thank you for your response

Dear Authors,

Thank you for your response and the additional experiments. They address most of my concerns.

One point remains: the pre-training status of BindNet in Table 13 differs from what is reported in your response to my review. Please verify which designation is correct and update the table or response accordingly.

The experiments provide a convincing answer to the paper’s central research question, and I am therefore raising my score.

Best regards.

Dear reviewer RiUR, We are happy to hear that the additional experiments have addressed most of your concerns.

Regarding the pre-training status of BindNet in the response, the entry should be ✓, as shown in Table 13 of the revised paper. We have updated the table in our response and noted it in dark green.

Thank you again

Reviewer_2EGhOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper pre-trains a protein-ligand system by co-encoding the atomic coordinates, experimental electron density, and density gradient, and then uses the pre-trained output to assist ligand generation. Experiments show that this pre-training improves affinity prediction and docking scoring.

Critical strengths
  1. This paper uses the modality of experimental electron density and employs a pretraining method to facilitate its transfer across different tasks.
Critical weaknesses
  1. The authors need to provide a model using the B factor/occupancy for comparison. Currently, the values ​​compare coordinates versus coordinates + complete electron density. However, is the information provided by complete electron density more valuable than directly providing information about structural uncertainties? This point has not yet been discussed.

  2. The authors considered the potential data leakage issue during pre-training, which is good, but I would prefer to see an evaluation of sequence similarity, ideally accurate to a certain similarity ratio. If this experiment is conducted, I would be willing to raise the score.

  3. There's a numerical issue with the RMSE in Table 7: Based on the interpretation of "selecting a subset of test results from identical predictions," the values ​​are unlikely to hold true.

CL1: 1166 samples, CL2: 1149 samples, a difference of only 17 samples. While removing samples with smaller errors might indeed increase the average error, the sum of the squared errors of the remaining 1149 samples couldn't possibly exceed the original sum by that many times. Therefore, the upper limits for HBGSA and BindNet are approximately 1.640 and 1.589, respectively. I hope the authors can explain the reasoning behind these calculations. An unreasonable explanation could lower the credibility of the entire paper.

  1. The main text §3, p.4, L206–210 states that the generator and projection are trained after pre-training and the reconstructed head generation is discarded. Appendix B.3, p.23, L1195–1199 also states that the frozen patch decoder converts the encoder tokens into voxels and then fuses them through trainable convolutions. Please correct this contradiction. Which decoder is it exactly?

  2. The main experiment's encoder pre-training, according to the paper, involved 100 epochs of training, with the main affinity results using the epoch 25 checkpoint. However, the ablation experiment's encoder reportedly trained for 50 epochs, indicating the use of different pre-training checkpoints. Nevertheless, the complete model CoDE and CoE (with density and gradient removed) in Table 2 are consistent with the results of the corresponding pre-trained model in Table 1 on the CL3 test set.

Page 22, Appendix B.2, lines 1145–1146: The main experiment explicitly uses the epoch 25 checkpoint. “We pre-train for 100 epochs with the settings of Table 6 and use the epoch-25 checkpoint for the main affinity results.” Page 22, Table 6, Pre-training section, approximately lines 1172–1173: This is confirmed again. This clearly distinguishes between the total number of training epochs and the actual checkpoints used.

Page 24, Appendix B.5, lines 1244–1247: Ablation uniform training for 50 epochs. “All ablations change only the input or target channels of the CoDE recipe. They share the architecture, PLINDER data, masking ratio of 0.75, atom-biased masking, channel weighting, optimizer, EMA, and seed, and run for 50 epochs. Each row of Table 2 averages the five-seed CL3 probe means and standard deviations.”

The corresponding replication results are located in the CL3 column of Table 1 on page 4 and Table 2 on page 6.

I need a more reasonable explanation. Specifically, does the 50 epochs of ablation training refer to post-training?

