Software Evaluation

Links

• Progress Tracker: ./docs/sources/CHANGELOG.md
• Error Log: ./docs/sources/ERRORLOG.md
• Ligand Selection ./docs/sources/LIGANDSELECTION.md
• Docking Instructions ./docs/sources/DOCKINGINSTRUCTIONS.md

Section 1: Virtual Docking

Background: To improve the accuracy and consistency of current virtual docking algorithms, we are developing DeepDockR, a machine learning powered docking platform. Achieving that requires us to dock thousands of molecules into a target of interest and store all associated scores and metadata in our database. While we have automated much of the docking pipeline, a very important step is the choice and preparation of the structures to use for docking.

Goal: Evaluate the performance of a docking software in estimating the binding affinity of compounds to Carbonic Anhydrase I, by comparing the virtual docking results to available bioactivity data. The evaluation should comprise of regression and as well as ranking metrics.

Protein Evaluation

To tackle this goal we need to evaluate the protein in question. Protein's can exist in several different states but for evaluation of a docking software the user should pick a protein that is already in a docked state. Here we can determine what what exactly the pocket size looks like and how it interacts with a possible therapuetic inside of the protein.

The problem asks for evaluation of docking for CA1, to initialize we are going to big an undocked ligand into the protein which stems from the x-ray structure 2CAB. We are luxurious enough to have a well defined active site that has the correct protein conformation that will be more relevant to the binding pose for the potential ligands binding to the complex. 2cab.pdb which has the zinc ion in the pocket of the protein in the protein docked state which can be the most unbiased x-ray structure and prove as the docked protein template.

If we did not have knowledge of the active site we will use a software such as AutoLigand that will help us identify where the active site could potentially be.

Ligand Selection

For our ligand selection we need enough chemical space to explore known inhitors, negative inhibitors, and possible positive inhibitors that match functional groups revelant to the known inhibitors. The ligands that I have chosen for this project are listed in Ligand Selection

Reason behind the ligand choices are what is most avaliable to me! My choices of ligands also stem from the ZINC Database because I ran into complications with the CHEMBL database. Follow the error log to get more insight.

Known inhibitors: are great so we can evaluate what the Kd potentials will be from the Autodock and already pre-screened values that are available online. Also if we super impose images of docked ligands with docked proteins we can evaluate x-ray or nmr determined structures with the virtual docking software to evaluate correct poses.

Negative inhibitors: were determined through manual inspection, after determining the pocket size of the active-site of the protein I could infer what wouldn't fit so I selected large bulky phenyl compounds that also lacked in relative functional groups to reach the zinc site.

Positive inhibitors: were selected from a blend of what was available from the ZINC Database and similarity functional groups as th known inhibitors.

Grid Selection

On preparation of the protein-ligands for autodocking, we need to prep the protein using the AutoDock tools. We can use Autodock tools to select hydrogen atoms, solvent interactions, merging non-polar hydrogen atoms onto their respective heavy atoms, and assigning atom types.

Protein charges are implemented using force fields. For this experiment we will be using AMBER (AM1-BCC) that semi-emperical quantum mechanics to determine correct states. It's more accurate but slower than other force field rendering.

Using Chimera we will also be able to detect the sum of atomic charges lead to to an integer matching the summary charges.

Docking Site Preparation, I set up a 3D grid of where the active site of the protein complex is. Because we are fortunate enough to have this active site known we can narrow down our search for potential inhibitors. Usually the active site would be determined through a 2D NMR approach or X-ray to evaluate where ligands are binding in corresponding to amino acids.

Here you can see the grid formation of the ligand.

For docking instructions please follow the

Results

This was a challenge for analysis of the docking results. We are limited by both the software and the results of the complexes and also how the software calculated it's scoring function.

AutoDock Vina seems to be running it's own scoring function that doesn't directly correlate to a Kd value. The Kd will be widely used as the primary docking measurement to detect the binding efficiency of the ligand.

To calculate the Kd Value we have a couple of parameters that are not in our control

• Concentration of the Ligand | Autodock Vina assumes it is at 1nM
• PH Level | AutoDock Vina assumes it is at 7~8
• Temperature | AutoDock Vina assumes it is at 25 C or 298K

We are also limited by the data available:

• online pdb banks
• publication restrictions
• ligands that haven't been tested before.

Results of the file are labeled here in this Excel Doc with only Kd Values caluclated and one correlation based on the ZINC ID of the compound.

We can test the docking performance of Autodock Vina on 3 main categories.

• Predicted Free Energy of the Compound
• Clustered Docking Conformations based on the RMSD Coordindates
• Visual Inspection of the Docking Poses

I wasn't able to find the protein in the undocked state as free, the best I was able to receive was a Zinc Ion Core in a salt solution.

Here you can say the overlay which matches 100% Identity to most structures:

I also ran into a lot of problems in regards to Chimera, MatchMaker Suite, and Autodock Vina. I was unable to write the PDB files from the software and do a direct comparison to the X-Ray Structure because of this conundrum.

To achieve the results I capture the coordinates of the atoms did a direct coorelation to the crystal structure to achieve the RMSD calculations. Although if I had to work alignment into the equation the error margins would be massive.

Here is an example of those grueling input of xyz coordinates....

Opinion

Based on current efforts and results from this experiment, it proved to be not entirely inconclusive. Running experiments on negative controls helped the correlation to discover that AutoDock Vina showed what can be potentially bound and what cannot be. With the positive random inhibitors based on the functional groups of the known inhibitors we can predict to see what are lead potential therapeutics based solely on the AutoDock Vina score and free energy calculations.

Autodock Vina also favors strong hydrogen bond donors binding into the Zinc Core. Which is skewing the posing results and eliminating posing factors.

Autodock Vina also doesn't handle match making as well as it should, we run into a lot of issues regarding the file formats that are produced to do manual coordination determination. I would rather use MOE or schrodinger something a little high calibur to analze docking results.

Although we have some limiting factors

• The lack of data that is publicaly available or the lack of data in general between ligands and proteins. Because of how expensive it can be to run NMR, ITC, and X-Ray Diffractions on protein-ligand complexes we lack a lot of chemical data which would prove useful in the effectiveness of the software
• The force fields used in the docking are only used to calculate using a single conformation of the complex with fixed stable parameters.

Because of the complications with the docking software, at most, in it's current form we can use to identify potential binding functional groups and potential poses. It could serve as a beginning start to early stage drug-discovery but not capable enough of pushing out a drug. A major advancement in scoring functions will take in more parameters such as liver clearance, higher force field calculations with protein-ligand flexibilty, and cross chain pocket evaluation. As of right now we do need more manual inspection to drive AutoDock Vina futher.