fingerprint_comparison

Comparing E3FP and ECFP4 Tanimoto Coefficients (TCs)

• Note: Here we assume that E3FP and ECFP4 fingerprints have already been generated and saved to e3fp_molecules.csv.bz2 and ecfp4_molecules.csv.bz2, respectively. See the discussion in crossvalidation for fingerprinting instructions.

Computing Pairwise Fingerprint TCs

For both sets of molecules files, run the following series of scripts. Where multiple fingerprints exist for a given molecule pair (e.g. multiple conformers for E3FP or tautomers for ECFP4), the maximum TC between all fingerprint pairs across the two molecules is computed. The results are multiple flat binary files, each consisting of a flattened lower (below- diagonal) pairwise TCs triangle matrix, corresponding to a contiguous chunk of the triangle matrix.

python get_fingerprint_tril_tcs.py <molecules_file> --merge_confs

Computing Pairwise FastROCS Shape and Combo Tanimotos

Run FastROCS on the same set of conformers used for computing pairwise E3FP TCs above. First, concatenate the SDF files:

python sdf_files_to_file.py $E3FP_PROJECT/conformer_generation/conformers_proto_rms0.5 combined.sdf

We must manually split the lower triangle matrix into batches consisting of roughly equal numbers of pairs.

python get_triangle_indices.py <mol_num> <batch_num>

This script prints out start and stop row indices (inclusive) for each batch. For each batch, run

python get_fastrocs_tril_tcs.py combined.sdf <start_index> <stop_index> --merge_confs

Merging Batch Files

Each of the above scripts produces multiple binary files. We need to assemble these into into a NumPy memmap file. A memmap enables relatively quick look-up without needing to read many doubles into memory. The format for a NumPy memmap file on disk is identical to the binary files generated above, so we only need to concatenate them.

cat $(ls *bin | sort -n -t - -k 2) > <out_memmap_file>

Binning TCs Comparisons

For comparing two sets of pairwise TCs, first use the above approach to build the two pairwise TCs memmaps. Then, count the number of TCs pairs that map to specific values at a specified precision (number of decimal places). For a given memmap pair, run

python count_pairwise_tcs.py <mmap_file1> <mmap_file2> --names Name1 Name2

which produces an output file name1_name2_tcs.csv.gz.

Computing Enrichment Curves

FastROCS output are pairwise shape and combo (shape + color) Tanimotos, as opposed to ECFP4's and E3FP's individual fingerprints. To compare the performance of all three approaches at enriching for actives, we perform k-fold cross-validation as previously described, using as the threshold for fingerprints the maximum pairwise TC computed between the set of query molecule fingerprints and target molecule fingerprints, and for ROCS the shape or combo max TC between the query molecule conformers and target molecule conformers. The enrichment curve plots fraction of actives recovered (sensitivity) vs fraction of database screened.

The MaxTC "classifier" can either dynamically compute the Tanimotos or read from a pre-computed memmap (see above). The former is significantly faster for fingerprints, due to the efficiency of large-scale TC computation.

For fingerprints, run

python $E3FP_PROJECT/scripts/run_cv.py <molecules_file> $E3FP_PROJECT/data/chembl20_binding_targets.csv.bz2 --method maxtc --reduce_negatives -l cv_log.txt

For FastROCS comparisons, run

python $E3FP_PROJECT/scripts/run_cv.py <molecules_file> $E3FP_PROJECT/data/chembl20_binding_targets.csv.bz2 --method maxtc --tc_files <tcs_memmap_file> <mol_names_file> --reduce_negatives -l cv_log.txt

All-By-All Comparisons of Random Conformers from Different Molecules

See random_conformers for specific instructions.

Pairwise Comparisons of Random Conformers of the Same Molecules

See random_conformer_pairs for specific instructions.