Benchmarking Zinc-Binding Site Predictors: A Comparative Analysis of Structure-Based Approaches

Abstract

Metalloproteins play crucial physiological roles across all domains of life, relying on metal ions for structural stability and catalytic activity. In recent years, computational approaches have emerged as powerful and increasingly reliable tools for predicting metal-binding sites in metalloproteins, enabling their application in the challenging field of metalloproteomics. Given the growing number of available tools, it is timely to design a reproducible approach to characterize their performance in specific usage scenarios. Thus, in this study, we selected some state-of-the-art structure-based predictors for zinc-binding sites and evaluated their performance on two data sets: experimental apoprotein structures and structural models generated by AlphaFold. Our results indicate that apoprotein structures pose significant challenges for predicting metal-binding sites. For these systems, the predictors achieved lower-than-expected performance due to the structural rearrangements occurring upon metalation. Conversely, predictions based on AlphaFold models yielded significantly better results, suggesting that they more closely resemble the holo forms of metalloproteins. Our findings highlight the great potential of metal-binding site predictions for advancing research in the field of metalloproteomics.

1

Percentage distribution of sites across CLESs and structures. (Top) The histogram shows the percentage distribution of the number of sites across CLESs. The number above each bar indicates how many CLESs have the corresponding number of sites. (Bottom) The histogram shows the percentage distribution of the number of sites per structure. The number above each bar indicates how many structures harbor the corresponding number of sites.

2

Violin plot for CLES 22022, composed of 31 sites. Each violin represents the distribution of the distance between two selected donor atoms in the binding site. The white line at the center of each violin indicates the median value, while the bars around it mark the first and third quartiles. The circles and the triangles of the same color correspond to the sites whose distance matrices contain at least two elements with values beyond two standard deviations from the mean. the site of this cluster consists of 1 cysteine (Sγ), 1 aspartate (Cγ) and 1 histidine (Cε1), where Sγ, Cγ and Cε1 are the donor atom or donor atom proxy.

3

Performance of the predictors, based on the top 5 outputs. True positive (TP), redundant true positive (RTP), false positive (FP) and recall (R) values are displayed for all the predictors.

4

Percentage of CLESs identified by the predictors.

5

(A) Percentage of correctly predicted sites in CLESs featuring large rearrangements of the MBS. The performance of the three best predictors, each shown in a different color, is computed as the percentage of correctly identified sites over the structures in each CLES. Five different CLESs characterized by a high spread of side chains were selected. (B) Violin plot of CLES 18431. The site of this cluster consists of 4 cysteines, whose donor atoms are the Sγ atoms. (C) View of the 3D structure of the sites in CLES 18431. This panel shows the spread of side chains for three different apo-sites belonging to CLES 18431. Each apo-site is represented by a different color.

6

Precision and recall (R­(top)) of the selected predictors as a function of the number of outputs considered.

7

Precision and recall (R (top)) of the selected predictors as a function of the number of nonredundant outputs.

8

F1-score of the selected predictors as a function of the number of considered outputs. The F1-score has been calculated using the data of Figure .

9

Venn diagram representing the agreement of the predictors for the second benchmark. The diagrams represent the intersections of the true positives (TP) identified by the four tools. The percentages in the left panel were computed with respect to the total of 62 sites identified by at least two predictors.

10

Recall compared with precision for all the predictors. Recall is defined as the ratio between TP and the total number of expected sites in the data set.