Questions for authors
  1. Additional experiments using B factor/occupancy as network input.

  2. Explain the extent of this data leak (sequence similarity).

  3. Explain the contradictions among the three data points.

2 comments

We thank the reviewer for the careful reading and the detailed checks that uncovered errors. We have corrected all of them and addressed each point below. We have highlighted additional content and modifications in the revised paper in dark green.

[W1, Q1] Structural uncertainty baselines

We have added a B-factor-based baseline as a structural-uncertainty signal, CoBE, and report it in Appendix F.1. We encode each atom's B-factor on its atom blob so that per-atom size is respected, and keep all other configurations identical. CoBE matches or improves over the coordinate-only CoE, and it still falls behind CoDE.

MethodPearson r↑r \uparrowSpearman ρ↑\rho \uparrowRMSE ↓\downarrow
CoE0.614 ±\pm 0.0020.574 ±\pm 0.0011.416 ±\pm 0.045
CoBE0.617 ±\pm 0.0030.595 ±\pm 0.0021.390 ±\pm 0.027
CoDE0.643 ±\pm 0.0020.620 ±\pm 0.0031.342 ±\pm 0.009

(CL3 filtered test, N=733)

[W2, Q2] Sequence similarity considered results

In Appendix F.2, we report the sequence similarity between the pre-training data and the CL3 test complexes, computed with BLAST+ [1] following ATOM3D [2]. We found that test complexes highly similar to the pre-training corpus exist; therefore, we additionally pre-trained on a corpus filtered at 30% sequence identity against the CL3 test set. Approximately 12% of the pre-training data corpus was removed, and the downstream train split was also filtered. As seen in Table 13, this results in lower performance for most models, but the ordering is preserved. CoDE still stays competitive against baselines, and improves over CoE. Models marked ‡\ddagger are pre-trained baselines containing proteins above 30% identity to the test set and are reported for context.

MethodPre-trainingPearson r↑r \uparrowSpearman ρ↑\rho \uparrowRMSE ↓\downarrow
DeepDTA✗0.481 ± 0.0120.448 ± 0.0131.536 ± 0.022
MolTrans✗0.363 ± 0.1760.344 ± 0.1421.629 ± 0.106
HonestAffinity$^{\ddagger}$✓0.331 ± 0.0400.309 ± 0.0441.731 ± 0.075
Nesso-1$^{\ddagger}$✓0.6730.6591.367
HBGSA✗0.522 ± 0.0160.496 ± 0.0211.547 ± 0.043
EGNN✗0.561 ± 0.0210.532 ± 0.0211.486 ± 0.058
EGNN$_{\text{TargetDiff}}$$^{\ddagger}$✓0.557 ± 0.0220.527 ± 0.0221.483 ± 0.029
GET✗0.585 ± 0.0050.553 ± 0.0051.421 ± 0.021
CheapNet✗0.567 ± 0.0240.529 ± 0.0221.496 ± 0.061
IPNet✗0.540 ± 0.0280.515 ± 0.0151.476 ± 0.038
IPNet$^{\ddagger}$✓0.487 ± 0.0070.432 ± 0.0051.609 ± 0.017
AEV-PLIG✗0.452 ± 0.0310.408 ± 0.0311.637 ± 0.033
DSMBind$^{\ddagger}$✓0.4640.480N/A
BindNet✗0.435 ± 0.0320.439 ± 0.0151.655 ± 0.068
ProFSA$^{\ddagger}$✓0.616 ± 0.0060.582 ± 0.0061.461 ± 0.015
CoE✓0.617 ± 0.0020.581 ± 0.0021.488 ± 0.054
CoDE (Ours)✓0.630 ± 0.0010.610 ± 0.0011.452 ± 0.060

[1] BLAST+: architecture and applications, BMC Bioinformatics 2009

[2] ATOM3D: Tasks On Molecules in Three Dimensions, NeurIPS 2021

[W3, Q3] Numerical issue of RMSE in Table 7

We thank the reviewer for the arithmetic check, and apologize for the confusion. The CL tiers are not filters applied only to the test split; each tier restricts the whole LP-PDBBind split, including training and validation. While CL3 ⊂\subset CL2 ⊂\subset CL1 ⊂\subset Full holds, moving from CL1 to CL2 removes 17 test complexes and also 78 training complexes. Therefore, results in Table 7 come from separately trained models, and a bound that assumes identical predictions on the shared complexes does not apply. We have added the table below in Appendix A.1 to make the data counts explicit.

