One chiral fingerprint to find them all

Abstract

Background: Molecular fingerprints are indispensable tools in cheminformatics. However, stereochemistry is generally not considered, which is problematic for large molecules which are almost all chiral. Results: Herein we report MAP4C, a chiral version of our previously reported fingerprint MAP4, which lists MinHashes computed from character strings containing the SMILES of all pairs of circular substructures up to a diameter of four bonds and the shortest topological distance between their central atoms. MAP4C includes the Cahn-Ingold-Prelog (CIP) annotation (R, S, r or s) whenever the chiral atom is the center of a circular substructure, a question mark for undefined stereocenters, and double bond cis-trans information if specified. MAP4C performs as good as the achiral MAP4, ECFP and AP fingerprints in non-stereoselective virtual screening benchmarks. Furthermore, it readily distinguishes between thousands of stereoisomers in complex natural products and peptides. Conclusion: MAP4C is recommended as a generally applicable chiral molecular fingerprint.

Can Large Language Models Predict Antimicrobial Peptide Activity and Toxicity?

“…Starting from the DBAASP dataset of 9,548 peptide sequences annotated with antibacterial activity and 2,262 peptide sequences annotated with hemolysis effect, we had previously evaluated NB, RF, SVM and RNN models, and found the latter to perform best for predicting both activity and hemolysis from sequence data. 13,14 For additional reference, we trained an SVM on the fraction of helical residues and the hydrophobic moment, two properties commonly known to correlate with antimicrobial activity, as well as another SVM on MAP4C, a molecular fingerprint that can reliably encode large molecules such as natural products and peptides including their chirality, 34 a parameter which we considered important since our data listed sequences containing both L-and D-amino acids.…”

Navigating a 1E+60 Chemical Space

“…We modified our previously reported PDGA 44 by computing fitness functions either as the Jaccard distance (dJ) to the target molecule computed using the molecular fingerprint MAP4C, 49 saving all generated molecules at each generation as trajectory molecules, or as the City Block Distance (dCBD)…”

“…In each case, we performed three PDGA runs of maximum 10,000 generations starting from 50 random sequences using the chiral fingerprint MAP4C, which encodes pairs of circular substructures with high precision including chirality. 48,49 PDGA identified the target molecule in less than 10,000 generation in at least one of the three runs for each of these six peptides, including the two 30-mer peptides 5 and 6, which required exploration of the full 1E+60 chemical space (Table 2). Since each generation only amounted to 35 new molecules, which were evaluated against the 15 best scoring molecules of the previous generation used as parents, the cumulative number of molecules generated in each trajectory only amounted to a few thousands, which is remarkably low considering the size of the explored chemical space.…”

“…To demonstrate LBVS, we modified our recently reported peptide design genetic algorithm (PDGA), 44 which evolves analogs of any target molecule by searching a topologically diverse oligomer space using molecular fingerprint similarity as fitness function, and can be used to design new analogs of known peptides as recently demonstrated experimentally for antimicrobial peptide dendrimers. 45 Specifically, we computed the fitness function using the macromolecule extended atom pair fingerprint (MXFP) 46,47 and the chiral MinHashed atom pair fingerprint (MAP4C), 48,49 both designed for large molecules.…”

Can large language models predict antimicrobial peptide activity and toxicity?

“…Starting from the DBAASP dataset of 9548 peptide sequences annotated with antibacterial activity and 2262 peptide sequences annotated with hemolysis effect, we had previously evaluated NB, RF, SVM and RNN models, and found the latter to perform best for predicting both activity and hemolysis from sequence data. 13,14 For additional reference, we trained an SVM on the fraction of helical residues and the hydrophobic moment, two properties commonly known to correlate with antimicrobial activity, as well as another SVM on MAP4C, a molecular fingerprint that can reliably encode large molecules such as natural products and peptides including their chirality, 34 a parameter which we considered important since our data listed sequences containing both Land D-amino acids.…”