MF-LAL: Drug Compound Generation Using Multi-Fidelity Latent Space Active Learning

We thank the reviewer for their helpful feedback and positive comments about the work.

Q1: Because the authors opt to use a setting close to real-life, with MD as the highest-fidelity oracle, they are limited to sampling very few samples in the end (15 per method). This is somehow understandable given high computational cost, but limits the reliability of the results.

A1: We would like to emphasize the statistical tests in Section 4.3 and Appendix B.3, where we show that the ABFE scores of MF-LAL compounds are significantly better than those from baselines in terms of both mean and top scores for both targets. So while we do have a relatively small number of samples, we show that MF-LAL makes statistically significant and reliable improvements over baselines.

Q2: I wonder if it would be possible to instead do a simulated multi-fidelity setup, as in some earlier papers; or perhaps to pick a lower-cost highest-fidelity oracle.

A2: While many previous works have indeed focused on a simulated multi-fidelity setup, we do not think that such a setup is a very good proxy for real-world simulators. This is because the method of ABFE calculation is fundamentally different from docking or other activity prediction techniques, and so it is difficult to substitute it with other cheaper techniques and still reach useful conclusions. If it is possible to reach statistically significant conclusions using the real-world MD simulator, which we believe to have done in this paper, then we think it is most valuable to use the real simulator instead of an artificial setup.

Q3: ...for the scaffolds, the authors perform statistical analysis only for two best methods (the information about which is hidden away in the appendix), which might be misleading. I would strongly suggest limiting the results in this table to the same budget of oracle calls (15 each), and perhaps having a separate table / figure for 40 compound setting.

A3: We agree that making two separate tables for 15 and 40 compounds would increase the clarity of the results. In the updated draft, we will include these two tables, as well as statistical tests comparing MF-LAL to all applicable baselines in both settings. As noted in Appendix B.3, statistical tests comparing MF-LAL to all baselines where all methods were limited to 15 compounds still showed statistically significant results in favor of MF-LAL (except for the binomial test for the number of active scaffolds between MF-LAL, which generated 4 scaffolds, and REINVENT, MF-AL-PPO, and DecompDiff which generated 1 scaffold, where p=0.07). The results for MF-LAL and the top baseline for each target using only 15 compounds are shown below, because they were not included in the original draft. These results remain consistent with our conclusions about the strong performance of MF-LAL.

Q4: I’d be curious to see some citation backing up this claim [about query synthesis] (that’s also more recent).

A4: [1, 2, 3, 4] are recent papers that explore query synthesis approaches, and many of them find that query synthesis outperforms traditional pool-based methods in active learning. We will include these citations in the updated draft.

[1] Morand et al. “Efficient Exploration of Microstructure-Property Spaces via Active Learning.” Frontiers in Materials 2022.

[2] Schumann and Rehbein. “Active Learning via Membership Query Synthesis for Semi-Supervised Sentence Classification.” CoNLL 2019.

[3] Guo et al. “Dual generative adversarial active learning.” Applied Intelligence 2021.

[4] Zhu and Bento. “Generative Adversarial Active Learning.” arXiv 2017.

Q5: “Ensembled AutoDock4 (cost: 44s, ROC-AUC BRD4(2): 0.80, c-MET: 0.80)” - one cost is provided for both targets, while c-MET seems to use 5 crystal structures, whereas the other uses 8; could authors explicitly confirm that the cost is supposed to be the same in both cases?

A5: Thank you for pointing out this omission, the shown cost is indeed only for BRD4(2). The correct cost for c-MET is 68s, which we will include in the updated draft.

Q6: “Additionally, the pairwise Tanimoto similarity among the 40 compounds generated by MF-LAL is less than 0.2” - does this refer to mean pairwise similarity?

A6: Yes, mean pairwise similarity. We will update the draft to clarify.

We thank the reviewer for their helpful feedback and positive comments about the work.

Q1: In table 1, POCKET2MOL and MF-LAL generated 40 molecules while other baseline methods generated 15 molecules. It seems that this is unfair as generating more molecules would definitely result in better ABFE scores for the top-3 molecules and more active scaffolds. But the increase in mean value is solid, so I think the author should show the separated results mean value for 40 molecules ; number of active scaffolds and top-3 values for 15 molecules

when comparing the top-3 values, it should also generating 15 molecules as for other methods for fair comparasion

A1: Based on this and other comments from reviewers, we will make two separate tables in the updated draft reporting results on both 15 (for all methods) and 40 compounds (for the methods that have them). See the response to Reviewer QauA for the MF-LAL and baseline results on only 15 compounds, including the top compound scores and the number of active scaffolds, which remain consistent with our conclusions about the strong performance of MF-LAL.