TierTrainValidationTest
Full38508171320
CL127216801166
CL226436591149
CL31559410733

Separately, we found that our CL3 runs had been trained on the training split of a different tier. We have retrained them on the correct CL3 training split and updated all CL3 results throughout the paper; the conclusions are unchanged, and CoDE remains competitive.

[W4] Decoder configuration specification

We apologize for the confusion on the decoder configuration, and have updated the manuscript to remove this ambiguity. The two parts refer to different modules. The frozen patch decoder of Appendix B.3 is the patch-to-voxel projection that belongs to the encoder: it maps each patch token back to its voxel block and yields the 16-channel grid used as xdensx_\text{dens}, and stays frozen for downstream tasks. The reconstruction head is a separate two-layer MLP predicting the 13 input channels, and it is discarded after pre-training. The trainable convolution in Appendix B.3 is a third module, the fusion convolution combining xdensx_\text{dens} with VoxBind's features; the main text previously referred to this module as a projection as well.

[W5] Training epoch configuration specification

We apologize for the confusion, which stems from an error in our manuscript. All runs, including the main encoder and every ablation, use the same 50-epoch recipe, and the best validation checkpoint is selected (for CoDE, this was the epoch-25 checkpoint). The 100-epoch setting listed in the Appendix table was carried over from an additional experiment to confirm that the encoder was not under-trained, and should not have been reported as the main recipe. We have corrected both places, and note that there is no post-training stage. CoDE and all other ablations differ only in the selection of input channels; all other configurations are identical.

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

"BLENDING EXPERIMENTAL ELECTRON DENSITY INFORMATION TO PROTEIN-LIGAND INTERACTION" investigate whether electron density maps can compliment atomic coordinates and the combined info carry information about protein–ligand interactions that atomic coordinates alone do not as electron densities measures uncertainty of atomic positions, thus taking conformational and positional ambiguity into account. They train a channel-wise vision transformer to jointly reconstruct atomic and density channels that they call CoDE. Then they show promising results on binding-affinity regression, ligand generation conditioned on pockets, and more such as B-factor decoding.

Critical strengths

1: Authors appear to take very great care to avoid leakage in their evaluations, have many controls, ablations, and benchmarks (not the boltz family). As bonus the paper is well-written with good figures underlying their points.

2: Author provide an interesting story on how electron density might provide useful information towards ligand binding and ligand generation.

Critical weaknesses

1: While, the authors ablate the effects of density etc an interesting test to see if density and grad mag gives information (and thus the gain for CoDE over other methods) due to either a) giving more information on the positional uncertainty etc as mentioned in abstract or b) it is because densities were made together with ligands and this gives implicit information about the ligand and its order which correlates with affinity. Swapping the electron density maps of proteins, if possible, to maps they have with other ligands would support the claims of the abstract.

2: Figure 3 would also benefit from showing readers the actual distributions and not just mean and median to understand the statistics. Also, as a measure of a strong good-binder generator is to be able to generate the best binders, thus showing the statistics and provide metrics for best in class would show whether CoDE provides an advantage.

Questions for authors

Could benchmarking to additional models eg Boltz family be done?

Can figure 3 be extended to show metrics for best binders?

1 comment

We thank the reviewer for the thoughtful comments on what the density channels actually encode, and address each point below. We have highlighted additional content and modifications in the revised paper in dark green.

