Closed benstaf closed 7 years ago
Hi.
You're right: there is an error in the paper. The _drugcandidate metric is a combination of solubility, novelty, synthesizability and conciseness. Just citing the corresponding function in the mol_metrics.py
file:
def drug_candidate(smile, train_smiles):
good_logp = constant_bump(logP(smile), 0.210, 0.945)
sa = SA_score(smile)
novel = soft_novelty(smile, train_smiles)
compact = conciseness(smile)
val = (compact + good_logp + sa + novel) / 4.0
return val
However, it is possible than in a previous stage of development this function was as said in the paper. You could try changing the previous snippet for this one, and optimizing:
def drug_candidate(smile, train_smiles):
good_logp = constant_bump(logP(smile), 0.210, 0.945)
sa = SA_score(smile)
novel = soft_novelty(smile, train_smiles)
diverseness = diversity(smile)
val = (diverseness + good_logp + sa + novel) / 4.0
return val
Anyway, if you are interested in generating drug-like molecules, check our recent work in chemically-oriented ORGANs. Among other things it implements Lipinski's rule-of-five and chemical beauty scores.
Feel free to ask any questions, Carlos
How do you compute Druglikeness in the paper?
Page 5, you say that Druglikeness is composed of a linear combination of Novelty, Diversity, Solubility and Synthetizability, without specifying the coefficients of this linear combination.
In the code (line 313 of mol_metrics.py), druglikeness seems to be the arithmetic mean of those parameters. However, this formula fails to give the figures announced in the paper.
To figure out those coefficients, I tried to solve the linear system of equations (with just 8 equations), but it is impossible again to find a solution. Can you provide clarifications please?![linearequations](https://user-images.githubusercontent.com/8183572/29006861-48af4368-7af8-11e7-866f-b9af5db43f29.png)