Reverse engineering molecules from fingerprints through deterministic enumeration and generative models.

Area of Science:

• Computational chemistry
• Cheminformatics
• Artificial intelligence in drug discovery

Background:

• Molecular design often uses fingerprints like Extended-Connectivity Fingerprints (ECFPs) to represent chemical structures.
• Reconstructing molecular structures from ECFPs is challenging due to information loss during vectorization.
• Existing AI methods highlight privacy risks in sharing ECFP data.

Purpose of the Study:

• To develop a deterministic algorithm for reconstructing molecular structures from ECFPs.
• To benchmark a Transformer-based generative model against this deterministic approach.
• To explore the potential of ECFP reconstruction for de novo drug design.

Main Methods:

• Developed a deterministic algorithm to reverse engineer molecular structures from ECFPs.
• Utilized MetaNetX and eMolecules databases for benchmarking.
• Trained and evaluated a Transformer-based generative model for predicting SMILES from ECFPs.

Main Results:

• The deterministic algorithm successfully reconstructs molecular structures from ECFPs.
• The Transformer model achieved 95.64% retrieval accuracy but had limitations in exhaustive enumeration.
• Reverse-engineered structures from a drug dataset included patented compounds and molecules with bioassay data.

Conclusions:

• The deterministic algorithm provides a conclusive method for molecular structure reconstruction from ECFPs.
• This technique enhances privacy in ECFP data sharing and offers a novel approach for de novo drug design.
• The findings demonstrate the feasibility of generating novel, potentially valuable molecular structures from ECFP data.