[W1] Ablation on density effects

Following QESD's suggestion, we swapped each complex's density map with maps of the same protein bound to a different ligand. This was possible for 336 complexes in the CL3 test split, and all settings below are evaluated on this subset.

Model (inference setting N=336N=336)Pearson r↑r \uparrowSpearman ρ↑\rho \uparrowRMSE ↓\downarrow
CoDE (true density map)0.629 ±\pm 0.0030.605 ±\pm 0.0031.281 ±\pm 0.014
CoE0.598 ±\pm 0.0050.574 ±\pm 0.0021.369 ±\pm 0.038
CoDE (swapped density map)0.564 ±\pm 0.0040.546 ±\pm 0.0041.362 ±\pm 0.013
CoDE (random density map)0.527 ±\pm 0.0060.485 ±\pm 0.0061.570 ±\pm 0.118

A swapped map is better than a random map from an unrelated complex, but worse than both the true map and the coordinate-only variant CoE in both correlation metrics. These results support the importance of density that matches the input complex for binding affinity prediction.

We would also like to note that (a) and (b) are not mutually exclusive; experimental density maps reflect both structural disorder and the ligand's presence and conformation. As a complementary control, the CoBE baseline in Appendix F.1 replaces the density with per-atom B-factor features. CoBE improves over CoE in rank correlation and RMSE, but remains below CoDE, suggesting that density provides useful information beyond the tested B-factor encoding. Together, these results support the predictive value of experimental density, while (a) and (b) cannot be strictly disentangled.

[W2, Q2] Additional distribution and best binder information for Figure 3

In Appendix F.4 (Figure 32), we extend Figure 3 with the interquartile range, the median, the pocket minimum, and the best-10 mean for each test pocket, for both VoxBind and CoDE. For the box convention, we follow the PoseCheck evaluation as in Figure 5. CoDE's interquartile range lies below VoxBind's in many of the pockets. For the best-10 mean, CoDE is lower in 46 of 79 pockets, while the single best binder is comparable between the two methods. Since the CrossDocked2020 benchmark focuses on the distribution of generated molecules rather than a single best sample, we regard the average and median as the primary evidence of CoDE's advantage and interpret the best-10 mean and the best binder as supplementary information.

[Q1] Boltz-like baseline for binding affinity

Nesso-1 is a Boltz-2 derived co-folding affinity model that reports improved affinity prediction over Boltz-2 [1], so the Boltz family is represented in our comparison. Note that we have marked it with †^\dagger, since its affinity training data overlaps with our test split. Adding Boltz-2 itself would introduce the same overlap and yield another row that can only be read as a reference rather than a comparison.

[1] Nesso-1: Accelerating Open-Source Binding Affinity Predictions, bioRxiv 2026

Reviewer_GxRFOpenReview ↗
Rating
4
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper highlights a limitation in existing generative models for molecules -- training only on atomic coordinates, while discarding the uncertainty in atom occupancy. Therefore the paper trains an auto encoder to reconstruct 2mFo-DFc maps. The authors show the benefit of this encoder using two downstream tasks 1) property prediction 2) pocket conditioned ligand generation.

The property prediction task uses the frozen encoder embeddings. For pocket conditioned generation, the encoder's features are projected and added with VoxBind architecture's standard features. On both tasks, CODE model variants show gains over CoE (coordinate only) models and other baselines.

Critical strengths
  • The paper makes a strong case for using electron density as another source of information as opposed to just modeling resolved atomic coordinates. The benefit is clearly shown on two tasks : 1) binding affinity prediction 2) pocket conditioned ligand design

  • For both the tasks, CODE achieves higher performance compared to its COE (identically pre-trained coordinate-only encoder) and baselines considered

  • The paper has done important ablations: such as assessing the generated ligands novelty (since the reference ligand's electron density is given) and assessing the cases in which density mostly helps.

