Escrever um comentário

The article by Liu et al introduces a deep learning approach for predicting protein-ligand affinity, which is an important field of study for both experimental and computational drug discovery. The authors introduce a CNN model, DeepDOX1, which takes a variety of physical-informed inputs such as RdKit’s ligand fingerprints and features of residues within the active site. The novelty, as highlighted by the authors, lies in the incorporation of quantum mechanical quantities on ligand entropy, pocket binding energy and interaction descriptors. These features are often computationally expensive and difficult to scale across diverse systems. The authors bypass some of these calculations by integrating predictions made by other previous models. The resulting DeepDOX1 therefore incorporates many descriptors from different levels of theory and AI.

We have the following major comments:

1. The authors benchmarked on 5 models built on different architectures and schemes to their method. It is not immediately clear to me why these models are selected as some of them are not considered recent SOTA models (eg d_{vina}XGB from 2019 and DeepDock from 2019). The authors should either provide their motivation on why these models are selected as their benchmarks in the manuscript or benchmark on more recent models (diffusion/flow-based methods such as Boltz-2 which are also mentioned in the introduction) given how fast this field moves.

2. The definition of the AI and QM module is unclear to me. These modules are only termed to in the ablation study and are not identified in the schematic in Figure 1 itself or its caption. Are the dotted lines in Figure 1b corresponds to different modules by colors? Does the removal of a QM module mean removing all QM inputs or just the dense layers for the QM residue-ligand interaction?

3. Following up to question 2, it could be useful to report the number of remaining trainable parameters in these ablation studies versus the total number of parameters in the entire model - which may affect the results when the model becomes too small.

4. This model is built upon QM predictions made by the DOX_DBW package; it would be helpful to include the original DOX_DBW predictions for benchmarking (or is that eventually the same idea as of the removal of the AI module ablation case?).

5. How does the definition of binding pocket affect the predicting ability - the DFT region is defined 3 Angstrom from the small molecules and the pocket residues are drawn from a radius cutoff of 3.5Å. It would be helpful to clarify why different metrics are used for characterizing the binding site.

6. I find the statement made by the authors important - “We also found that the predicted binding affinity exhibits significant fluctuations(~1pKi unit) depending on the distinct binding poses, highlights the importance of binding structure and necessity to discuss both AI generated binding structure and AI employed binding structure.”. The co-folded structure is an important input to the model for predicting the corresponding binding affinity. It would be interesting to see how sensitive the model is among predictions made by more recent structure prediction models (such as AF3 and its replicates/variants), or docking methods (diffdock[1] for example), or posing models (PLACER[2] for example).

7. Some details on training the model in the SI would be helpful. What optimizer was chosen? What learning rate was selected? What loss function was used to optimize the model?

8. In the evaluation table (such as 3,4), it would be preferable to highlight the best coefficient (for example in bold).

9. For the case study as in Fig. 5, it would be much easier for the readers to compare the predicting power of the model if the authors can report with the same unit - currently predictions are in pk_i’s and experimental measures are in K_i’s.

Refs.

1. DiffDock: Diffusion Steps, Twists, and Turns for Molecular Docking. https://arxiv.org/abs/2210.01776

2. Modeling protein-small molecule conformational ensembles with PLACER. https://www.biorxiv.org/content/10.1101/2024.09.25.614868v2

Competing interests

The authors declare that they have no competing interests.

Use of Artificial Intelligence (AI)

The authors declare that they did not use generative AI to come up with new ideas for their review.