Write a comment

Machine Learning-Enhanced Drug Discovery for BACE1: A Novel Approach to Alzheimer's Therapeutics

This paper examines possible inhibitors for BACE1, a critical mediator in the amyloidogenic pathway responsible for generating neurotoxic amyloid-beta (Aβ) peptides. It uses DrugGPT to identify potential inhibitors, filter inhibitor candidates, docking inhibitors, and perform MD simulation. They aimed to characterize one new potential inhibitor against three existing BACE1 inhibitors. They claim that MLC10 promotes controlled flexibility within the enzyme that is unique from other inhibitors. While they show that MLC10 does change the flexibility profile of BACE1, the argument that this changes the ability of substrates to bind is misguided. As I interpret it, MLC10 binds in the active site, which would prevent a substrate from binding due to direct competition. Changing the flexibility of BACE1 would not impact the ability of a substrate to bind or for a product to be released, as the inhibitor and MLC10 cannot be bound simultaneously (based on Figure 4). Further characterization of what areas of BACE1 are becoming more/less flexible, focusing on how this changes binding affinity or inhibition measurements, should be made to support the arguments in this paper. 

Major:

1) Given that ligands represented by SMILES can be generated into multiple conformations, yet DrugGPT only chooses one conformation, how does this potentially impact the output of ligands?

2) Please provide more details on the claim ‘generated ligands with varying levels of complexity’? Was this done using Tanimoto, 3D, or other characteristics?

3) While MCL10 had a better binding affinity, given that MCL11 was still very similar, what was the rationale for moving forward with just one of these compounds?

4) To help contextualize the statement ‘MLC10’s dynamic behaviour could enable induced-fit optimization, allowing the active site geometry to be fine-tuned, thereby maximizing complementarity between the inhibitor and BACE1’s active site’, please provide some shape complementarity metric of the four inhibitors you examined, such as a change in pocket volume size, number of hydrogen bonding, or van der Waals interactions. While this is listed in a table, having an interaction matrix or other graphic form would make it easier to understand.

5) As all of these inhibitors bind to the orthosteric catalytic site, I need clarification on how they would prevent catalysis beyond sterically blocking the natural substrate from binding. Please clarify the statements throughout the paper referring to changes in catalysis. 

6) The changes in RMSF and RMSD across all five structures look incredibly similar, making it difficult to qualify the claims of differences in fluctuations between inhibitors in different regions. Can you quantify and visually show the differences you are pointing out to back up these claims?

7) Please increase the resolution and font size of Figure 7. Currently, I cannot interpret it.

8) In Figure 9, please describe what the deformity is compared to. Is this compared to apo BACE1?

9) Please reword the claims regarding the inhibitor effect of MLC10. As you are only measuring computational estimates of binding affinity and not competition, you can make conclusions on the inhibitory effects of substrate binding.

10) Overall, as the three existing inhibitors and MLC10 have very similar binding affinities, it is unclear how the differences in flexibility impact binding affinities. Please use the discussion to outline this argument and what potential forward tests may be. 

Minor:

1) It would be useful to provide a flowchart and numbers through the manuscript of the number of compounds considered and removed at each step (i.e. number of compounds generated from DrugGPT, number of compounds removed using SMIANA, ect). 

2) Please describe what you mean by an in-depth analysis of the SMILES notations was performed and what tools were used. 

3) Please increase the resolution of Figure 4. 

4) In Figure 4, it would be helpful to orient the figure so we can see where the four compounds bind in relation to each other. Orienting the protein the same way and/or overlaying transparent spheres of the ligand onto the same structure would be helpful.  

Competing interests

The author declares that they have no competing interests.