Q2: Another concern is that authors use Pocket2Mol and DecompDiff as baseline model. But they are 3D pocket-based molecule generation models instead of optimization models. There are some 3D optimization models, like, DecompOpt[1] and TagMol[2]. Also RGA[3] can be included

A2: Thank you for the references, we will cite them in the updated draft. We agree that comparing MF-LAL to a 3D optimization model would be valuable, so we are currently running TAGMol as a baseline. While the results are not yet finished due to computational cost, we will post them when they are finished during the author-reviewer discussion period. Following the reviewer’s suggestion, we have also run RGA. We ran RGA using a single fidelity (docking) as the oracle, similar to how we implemented the SF-VAE (only docking) and REINVENT (only docking) baselines in our paper. We chose to use docking, instead of ABFE, as the single fidelity oracle due to computational cost and the presumed inability of a genetic algorithm to make use of a very small number of ABFE oracle calls. We used the default parameters of RGA, with random ZINC250k compounds as the starting population for the genetic algorithm. The results from RGA are as follows:

The results from MF-LAL are also shown for comparison. MF-LAL significantly outperforms RGA, with the latter having performance similar to that of the REINVENT baseline. We will include these results in the updated draft.

Finally, we do not think DecompOpt is a good baseline to compare with because it requires knowledge of existing binders for generation (to compute the “reference arms”). Since MF-LAL and our other baselines are de novo generation methods that do not use knowledge from existing binders, we do not think a comparison with DecompOpt would be fair.

Q3: tested targets are limited

A3: The number of targets we tested is limited by the high cost of ABFE calculations. In addition, the ABFE framework we use is only validated and configured for a few targets, and BRD4(2) and c-MET are the only of those targets that are of interest from a biological/drug discovery perspective.

Q4: From the ablation it seems that even the Linear regression fidelity contributes to the performance. Removing it would results in significant decrease of number of active scaffolds from 4 to 0. The ABFE scores for top 3 molecule also significantly decreased. Can you explain more on why this happens? My understanding is that it should have some kind of influence but should not be that obvious, as ABFE is much accurate than a linear regression model.

A4: The lowest fidelity oracle is critical for performance because it provides the majority of the data for the multi-fidelity model. As stated in Appendix B, we provide the model with an initial dataset of 200,000 ZINC250k compounds and associated oracle outputs at the first fidelity level (linear regression). Thus, removing this fidelity greatly reduces the total data available to the model. Even if this data is significantly less accurate than ABFE, the quantity of data is still important for model performance, and so it is expected that removing this data significantly reduces performance.

We thank the reviewer for their helpful feedback and positive comments about the work.

Q1: There are no details on how GCN was implemented for generating molecules.

A1: We will include details of our GCN implementation in the Appendix in the updated draft. Briefly, we used a three-layer graph convolutional network with one-hot encoded atom types for the encoder, and an inner product decoder as described in Kipf and Welling 2016.

Q2: The SELFIES representation is used to ensure that the string representation can be decoded to molecules, but rather heavy filtering criteria are applied (QED > 0.4, SA < 4, no rings with at least seven atoms) - what percentage of the generated compounds match these criteria for the tested models?

A2: Among all compounds generated after training with active learning, 55% for BRD4(2) and 68% for c-MET fulfilled the filtering criteria. We find these percentages high enough such that there is no need to do multi-objective optimization, especially because generation is very fast so discarding 25-50% of the generated molecules is not problematic. We will include these numbers in the updated draft.

Q3: It is unclear whether the number of active scaffolds and the results of the top compounds are computed from all 40 compounds or just 15.

A3: The number of active scaffolds and the top 3 compounds in Table 1 are computed from all 40 compounds (for MF-LAL and the top baselines). Based on this and other comments from reviewers, we will make two separate tables in the updated draft reporting results on both 15 and 40 compounds for clarity. See the response to Reviewer QauA for the MF-LAL results on only 15 compounds, including the top compound scores and the number of active scaffolds.

Q4: the initially generated compounds prior to filtering could be assessed in terms of validity, synthetic accessibility, and drug-likeness.

A4: Prior to filtering, the mean SAscore of generated compounds is 3.9 for BRD4(2) and 3.6 for c-MET. The mean QED is for 0.48 BRD4(2) and 0.50 for c-MET. The validity is 100% because we used SELFIES strings, which are guaranteed to be valid. We will include these numbers in the updated draft.

