Molecular identification via molecular fingerprint extraction from atomic force microscopy images

Area of Science:

• Chemical Physics
• Computational Chemistry
• Materials Science

Background:

• High-resolution atomic force microscopy (HR-AFM) with CO-functionalized tips offers unprecedented resolution for imaging molecular structures.
• Deep learning (DL) models have shown potential in extracting chemical and structural information from HR-AFM images for molecular identification.
• Previous DL approaches utilized human-intuitive descriptors, which were sub-optimal for neural network performance.

Purpose of the Study:

• To develop a novel DL model for molecular identification from HR-AFM images using optimized topological fingerprints.
• To improve the accuracy and reliability of molecular identification compared to existing methods.
• To assess the feasibility of applying this method to experimental HR-AFM data.

Main Methods:

• Utilized 1024-bit Extended Connectivity Chemical Fingerprints of radius 2 (ECFP4) as molecular descriptors.
• Developed a DL model to extract ECFP4 from 3D HR-AFM image stacks.
• Implemented virtual screening for molecule identification based on predicted ECFP4.
• Integrated a second DL model to predict chemical formulas from HR-AFM data to complement ECFP4 information.
• Assigned Tanimoto similarity scores to quantify identification confidence.

Main Results:

• The DL model achieved 95.4% retrieval accuracy for molecular identification using predicted ECFP4 from theoretical HR-AFM images.
• Combining ECFP4 prediction with chemical formula prediction boosted identification accuracy to 97.6%.
• The model provides a confidence score (Tanimoto similarity) for each identification.
• Initial tests with experimental HR-AFM images showed promising results.

Conclusions:

• The proposed method effectively extracts ECFP4 fingerprints from HR-AFM images for accurate molecular identification.
• Complementing fingerprint-based identification with chemical formula prediction significantly enhances accuracy.
• This approach offers a reliable and confident method for molecular identification from AFM data.
• The findings support the potential application of this pipeline in real-world experimental conditions.