Substrate Activation Efficiency in Active Sites of Hydrolases Determined by QM/MM Molecular Dynamics and Neural Networks

Abstract

The active sites of enzymes are able to activate substrates and perform chemical reactions that cannot occur in solutions. We focus on the hydrolysis reactions catalyzed by enzymes and initiated by the nucleophilic attack of the substrate's carbonyl carbon atom. From an electronic structure standpoint, substrate activation can be characterized in terms of the Laplacian of the electron density. This is a simple and easily visible imaging technique that allows one to "visualize" the electrophilic site on the carbonyl carbon atom, which occurs only in the activated species. The efficiency of substrate activation by the enzymes can be quantified from the ratio of reactive and nonreactive states derived from the molecular dynamics trajectories executed with quantum mechanics/molecular mechanics potentials. We propose a neural network that assigns the species to reactive and nonreactive ones using the Laplacian of electron density maps. The neural network is trained on the cysteine protease enzyme-substrate complexes, and successfully validated on the zinc-containing hydrolase, thus showing a wide range of applications using the proposed approach.

Keywords: AI; Laplacian of electron density; QM/MM MD; hydrolases; neural network; substrate activation.

Figure 1

(A) The first step of the reaction initiated by the nucleophilic attack of the carbonyl carbon atom of the substrate by the catalytic moiety Nu: X = O,C,N. Dashed violet lines depict hydrogen bonds. ES stands for the enzyme-substrate complex, and TI for the tetrahedral intermediate. (B) Gibbs energy profile of the imipenem hydrolysis in the active site of the metallo-β-lactamase NDM−1 (the upper panel) [6]; heterogeneity of the reactant states of the NDM-1-imipenem complex with respect to the distance of the nucleophilic attack [7] (the lower panel).

Figure 2

The workflow. QM/MM molecular dynamic trajectories are simulated. Laplacian of electron density maps are calculated in the plane of the carbonyl group and a nucleophilic atom. The inset shows the Laplacian of electron density map with zero (green bold lines), positive (blue dashed lines), and negative (red solid lines) isovalues. Maps are reduced to only zero value contour lines. CNN analyzes only a part of the images comprising a carbonyl group (highlighted green for reactive, and red for non-reactive species) to make a binary classification: reactive or nonreactive.

Figure 3

(A) The scheme of the CNN for discrimination of reactive and nonreactive states in the active sites of hydrolases. (B) The CNN fitting procedure; weights obtained in epoch 10 were utilized.

Figure 4

Analysis of images obtained from the frames of QM/MM MD trajectories of enzyme substrate complexes of the M^pro and substrates with Pro, Ser, or Thr at P2. (A) Confidence of reactivity determination for the reactive (marked with R, green circles) and non-reactive (marked with N, red circles) species. Arrows demonstrate the points corresponding to the images from panels (B,C) with the lowest values of confidence. (B) Images with the lowest confidence of determination among the reactive species. (C) Images with the lowest confidence of determination among the non-reactive species. (D) Alternation of reactive (with 1 value at y-axis) and non-reactive (with 0 value at y-axis) species along the QM/MM MD trajectory of the reactant complex of the M^pro and a substrate containing an Ile residue at P2.

Figure 5

(A) Confidence of reactivity determination for the dataset from the NDM-1-imipenem complex containing both reactive (green circles) and non-reactive (red circles) species. A nucleophile (yellow) and an oxyanion hole (lavender) in the cysteine protease M^pro (B) and metallo-β-lactamase NDM-1 (C).