Hey everyone, this week was pretty fast-paced with regards to experimentation. I began by getting acquainted to the DiffDock molecular docking software, and then conducted some fun simulations. This isn’t my first time doing a project using DiffDock, because I have previously researched similar models to understand how we can optimize them and use them on most computers (potato PCs versus large H200 GPU clusters). I did, however learn some new steps, so before we carry on, here is the google colab notebook that I used to perform some elementary simulations: https://colab.research.google.com/drive/1CTtUGg05-2MtlWmfJhqzLTtkDDaxCDOQ
Phase 1: The parameters.
DiffDock is a model and has many parameters. To draw an analogy, DiffDock is like a radio with many switches and knobs that can be modified to play different songs. In the context of DiffDock, these knobs are parameters which can change how accurate/efficient the model can be at predicting protein-ligand interactions. While there are technically 16 parameters, many of these aren’t applicable to the interface (which I will be using). There are ultimately three parameters (Number of samples per complex, number of inference steps, and batch steps). The first two deal with how the model predicts the protein-ligand interaction poses. The final one is helpful for speeding up the run-time, with a marginal decrease in accuracy.
I had previously determined that using 15 inference steps, 30 samples per complex, and 4 batch steps increased the accuracy and efficiency of DiffDock. I will be using the same parameters for my simulations.
Phase 2: PubChem.
Over the past few weeks I introduced the different types of Rhodopsin states (dark-state, meta-II, etc.), and I also mentioned that there is a ‘retinal ligand’ that binds to Rhodopsin and absorbs photons to trigger conformational changes. Using DiffDock, I will be taking each of the Rhodopsin states (dark-state, meta-II, photo-intermediate, etc.) and binding them each to the retinal ligand using DiffDock. As mentioned previously, retinal begins in a dark-state too (11-cis-retinal) and absorbs photons to turn into an active retinal ligand (all-trans-retinal), which is responsible for turning Rhodopsin from dark into meta-II states. I will be using DiffDock to see how the binding affinity (or how well the interaction is) changes with respect to the state.
PubChem, a database, contains many PubChem IDs for small molecules and ligands. Inside Pubchem, there are two IDs (5280490 and 638015) which represent these two states of Retinal that we can use for docking.
Phase 3: Some elementary experimentation!
I conducted some simulations to get acquainted with how DiffDock predicts Rhodopsin-retinal poses, and began learning where retinal docks within Rhodopsin. I will be sharing some figures and analysis next week, so stay tuned!
All viewpoints are welcome but profane, threatening, disrespectful, or harassing comments will not be tolerated and are subject to moderation up to, and including, full deletion.
mikyle_h
Very fascinating work, Aditya! How do you think these parameters would change for more complex protein-ligand interactions beyond Rhodopsin, if at all? Would the same settings be effective for other GPCRs or proteins with more intricate binding sites, or do you think a different approach would be needed?
Comments:
All viewpoints are welcome but profane, threatening, disrespectful, or harassing comments will not be tolerated and are subject to moderation up to, and including, full deletion.