Method and device for designing target-specific drug in which deep-learning algorithm is combined with water pharmacophore model

1 . A device for designing a target-specific drug, the device comprising:

a memory to store instructions; and

a processor configured to, by executing the instructions:

construct a water pharmacophore (WP) model by identifying hydration sites at a binding site of a target protein through molecular dynamics simulation,

perform a first WP screening on commercially available compound database based on the WP model to screen WP complementary compounds that are capable of complementarily binding to the binding site of the target protein;

construct a target deep-learning model that is trained specifically to the target protein, by performing deep learning, using the WP complementary compounds as a learning target, on a deep-learning model which encodes an input compound into a latent vector and decodes the latent vector into an output compound;

generate novel-compound candidates by decoding a sum vector obtained by summing a WP complementary latent vector and a random vector of each of the WP complementary compounds through the target deep-learning model; and

generate a library that is formed specifically to the target protein, by performing a second WP screening on the novel-compound candidates based on the WP model.

2 . The device of claim 1 , wherein

the processor is further configured to, when constructing the WP model,

calculate thermodynamic properties of the hydration sites, and

assign WP features to the hydration sites based on the thermodynamic properties.

3 . The device of claim 2 , wherein

the thermodynamic properties comprise energy properties, entropy properties, and hydrogen bonding properties of the hydration sites, and

the WP features comprise a positive charge feature, a negative charge feature, a hydrophobic feature, an aromatic feature, a hydrogen bond acceptor, and a hydrogen bond donor.

4 . The device of claim 2 , wherein

the processor is further configured to, when screening the WP complementary compounds, perform a first WP screening based on a tolerance and a matching number of the WP features.

5 . The device of claim 4 , wherein

the tolerance and the matching number are adjusted such that the WP complementary compounds are screened in such a number that is equal to or greater than 20,000 and equal to or less than 100,000 depending on the kind of the target protein.

6 . The device of claim 1 , wherein the processor is further configured to, when constructing the target deep-learning model, build such that a SMILES string of the output compound becomes as identical to possible to a SMILES string of the input compound, and the training is considered successful when a reconstruction loss, which quantifies a difference between the two SMILES strings, decreases and converges to a specific value.

7 . The device of claim 1 , wherein

the processor is configured to, when generating the novel-compound candidates,

form a latent space including WP complementary latent vectors of the WP complementary compounds, and

sum a latent vector randomly extracted from the latent space and the random vector.

8 . The device of claim 7 , wherein

the processor is further configured to, when setting as the random vector,

adjust a variance of a Gaussian random distribution based on diversity and validity of the novel-compound candidates.

9 . The device of claim 1 ,

wherein a screening criteria of the first WP screening and a screening criteria of the second WP screening are determined based on an allowable proximity distance of WP features and a number of matches of WP features, and

wherein the screening criteria of the second WP screening has a smaller allowable proximity distance or a greater number of matches than the screening criteria of the first WP screening.

10 . The device of claim 1 , wherein

the processor is, by executing the instructions, further configured to:

verify ease of binding of novel compounds of the library to the target protein by comparing a first docking score with respect to the active compounds of the target protein, a second docking score with respect to decoy compounds, and a third docking score with respect to the novel compounds of the library, and

verify novelty of the novel compounds of the library based on a tanimoto coefficient of active compounds and a tanimoto coefficient of the novel compounds of the library.

11 . The device of claim 1 , wherein

the target protein has the crystal structure of any one of androgen receptor (AR), glucocorticoid receptor (GR), progesterone receptor (PR), poly ADP-ribose polymerase (PARP), acetylcholinesterase (AChE), and peroxisome proliferator-activated receptor-gamma (PPARγ).

12 . A method of designing a target-specific drug, which is performed by a processor executing instructions stored in a memory, the method comprising:

constructing a water pharmacophore (WP) model by identifying hydration sites at a binding site of a target protein through molecular dynamics simulation;

performing a first WP screening on commercially available compound database based on the WP model to screen WP complementary compounds that are capable of complementarily binding to the binding site of the target protein;

constructing a target deep-learning model that is trained specifically to the target protein, by performing deep learning, using the WP complementary compounds as a learning target, on a deep-learning model which encodes an input compound into a latent vector and decodes the latent vector into an output compound;

generating novel-compound candidates by decoding a sum vector obtained by summing a WP complementary latent vector and a random vector of each of the WP complementary compounds through the target deep-learning model; and generating a library that is formed specifically to the target protein, by performing a second WP screening on the novel-compound candidates based on the WP model.