Combination of Coevolutionary Information and Supervised Learning Enables Generation of Cyclic Peptide Inhibitors with Enhanced Potency from a Small Data Set

Abstract

Computational generation of cyclic peptide inhibitors using machine learning models requires large size training data sets often difficult to generate experimentally. Here we demonstrated that sequential combination of Random Forest Regression with the pseudolikelihood maximization Direct Coupling Analysis method and Monte Carlo simulation can effectively enhance the design pipeline of cyclic peptide inhibitors of a tumor-associated protease even for small experimental data sets. Further in vitro studies showed that such in silico-evolved cyclic peptides are more potent than the best peptide inhibitors previously developed to this target. Crystal structure of the cyclic peptides in complex with the protease resembled those of protein complexes, with large interaction surfaces, constrained peptide backbones, and multiple inter- and intramolecular interactions, leading to good binding affinity and selectivity.

Figure 1

In silico molecular evolution of bicyclic peptide inhibitors of huPA. MSA logo of 37 phage-encoded bicyclic peptides (input data) selected in vitro against huPA (top left). Training and validation data set were generated using amino acid sequences of all selected bicyclic peptides (“sequences”), their biochemical and biophysical properties (“features”), and the K_i values (“label”) determined for 37 bicyclic peptide molecules (top right). Combination of pseudolikelihood maximization direct coupling analysis (plmDCA) and Monte Carlo (MC) methods (left) with the Random Forest Regression (RFR) algorithm (right) yielded new peptide sequences with a preferential frequency of amino acids at each position (MSA logo, bottom right; Supplementary Table 1). MSA: multiple sequence alignment; K_i: inhibitory constant.

Figure 2

Biochemical characterization of in silico evolved bicyclic peptide inhibitors of huPA. a) MSA logo of bicyclic peptides derived from the iterative in silico process and predicted to have K_i values below 0.92 μM (that corresponds to 50th percentile); b) amino acid sequences of bicyclic peptides designed according to the sequence logo graph. The residues with the highest frequency (larger letters) were placed in each position. The sequences are arranged in groups according to sequence similarities; c) amino acid sequences of bicyclic peptides variants in which the in silico selected residues were reverted to those present in the parental phage-selected UK18 molecule. Identical or similar amino acids between different bicyclic peptide sequences are highlighted in color. The K_i values were determined at 25 °C and physiological pH (7.4) using the suitable substrate at the concentration of 50 μM. Mean values of at least three measurements are indicated S.E., standard error; d) column graph comparing the determined K_i values; e) scheme representing the contribution of mutated amino acid residues to the potency of inhibition; f) residual activities of huPA and a series of homologous human and murine trypsin-like serine proteases incubated with synthetic bicyclic peptide UK970 were determined at 25 °C, at physiological pH (7.4) using the suitable substrates at a concentration of 50 μM. The shown values are the means of three independent experiments. Data are presented as the mean (symbol). S.E., and standard error. The K_m values of each protease were determined by standard Michaelis–Menten kinetics and used in the calculation of the reported K_i values (Supplementary Table 2).

Figure 3

Structural comparison of the binding mode of bicyclic peptides UK18, UK965 and UK970 in complex with huPA. a) Molecular surface representation of the overall huPA-UK18, huPA-UK965, and huPA-UK970 superimposed complexes are shown in two orientations (90° rotation). Surface of huPA is colored in gray, while the peptide ribbon and mesitylene scaffold of UK18, UK965, and UK970 are colored in blue, pale green, and salmon, respectively; b) column graph reporting the total number of polar (both direct and H₂O-mediates) and nonpolar interactions of huPA with bicycle peptides UK18 (blue), UK965 (pale green) and UK970 (salmon); c) comparative analysis of the buried surface area (BSA) covered by UK965 in respective to UK18 (pale green) and that covered by UK970 in respective to UK18 (salmon); d) schematic representation of molecular interactions between huPA and UK970. Residues of huPA are labeled according to the chymotrypsin numbering system. Intermolecular salt bridges and hydrogen bonds are shown as red and blue dashed lines, respectively. Bicyclic peptide intramolecular hydrogen bonds are shown as green dashed lines. Bent gray lines indicate residues of UK970 in close contact with human uPA (distances shorter than 4.0 Å that are not polar intermolecular interactions).

