I am training machine learning models to predict the binding affinity of small molecules against a protein receptor. During the training set preparation I remove very small and very big moelcules, as they are more likely to binding non-specifically to the receptor and abolish its funtion. Currently, I use z-score thresholds (-2.5 for the very small and 2.5 for the very big molecules). However, the z-score depends each time on the distribution of MW in the available data and hence I believe that actual MW lower and upper thresholds would be more accurate. According to your experience, what should be these MW thresholds?
PS: please don't point me to Lipinski's rule of 5 (180 to 500) or some related rule. When training a machine learning model on known data you have to be more "generous" otherwise you will discard lots of precious experimental data.
If you don't like lipinski's rules, or the various derivatives thereof, you're sort of out of luck. The molecular weight range that's likely to be viable is going to vary by target: ATP competitive inhibitors will, by and large, be smaller than Protein-protein-interface inhibitors. And if you're looking at inhibitors of, say, PD-1, you're talking antibodies, and should multiply your molecular weight by a few thousand. If you're looking at metallo-drugs, a different choice of metal could change the weight by 10-100, while doing little to the size and properties.
Personally, I wouldn't discard data at all unless I KNEW it was an off-target interaction. As you say: you risk your model missing some key information that way. But if you want a narrower model? I'd take an inhibitor you like, get chemical fingerprints (or whatever descriptor you prefer), and compare to the fingerprints for the rest of the dataset, then get rid of the 4-sigma outliers when comparing the distance between fingerprints. That should account for weight AND a variety of other properties that might fall outside the model.
Hi William,
You made some good points! I don't work with antibodies, just with small molecule inhibitors. But in the case of metallo-drugs, the actual MW threshold would fail as you correctly pointed out. I was thinking instead to use the SMILES of INChI string and set a minimum and maximum lengths to filter out putative non-binders. But, again what would be good values for these limits?
PS: the proceedure that you described about fingerprint similarity upraises the risk of low training data diversity that will result to a poor machine learning model.
True enough, but you're reducing your training diversity by limiting molecular weight as well. Besides, at four-sigma, assuming well distributed data, you're getting rid of less than 1% of the samples. I'd expect more like 5% for real data, but still - if your dataset diversity hinges on that small number of samples, something is probably very wrong somewhere.
Still - you appear to want somebody to give you arbitrary molecular weight values, but aren't happy with the Lipinski set. Literally nobody can do better than that, ESPECIALLY without knowing your binding site. If you don't like just eliminating the molecules most unlike the others, either for all properties or for weight alone, the best I can suggest is to do an alpha-sphere search in the binding site, figure the amount of, say, diamond that would fit into that volume, and use that as, say, 80% of your maximum mass. Then figure that anything that doesn't weigh at least, oh, 20% of that mass, won't make enough contacts to matter. You'll probably have to fiddle with that more than a little bit, but it would at least give you an algorithmic approach to your cutoffs, rather than arbitrary numbers.
As for smiles string lengths - I'd just do the mW calculation properly. It's a simple calculation, quick, and avoids the problem of strings with very little markup 'weighing' the same as strings with lots of rings, branch points, etc.
- Badges
- Science topic
- Similar topics
- Organizational Psychology
- Training