Critical weaknesses
  • It would be beneficial for the audience if several recent work [1,2] are assessed to compare in the pocket conditioned ligand generation tasks

[1] Cremer, Julian, et al. "FLOWR. ROOT–A flow matching-based foundation model for joint multi-purpose structure-aware 3D ligand generation and affinity prediction." Nature Communications 17.1 (2026): 5883.

[2] Dunn, Ian, et al. "OMTRA: A Multi-Task Generative Model for Structure-Based Drug Design." arXiv preprint arXiv:2512.05080 (2025).

Questions for authors

Clarification Requests:

  1. Is there a reason why PLINDER test sets are not used for the generative tasks?

  2. For pocket conditioned ligand generation test set, how would removing just the ligand's electron density affect performance?

  3. Say a task does not have a density map, for such tasks, during inference time, would the CODE model still be able to work reasonably? or would a CoE variant be needed?

  4. Rather than a from scratch training run, If VoxBind architecture is fine-tuned to use electron densities, would such a pertained model still be able to adapt and learn from electron density data?

Suggestions For future work:

  1. For the generative task, does providing the reference ligand's electron density, bias the generative model's ability towards the reference's binding modes / poses?
6 comments

We thank the reviewer for the insightful comments and address each question. We have highlighted additional content and modifications in the revised paper in dark green.

[W1] Additional baseline for ligand generation

We have added FLOWR.ROOT and OMTRA as baselines for the ligand generation task in Appendix F.5, with 100 ligands generated from the official checkpoints for 79 pockets. For FLOWR.ROOT, we used the v2 checkpoint and the reference ligand atom numbers at generation, and it failed to create valid molecules for 5 pockets. For OMTRA, a number of generated poses clash with the pocket, which results in poor Vina Score and Vina Min metrics.

ModelnnVina Score ↓\downarrowVina Min ↓\downarrowVina Dock ↓\downarrowHigh aff. %↑\% \uparrowQED ↑\uparrowSA ↑\uparrowDiv. ↑\uparrow# atoms / mol
VoxBind7887-6.15 / -6.83-7.12 / -7.22-7.97 / -8.0561.7 / 65.70.54 / 0.550.70 / 0.710.76 / 0.7624.0 / 25.0
VoxBind + CoDE (paper)7873-6.58 / -7.20-7.64 / -7.70-8.48 / -8.5168.1 / 78.00.52 / 0.530.68 / 0.680.71 / 0.7024.9 / 25.0
FLOWR.ROOT6968-4.05 / -5.51-5.77 / -5.96-7.25 / -6.8852.6 / 53.00.48 / 0.480.73 / 0.750.65 / 0.6822.5 / 21.0
OMTRA703432.75 / 9.889.56 / -4.02-6.49 / -7.1447.3 / 48.90.50 / 0.510.79 / 0.820.79 / 0.8122.7 / 22.0

(We report the average and median in "Avg. / Med." format)

[Q1] Ligand generation for PLINDER test

We used all of PLINDER as the pre-training dataset regardless of its train/validation/test partition, so PLINDER's test sets cannot serve as a downstream evaluation set. Our focus was instead on removing the downstream test sets from the pre-training data, and CrossDocked2020 provides a generative evaluation set whose test complexes were excluded from pre-training.

[Q2] Removing ligand density for the ligand generation

Removing just the ligand's density at generation degrades the Vina metrics, and the generated molecules become smaller. Generation quality therefore depends on the ligand's density and not only on the pocket density.

