Systems and methods for predicting potential inhibitors of target protein

A system for predicting potential inhibitors of a target protein, comprising:

• a memory storing instructions;
• one or more communication interfaces;
• and
• one or more hardware processors coupled to the memory via the one or more communication interfaces, wherein the one or more hardware processors are configured by the instructions to:
• receive an input comprising the target protein;
• apply a transfer learning technique on a pre-trained generative model using a target protein specific dataset (TPSD) to obtain a target-specific generative model (TSGM) for the target protein and learn target-specific features of the target protein specific dataset and design new chemical molecules, and the pre-trained generative model is obtained by:
• obtaining an input dataset comprising molecules in a pre-defined format and training a generative model with the input dataset to obtain the pre-trained generative model, and wherein the target protein specific dataset (TPSD) is a curated target protein specific dataset (CTPSD) obtained by:
• obtaining a plurality of other proteins having one or more properties similar to one or more corresponding properties of the target protein;
• and
• iteratively apply a reinforcement learning (RL) technique on the target-specific generative model and a predictive model to obtain a set of target-specific small molecules, wherein each target-specific small molecule from the set of target-specific small molecules comprises one or more associated physico-chemical properties and the one or more associated physico-chemical properties comprise a Synthetic Accessibility Score (SAS), a biological activity, a partition co-efficient (log P), a molecular weight, and the docking score, and the predictive model is pre-trained using a protein dataset, and the predictive model predicts the one or more associated physico-chemical properties for each target-specific small molecule from the set of target-specific small molecules, wherein the RL technique (i) obtains the set of target-specific small molecules based on a reward assigned to the target-specific generative model and (ii) increases binding affinity of the set of target-specific small molecules based on the reward assigned to the target-specific generative model, and the reward is used to optimize the target-specific generative model iteratively using a regularized reward function, and the target-specific generative model is subjected to the docking score optimization with the predictive model using a regularized reinforcement learning (RL) technique, and the regularized reward function comprises:
• r
• ⁡
• (
• x
• )
• =
• exp
• ⁡
• (
• -
• x
• 3.
• )
• where x is a docking score from the predictive model and r (x) is the reward for reinforcing the target-specific generative model;
• apply physico-chemical property filters on each target-specific small molecule from the set of target-specific small molecules to obtain a reduced set of target-specific molecules, wherein the physico-chemical property filters relate to an octanol-water partition coefficient (log P), a molecular weight (MW) and the synthetic accessibility score (SAS) associated with each target-specific small molecule, and the SAS avoids molecules which are difficult to synthesize, and each molecule from the reduced set of target-specific molecules comprises one or more desired physico-chemical properties, and small molecules are docked at an active site of the target protein are chosen to curate a target specific training dataset and molecules with high virtual screening scores (<=−7.0) are chosen to form a final target-specific small molecule dataset for transfer learning;
• perform a virtual screening of the reduced set of target-specific molecules against the target protein to obtain a screened set of target-specific molecules, wherein the screened set of target-specific molecules serve as one or more potential inhibitors of the target protein;
• and
• validate the screened set of target-specific molecules by comparing the screened set of target-specific molecules with a dataset of target-specific small molecules set aside for validation, wherein selective and diverse small molecules similar to the dataset of the target-specific small molecules are generated using the target-specific generative model by optimizing affinity of the small molecules towards the target protein by utilizing inhibitors of proteins with similar active site, thereby reproducing the existing target-specific small molecules from a validation dataset.

A non-transitory computer readable medium on which a computer readable program is stored, wherein the computer readable program, when executed on a computing device causes the computing device to perform a processor-implemented method for predicting potential inhibitors of a target protein comprising:

• receiving an input comprising the target protein;
• applying a transfer learning technique on a pre-trained generative model using a target protein specific dataset (TPSD) to obtain a target-specific generative model (TSGM) for the target protein and learn target-specific features of the target protein specific dataset and design new chemical molecules, wherein the pre-trained generative model is obtained by:
• obtaining an input dataset comprising molecules in a pre-defined format and training a generative model with the input dataset to obtain the pre-trained generative model, and the target protein specific dataset (TPSD) is a curated target protein specific dataset (CTPSD) obtained by:
• obtaining a plurality of other proteins having one or more properties similar to one or more corresponding properties of the target protein;
• and
• iteratively applying a reinforcement learning (RL) technique on the target-specific generative model and a predictive model to obtain a set of target-specific small molecules, wherein each target-specific small molecule from the set of target-specific small molecules comprises one or more associated physico-chemical properties and the one or more associated physico-chemical properties comprise a Synthetic Accessibility Score (SAS), a biological activity, a partition co-efficient (log P), a molecular weight, and the docking score, and the predictive model is pre-trained using a protein dataset, and the predictive model predicts the one or more associated physico-chemical properties for each target-specific small molecule from the set of target-specific small molecules, and the RL technique (i) obtains the set of target-specific small molecules based on a reward assigned to the target-specific generative model and (ii) increases binding affinity of the set of target-specific small molecules based on the reward assigned to the target-specific generative model, and the reward is used to optimize the target-specific generative model iteratively using a regularized reward function, and the target-specific generative model is subjected to the docking score optimization with the predictive model using a regularized reinforcement learning (RL) technique, and wherein the regularized reward function comprises:
• r
• ⁡
• (
• x
• )
• =
• exp
• ⁡
• (
• -
• x
• 3.
• )
• where x is a docking score from the predictive model and r (x) is the reward for reinforcing the target-specific generative model;
• applying physico-chemical property filters on each target-specific small molecule from the set of target-specific small molecules to obtain a reduced set of target-specific molecules, wherein the physico-chemical property filters relate to an octanol-water partition coefficient (log P), a molecular weight (MW) and the synthetic accessibility score (SAS) associated with each target-specific small molecule, wherein the SAS avoids molecules which are difficult to synthesize, and each molecule from the reduced set of target-specific molecules comprises one or more desired physico-chemical properties, and small molecules are docked at an active site of the target protein are chosen to curate a target specific training dataset and molecules with high virtual screening scores (<=−7.0) are chosen to form a final target-specific small molecule dataset for transfer learning;
• performing a virtual screening of the reduced set of target-specific molecules against the target protein to obtain a screened set of target-specific molecules, the screened set of target-specific molecules serve as one or more potential inhibitors of the target protein;
• and
• validating the screened set of target-specific molecules by comparing the screened set of target-specific molecules with a dataset of target-specific small molecules set aside for validation, wherein selective and diverse small molecules similar to the dataset of the target-specific small molecules are generated using the target-specific generative model by optimizing affinity of the small molecules towards the target protein by utilizing inhibitors of proteins with similar active site, thereby reproducing the existing target-specific small molecules from a validation dataset.