peastman
commented
1 year ago That's not bad then. We're probably looking at under 100,000 molecules. I suggest we use a sequence of two scripts. The first one will generate the states, apply the free energy filter, and write out SMILES strings to a file. Once we see how many molecules we're dealing with, we can decide how many conformations to include for each one. A second script will read in the first file, generate the conformations, and create the HDF5 file.
I can create a draft PR with an outline of the first script, if that sounds reasonable.
jchodera
jthorton
giadefa
pavankum
I've been talking with lots of people about what would be most useful to add in version 2 of SPICE. I've gotten lots of great suggestions. Here are some ideas in no particular order.
Solvated PubChem Molecules. This would provide data on ligand-water interactions. They could be generated the same way we did the solvated amino acids: simulate the molecule in a box of water, periodically saving conformations along with ~20 of the nearest water molecules. These will be largish, and therefore expensive to compute. We'll probably need to limit the number of molecules we include and the number of conformations for each one. I think this is a case where a little data can go a long way, since each conformation has many water molecules.
Water Clusters. This would provide some conformations that are more representative of bulk water. How large should the clusters be? What's the best way to generate them? I've been advised that classical water models often don't produce very realistic water structure.
PubChem Molecules with Boron (and possibly Silicon). I excluded these elements from the original PubChem set. They only appear in a tiny fraction of molecules, so there wouldn't have been enough data on them to be useful. But if we specifically select molecules that include them, I think we could assemble a useful amount of data. There's around 1500 molecules with B and around 1900 with Si.
Amino Acid / Small Molecule Pairs. This would provide more data on protein-ligand interactions. One way of generating them would be to use PDB structures. We could identify which amino acids contact the ligand and generate conformations for them, each conformation including the ligand plus a single amino acid.
Transition States. If you're interested in rates, it's important to model transition states accurately. Because they're higher energy, they won't be very well sampled in a thermal distribution. Can we add conformations specifically to sample transition states? One suggestion was to scan dihedral rotations while minimizing all other degrees of freedom. What molecules would it be useful to do this with? Of course, not all transition states involve dihedral rotations. What else could we do?
Metals. The goal would be to allow simulating metalloproteins. These could also be generated from PDB structures, taking the amino acids that interact with the metal. Would it be sufficient to consider only one amino acid at a time, or do you really need to include the whole environment? If the latter, these could be quite expensive to calculate.
Active Learning. This is the idea discussed in #35, to add more conformations for existing molecules to improve the accuracy of trained models.
More PubChem Molecules. More chemical diversity is always good. The number included in version 1 was mostly arbitrary, just based on how long it took to do the calculations. It would be easy to add more.