ModelnnVina Score ↓\downarrowVina Min ↓\downarrowVina Dock ↓\downarrowHigh aff. %↑\% \uparrowQED ↑\uparrowSA ↑\uparrowDiv. ↑\uparrow# atoms / mol
VoxBind7887-6.15 / -6.83-7.12 / -7.22-7.97 / -8.0561.7 / 65.70.54 / 0.550.70 / 0.710.76 / 0.7624.0 / 25.0
VoxBind + CoDE (ligand density masked)7785-5.30 / -6.66-6.79 / -7.15-7.92 / -7.8866.4 / 71.50.55 / 0.560.69 / 0.690.73 / 0.7322.8 / 23.0
VoxBind + CoDE (paper)7873-6.58 / -7.20-7.64 / -7.70-8.48 / -8.5168.1 / 78.00.52 / 0.530.68 / 0.680.71 / 0.7024.9 / 25.0

(We report the average and median in "Avg. / Med." format)

[Q3] Density absence task cases

Without the complex's own density map, CoDE falls back to roughly the level of CoE, so a CoE variant would be preferable when no map is available. The table below reports CoDE with the density and gradient channels zeroed at inference. A single model that operates with and without density would require techniques such as density dropout during pre-training, which we leave to future work.

Model (inference setting N=733N=733)Pearson r↑r \uparrowSpearman ρ↑\rho \uparrowRMSE ↓\downarrow
CoDE (paper)0.643 ±\pm 0.0020.620 ±\pm 0.0031.342 ±\pm 0.009
CoDE (zeroed density)0.621 ±\pm 0.0040.584 ±\pm 0.0031.476 ±\pm 0.044
CoE0.614 ±\pm 0.0020.574 ±\pm 0.0011.416 ±\pm 0.045

[Q4] Fine-tuning rather than scratch training VoxBind with density

The fine-tuned model does adapt to the density conditioning; under an identical recipe, its Vina docking scores improve over the original VoxBind. It recovers part of the gain, but does not reach the model trained from scratch.

ModelnnVina Score ↓\downarrowVina Min ↓\downarrowVina Dock ↓\downarrowHigh aff. %↑\% \uparrowQED ↑\uparrowSA ↑\uparrowDiv. ↑\uparrow# atoms / mol
VoxBind7887-6.15 / -6.83-7.12 / -7.22-7.97 / -8.0561.7 / 65.70.54 / 0.550.70 / 0.710.76 / 0.7624.0 / 25.0
VoxBind + CoDE (fine-tuned)7774-5.51 / -7.02-7.09 / -7.55-8.23 / -8.3267.5 / 74.00.52 / 0.540.68 / 0.680.71 / 0.7024.9 / 25.0
VoxBind + CoDE (paper)7873-6.58 / -7.20-7.64 / -7.70-8.48 / -8.5168.1 / 78.00.52 / 0.530.68 / 0.680.71 / 0.7024.9 / 25.0

(We report the average and median in "Avg. / Med." format)

[S1] Whether reference ligand density bias exists for generated ligand binding modes or poses

CoDE's generated ligand poses resemble the reference ligand's interaction pattern to a similar degree as VoxBind, while the coordinate-only CoE resembles it far more. DecompDiff, which derives its priors from the reference ligand, shows higher similarity than VoxBind and CoDE, but remains below CoE. We compute the residue-level ProLIF fingerprints [1] over all generated ligands, and compare them by (i) the Tanimoto similarity, (ii) the fractions of the reference's total interactions and H-bond interactions recovered, and (iii) the fraction of distinct fingerprints among the poses. For more details, refer to Appendix F.7.

ModelTanimoto similarityInteraction recovery rateH-bond interaction recovery rateDistinct fingerprint fraction
DecompDiff0.5580.7150.4230.96
VoxBind0.4130.5840.2250.99
CoE0.7220.8180.7690.60
CoDE0.4370.6310.2530.99

[1] ProLIF: a library to encode molecular interactions as fingerprints, Journal of Cheminformatics 2021

Reviewer_GxRFOpenReview ↗

Follow up regarding tanimoto sim of CoE

Can you clarify what could be the reason CoE has a very high tanimoto similarity compared to VoxBind? If I understand right, conceptually the features in the CoE ablation is very similar to the VoxBind experiment right?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.