Figure 4

Differences in the binding mode of bicyclic peptides UK965 and UK970 to huPA with respect to UK18. a) Detail view of previously solved X-ray structure of bicyclic peptide UK18 in complex with huPA (blue and gray, top) and bicyclic peptide UK970 in complex with huPA (salmon and gray, bottom). The presence of a Pro instead of an Ala in position 15 of UK970 variant appears to induce a sharp turn in the local geometry that induce a different spatial arrangement of one arm of the linker arm and ultimately impose a different conformation on the backbones of the opposite loop; b) the large conformational change induced by the distal Pro15 cause a repositioning of the Arg4 side chain that instead of forming an intramolecular salt-bridge with Glu6 (top huPA-UK18 complex, gray and blue) now points toward huPA and engages in intermolecular contacts with huPA (bottom huPA-UK970 complex, gray and salmon); c) molecular surface representation of the bicyclic peptides UK18, UK965 and UK970 color-coded according to hydrophobicity. Most hydrophobic residues and the mesitylene scaffold are shown in raspberry, whereas the most hydrophilic ones are shown in white; d) view of the amino acids surrounding the central chemical linker. The mesitylene core and the side chains of the mutated residues are shown as spheres. Hydrophobic residues and the mesitylene scaffold are shown in raspberry, whereas the hydrophilic ones are colored in white.

Figure 5

Further round of in silico molecular evolution on an enriched family of bicyclic peptide inhibitors of huPA. a) MSA logo of 52 phage-encoded bicyclic peptides (input data) selected in vitro against huPA (top left). Combination of pseudolikelihood maximization direct coupling analysis (plmDCA) and Monte Carlo (MC) methods with Random Forest Regression (RFR) algorithm (top middle) yielded new peptide sequences with a preferential frequency of amino acids at each position (MSA logo, top right) and predicted to have K_i values below 0.38 μM (that corresponds to 50th percentile). b) Left, amino acid sequences of bicyclic peptides designed according to the MSA logo of the new peptide sequences. Identical or similar amino acids between different bicyclic peptide sequences are highlighted in color. Right, column graph comparing the determined K_i values. The K_i values were determined at 25 °C and physiological pH (7.4) using the suitable substrate at the concentration of 50 μM. Mean values of at least three measurements are indicated S.E., standard error; c) Structural comparison of the binding mode of bicyclic peptides UK970 and UK971 in complex with huPA. Molecular surface of huPA is colored in gray, while the peptide ribbon and mesitylene scaffold of UK970 and UK971 are colored in salmon and blue, respectively. Selected amino acid side chains (Phe5 and Val7 in UK970; Trp5 and Thr7 in UK971) are represented as ball-and-stick and colored by atom type (carbon = salmon for UK970 and olive for UK971, oxygen = firebrick, nitrogen = deep blue).

Figure 6

In silico molecular evolution on a different family of bicyclic peptide inhibitors of huPA. a) MSA logo of 31 phage-encoded bicyclic peptides (input data) selected in vitro against huPA (top left). Training and validation data set were generated using amino acid sequences of all selected bicyclic peptides (“sequences”), their biochemical and biophysical properties (“features”) and the K_i values (“label”) determined for 31 bicyclic peptide molecules (top right). Combination of pseudolikelihood maximization direct coupling analysis (plmDCA) and Monte Carlo (MC) methods (middle left) with the Random Forest Regression (RFR) algorithm (middle right) yielded new peptide sequences with a preferential frequency of amino acids at each position and predicted to have K_i values below 1.97 μM (that corresponds to 50th percentile; MSA logo, bottom right). The MSA logo obtained using statistical methods (plmDCA and MC) combined to computational (RFR) algorithm differs from that obtained when applying solely statistical methods (MSA logo, bottom left); b) amino acid sequences and K_i values of bicyclic peptides UK974–UK978 designed according to the sequence logo graph. Identical or similar amino acids between different bicyclic peptide sequences are highlighted in color. As a reference, the amino acid sequence and K_i value of the parental phage-selected UK140 are also reported. The K_i values were determined at 25 °C and physiological pH (7.4) using the suitable substrate at the concentration of 50 μM. Mean values of at least three measurements are indicated S.E., standard error; c) column graph comparing the determined K_i values of synthetic bicyclic peptide UK140 and UK978 against human uPA (huPA), murine uPA (muPA) and human trypsin (hTryp) proteases. Residual activities were determined at 25 °C, at physiological pH (7.4), using the suitable substrates at a concentration of 50 μM. The shown values are the means of three independent experiments. Data are presented as mean (symbol). S.E., standard error. The K_m values of each protease were determined by standard Michaelis–Menten kinetics and used in the calculation of the reported K_i values (Supplementary Table 2).