Statistical analysis tool to help identify molecular fragments that promote, or detract from, target properties.
Sepecifically, this tool calculates the "z-scores" of molecular substructures in a given sub-population of a database to identify fragments that are over- or under-represented in this sub-population relative to a reference population. These substructures can either be specified by the user, or automatically generated using Morgan fingerprints.
molZ relies heavily on RDKit, which I recommend installing via conda forge:
$ conda install -c conda-forge rdkit After that, molZ can be installed with pip:
$ pip install molz Using auto-generated fragments:
from molz import ZScorer # instantiate scorer class, optionally set length and radius of morgan fingerprint. # In this case, data.csv is a .CSV file of two columns: SMILES and computed LogP. scorer = ZScorer('data.csv', fp_rad=3, fp_bits=4096) # We are going to compute zscores of fragments present in high logp molecules. scorer.score_fragments('penalised_logp', [12, 25]) # We can plot a bar graph of zscores for the 15 highest and lowest scoring fragments. # Also, we can draw a given fragment by refering to its Morgan fingerprint bit index. scorer.plot(k=15, save_to='zscores_auto.png') scorer.draw_fragment(3595)Using user-defined fragments:
from molz import ZScorer # instantiate scorer class. In this case, data.csv is a .CSV file of two columns: # SMILES and computed LogP. scorer = ZScorer('data.csv') # We are going to compute zscores of fragments present in high logp molecules. scorer.score_fragments( 'penalised_logp', [12, 25], fragment_smiles=['CCCC', 'OC', 'N(C)(C)'] ) # We can plot a bar graph of zscores for the 15 highest and lowest scoring fragments. # Also, we can draw a given fragment by refering to its SMILES. scorer.plot(k=15, save_to='zscores_user.png') scorer.draw_fragment('CCCC')We will use the data from "Design Principles and Top Non-Fullerene Acceptor Candidates for Organic Photovoltaics" by Lopez et. al. as an example.
First, we need the data, which comes from the article supplementary info:
$ curl https://ars.els-cdn.com/content/image/1-s2.0-S2542435117301307-mmc2.csv > lopez_data.csv Now, we will use molz to detect over- and under-represented molecular fragments in molecues with a predicted HOMO energy of greater than -5 eV.
We will use a relatively large number of fingerprint bits, to minimize bit collisions.
from molz import ZScorer # we will use the 'HOMO_calc' data column. scorer = ZScorer('lopez-data.csv', fp_bits=8192, fp_rad=3) scorer.score_fragments('HOMO_calc', [-5, 10]) scorer.plot(k=40, figsize=(12, 3), save_to='example_pce.png', top_only=True, log_y=True)Which gives the following plot:
In this case the fragments (x axis) are expressed as bit vector positions. Drawing the top fragments reveals a series of particularly electron-rich substructures, which is what we'd expect for relatively high-energy HOMO orbitals. For instance:
scorer.draw_fragment(5773) 