Closed since about 3 years ago
Intermediate Overall DesignDesign a binding pocket for the FMN ligand! The Neural Net Objective awards a 2000 pt. bonus for AlphaFold predictions with both high confidence and high similarity to the current solution. After an AlphaFold prediction is complete, load the original or predicted structure from the AlphaFold panel to receive the bonus. You can continue to work on this solution, but the bonus will disappear if the similarity drops. Note that the table in the AlphaFold panel only displays the similarity of the original solution; the true similarity of your current solution is recalculated by the Neural Net Objective as you change its structure.
Once you've designed a binder for the target ligand, upload your solution for AlphaFold using the AlphaFold prediction tool. AlphaFold will predict the structure of your binder only (i.e. in the absence of the ligand). If you load this prediction, then Foldit will attempt to align the prediction with your solution. We are looking for designs where the raw AlphaFold prediction has a binding pocket that fits the ligand!
Flavin mononucleotide (FMN) is a naturally occurring molecule found across all kingdoms of life. It has a highly conjugated ring structure that readily accepts an extra electron, so FMN is often used by oxidizing/reducing enzymes to transfer electrons in the cell. The ring structure also makes FMN sensitive to visible light, and it is found in some light sensing proteins as well. We'd like to design a protein that can bind FMN and transfer electrons by absorbing light. The conjugated rings offer a large hydrophobic surface for tight binding, but FMN also has a lot of oxygen atoms that need to make H-bonds!
This puzzle starts with the backbone from LOV2, a natural FMN binder, with the ligand placed roughly in the correct position in the binding pocket. However, players may completely redesign the protein or reposition the ligand within the binding pocket. Try to design a binding pocket with high Contact Surface and zero BUNS! See the puzzle comments for Objective details.
Contact Surface Metric (max +2000)
Measures how much of your binder surface is in close contact with the target protein. The goal Contact Surface is 200 or more.
Buried Unsats (max +1000)
Penalizes 200 points for each polar atom that cannot make any hydrogen bonds.
Neural Net (max +2000)
Achieve an AlphaFold prediction confidence of 80% or more, and similarity to current solution is 80% or more.
You write that the AF-sim is recalculated in the background by the NN-algo. Would it be possible to get access to this data? Because for me the only reliable and performant way to improve AF-sim is to reload the prediction and that often screws up the sidechain- and backbone-positions and it for sure kills manually designed H-bonds. So I only do this only in the early phase of a design. With online-info from NN on AF-sim it should be possible to improve AF-sim late in the game by doing manual changes and monitor the NN-results for AF-sim. This should be far better than having to upload to AF and wait a few minutes for the results. This could be very helpful to improve sim for basically almost finished designs. I would appreciate ths.
@ichwilldiesennamen AlphaFold similarity cannot be calculated unless you have already gotten an AlphaFold prediction. Similarity is a measure of difference between your current solution and the AlphaFold prediction.
After you have gotten an AlphaFold prediction, then the Neural Net Objective can calculate similarity in the background (because the AlphaFold prediction is in memory on your local machine). But as soon as you change the sequence of your solution, the AlphaFold prediction is no longer valid so there is no way to calculate similarity. It is necessary to submit your new sequence to AlphaFold to get a new prediction.
@bkoep Thanks for explaining this. I didn't realize that the AF-sim is simply determined by a difference-check from the AF-prediction to the current solution. But of course it makes sense. But since any mutation in the sequence will basically invalidate the AF-result, I understand that the AF-sim can only be determined again if the current solution is uploaded again to AF. What I would like would be a tool to calculate AF-results in near-realtime but this will probably not be possible in the near future due to the high processing power needed. So you can forget my request.