Q5: How were thresholds -8.2 and -6.8 kcal/mol chosen for the two tested targets? Was it based on reference compounds or balance of the activity classes in training regression model for the lowest fidelity?

A5: The thresholds were chosen based on reference compounds from previous works that investigate BRD4(2) and c-MET (see lines 368-373, left side). We roughly picked these thresholds based on the typical experimental affinities of the best binders analyzed in these works. For BRD4(2), Liu et al. 2017 (see paper for citation) explore various BRD4(2) inhibitors and generally consider compounds active when they have submicromolar (<1 $\mu$M) activity, which is the cutoff we used in our paper. For c-MET, Naguib et al. 2024 states that “potent activity” is achieved with a 12 $\mu$M inhibitor, so we set our activity cutoff to <10 $\mu$M. We will include these details in the updated draft.

Q6: The choice of the decoder architecture seems nonstandard. Text representations such as SMILES and SELFIES are usually generated in the autoregressive fashion. Have you tried using RNNs? Was the transformer trained also to predict all characters at the same time, or autoregressively?

A6: Both Transformers and RNNs have indeed been used in such applications. We chose to only test the Transformer architecture, however, because it has demonstrated very similar or slightly superior performance relative to RNNs on molecular generation tasks [1, 2]. Our Transformer was trained to predict characters autoregressively using the standard Transformer decoder architecture.

[1] Chen et al.”Molecular language models: RNNs or transformer?” Briefings in Functional Genomics 2023.

[2] Xu et al. “REINVENT-Transformer: Molecular De Novo Design through Transformer-based Reinforcement Learning.” arXiv 2024.

We thank the reviewer for their helpful feedback and positive comments about the work.

Q1: The training details including the supplement are sparse and disjointed. It is not easy to understand how exactly the model is trained. I know space is limited, but even an algorithm in the supplement in addition to the loss function would help

A1: Thank you for the suggestion, we will consolidate all training details into a single section and add a training algorithm to the Appendix in the updated draft.

Q2: One concern would be that such an optimization causes adverse effects on the fidelity not being trained.

A2: The loss function we use while training considers molecules from all fidelity levels 1...K, regardless of the current level k. In other words, while k dictates which fidelity data we will add to the dataset, the entire dataset is used to train the model regardless of k. This ensures that training at level k does not degrade performance at the other fidelity levels. Analyzing the reconstruction accuracy, and the accuracy of the surrogate models, confirms that they retain their performance when other fidelities are being trained.

Q3: The usual generation quality measures such as validity, diversity, and synthesizability are not presented

A3: Among the generated compounds for both targets following filtering, the diversity (1 - mean pairwise Tanimoto similarity) is 0.81 for BRD4(2) and 0.83 for c-MET. The synthesizability (mean SAscore) is 3.6 for BRD4(2) and 3.5 for c-MET. The drug-likeness (mean QED) is 0.59 for BRD4(2) and 0.63 for c-MET. The validity is 100% for both targets, because we use SELFIES strings that guarantee validity. These metrics are in the range of typical drug compounds, so we consider them satisfactory. We will include these numbers in the updated draft.

Q4: Since all the baselines don’t have statistically significant results, it is a bit concerning to compare the results

A4: As reported in Section 4.3, compounds generated by MF-LAL had better ABFE scores than all baseline methods at a statistically significant level (see Appendix B.3 for statistical details). Specifically, the difference between both the mean ABFE scores of all generated compounds, as well as the scores of the top 3 compounds, is significant between MF-LAL and each baseline for both targets. For the “# of active scaffolds” test, MF-LAL also produced significantly more active scaffolds than other baselines. Since MF-LAL shows statistically significant improvements over baseline methods, we do not think these comparisons are concerning.

Q5: Substructure similarity is measured with Tanimoto distance of Morgan fingerprints which does not capture 3D information.

A5: If the reviewer is referring to the similarity metric we used to compute the number of active scaffolds, we do not think that capturing 3D information is critical in this case. Scaffold similarity, which is what we measure in our paper, is the most commonly used diversity metric by medicinal chemists and strongly relates to the overall structural diversity of a set of compounds [1]. Additionally, measuring the 3D shape diversity is a difficult task, and is not commonly done by practitioners [1].

[1] Galloway et al. “Diversity-oriented synthesis as a tool for the discovery of novel biologically active small molecules.” Nature Communications 2010.