SuperMetal: A Generative AI Framework for Rapid and Precise Metal Ion Location Prediction in Proteins

Abstract

Metal ions, as abundant and vital cofactors in numerous proteins, are crucial for enzymatic activities and protein interactions. Given their pivotal role and catalytic efficiency, accurately and efficiently identifying metal-binding sites is fundamental to elucidating their biological functions and has significant implications for protein engineering and drug discovery. To address this challenge, we present SuperMetal, a generative AI framework that leverages a score-based diffusion model coupled with a confidence model to predict metal-binding sites in proteins with high precision and efficiency. Using zinc ions as an example, SuperMetal outperforms existing state-of-the-art models, achieving a precision of 94 % and coverage of 90 %, with zinc ions localization within 0.52 ± 0.55 Å of experimentally determined positions, thus marking a substantial advance in metal-binding site prediction. Furthermore, SuperMetal demonstrates rapid prediction capabilities (under 10 seconds for proteins with ∼ 2000 residues) and remains minimally affected by increases in protein size. Notably, SuperMetal does not require prior knowledge of the number of metal ions-unlike AlphaFold 3, which depends on this information. Additionally, SuperMetal can be readily adapted to other metal ions or repurposed as a probe framework to identify other types of binding sites, such as protein-binding pockets.

Keywords: Generative AI; diffusion model; metal-binding sites; metalloprotein.

Fig. 1

Workflow of SuperMetal. The orange spheres represent the sampled Zn ions. The protein, shown in blue, is from 2J9R in the RCSB Protein Data Bank.

Fig. 2. Precision Versus Coverage for SuperMetal and Metal3D at Different Probability Cutoffs.

The graph compares SuperMetal (purple line) and Metal3D (green line), illustrating how precision varies with coverage at different cutoff values. The labels indicate the probability cutoffs, $p$ for the confidence model in SuperMetal and $t$ for Metal3D.

Fig. 3. Distribution of Mean Absolute Deviation (MAD) for SuperMetal and Metal3D at Increasing Probability Cutoffs.

This violin plot illustrates the MAD for metal ion location predictions by SuperMetal and Metal3D, with cutoffs ranging from lower to higher values for each model. SuperMetal is shown in purple and Metal3D in green. The plot employs kernel density estimation to display the data distribution, with the white circle indicating the median, the black box defining the first and third quartiles, and whiskers extending up to 1.5 times the interquartile range to capture the spread of typical data values.

Fig. 4. Computational Time Analysis for SuperMetal and Metal3D Across Protein Sizes.

This scatter plot compares the computational time required by SuperMetal and Metal3D to predict metal-binding sites against the number of protein residues. Polynomial regression curves (purple and green dashed lines) are only used to clarify the trends.

Fig. 5. Comparative visualization of Zn ion binding site predictions for the proteins 5IN2 (1) and 6BTP (2).

Zn ions are color-coded as follows: grey for experimentally determined Zn ions, cyan for Metal3D predictions, orange for SuperMetal predictions, and blue for AlphaFold 3 predictions. The protein structure is shown in green for the same input PDB file used in Metal3D and SuperMetal, while it is shown in yellow for AlphaFold 3. The transparent green region around the Zn ions highlights the protein’s atomic structure within a 5 Å radius of the metal ions. From left to right, the figure shows varying numbers of Zn ions specified in the AlphaFold 3 input, ranging from 1 Zn ion, to 2, and finally to 6 Zn ions.

Fig. 6. Fundamental Theory of the Score-based Generative Diffusion Model for Metal Ions in Proteins.

The forward continuous-time stochastic differential equation (SDE) transitions the true locations of metal ions (top left) to random locations (top right). The score at each intermediate time step is predicted by a deep learning neural network, enabling the reverse process of the SDE. The grey part of the protein (top) represents the atomic structure of the protein surrounding the metal ions at their original positions.