AI-Powered Drug Discovery and Interaction Prediction

Predicting drug efficacy and side effects using machine learning to integrate chemical, cellular, and clinical data for more accurate drug discovery. The method involves acquiring drug chemical structure, pharmacological properties, and cellular action data. A machine learning model estimates drug properties based on the chemical and pharmacological data. Then, retraining the model using the estimated properties predicts efficacy and side effects. This integrates chemical and cellular data to predict clinical outcomes.

Source

Method for predicting compounds with improved drug properties by analyzing structural differences between compounds. The method involves using an artificial intelligence model that takes two compounds as input instead of just one. The model predicts the difference in drug properties between the two compounds based on their structural differences. This allows finding compounds with improved properties by swapping certain moieties from a starting compound. The predicted property differences, reliability, and meeting desired ranges are used to prioritize and recommend alternative compounds.

Source

Predicting adverse drug interactions by recovering missing attribute data for drugs. The method involves constructing a recovery model to fill in missing attribute values for drugs based on shared and unique attributes. This recovered data is then used to train a prediction model for adverse drug interactions. By recovering missing attributes, the prediction accuracy improves compared to ignoring missing values. This enables more accurate adverse drug interaction prediction and safer medication.

Source

Method for predicting drug interaction effects using pre-trained molecular graph models to improve accuracy and scalability compared to traditional methods. The approach involves pre-training a multi-drug representation model using molecular graph structures and drug target information. This pre-trained model is then used to predict drug interactions and combinations by leveraging learned drug-drug and drug-internal latent feature representations. The pre-training allows the model to extract potential characteristics of drugs from multiple aspects and integrates potential characteristics of multiple drugs for scalable multi-drug molecular representation. It also enables the model to predict interactions using only structural and target data without complex feature engineering.

Source

Predicting the amount of pharmacokinetic change of one drug in a pair of drugs under the influence of the other drug. The prediction is based on machine learning models using drug property and interaction data. It involves receiving the two drugs as input, determining if they match interaction data, and using a learned prediction model to infer the change in pharmacokinetics of the first drug caused by the second drug. This allows predicting potential drug interactions and evaluating the impact of swapping one drug for another. The models leverage drug properties like classification, structure, and metabolism.

Source

A drug pathway prediction method based on network reasoning to predict potential drug-pathway relationships without relying on negative samples. The method involves constructing a heterogeneous network of substructures, drugs, and pathways from molecular and clinical data. It then uses network reasoning algorithms to calculate drug-pathway relationships by propagating initial resources through the network. This allows predicting potential pathways for new drugs and repurposing old drugs based on their substructure similarity. The method provides a drug pathway prediction model for finding new therapeutic applications of existing drugs by leveraging their similarity to known drugs with desired pathway effects.

Source

A drug-drug interaction prediction method based on tensor neural networks that improves interpretability and enables predicting multiple types of interactions between the same pair of drugs. The method decomposes drugs into substructures, constructs substructure interaction tensors, and uses a tensor neural network to learn the substructure interactions. This allows predicting drug-drug interactions by inferring the substructure interactions. The decomposed substructure representation provides interpretability compared to directly studying whole drug structures.

Source

Method for predicting the effect of combination drugs using a machine learning ensemble model to efficiently predict the effect of new drug combinations. The method involves converting gene-level cell and drug data into pathway-level data, learning the cell, drug, and correlation data using gradient boosting classifiers, and ensemble learning the classifiers to make combination drug effect predictions. This allows predicting drug synergy by inferring from cell, drug, and correlation data using machine learning.

Source

Predicting adverse drug reactions (ADRs) and identifying potential target proteins involved using only the molecular structure of a drug as input. The method involves calculating drug-target interaction features from the drug's structure and target protein structures, running classifier models to predict ADRs based on those features, and generating output indicating predicted ADRs. This allows automated prediction of ADRs for new drugs and identification of potential targets contributing to ADRs.

Source

Using deep learning models with Transformer networks to predict and explain drug metabolism interactions. The method involves learning drug molecular graph and substructure representations using Transformer networks, and then using joint attention mechanisms to capture the interaction between drugs. This allows predicting the type and magnitude of metabolic interactions between drugs while explaining the structural basis. The approach leverages deep learning to improve medication safety by accurately identifying and explaining drug metabolism interactions.

Source

Method for predicting drug-protein interactions using a convolutional neural network (CNN) that encodes drugs and proteins jointly to capture their consistent relationship. The method involves constructing a CNN model to calculate interaction probabilities between drugs and proteins based on their jointly encoded representations. This allows the model to analyze the potential identity between drugs and proteins, which can improve drug-protein interaction prediction compared to independent encoding methods. The CNN encodes drugs and proteins simultaneously to capture their consistent relationship, unlike previous methods that encoded drugs and proteins separately. This joint encoding helps the model analyze the potential identity between drugs and proteins, which can improve drug-protein interaction prediction compared to independent encoding methods. The method addresses limitations of previous network-based methods that independently encoded drugs and proteins.

Source

Drug screening method and system using machine learning to predict protein inhibition and in vivo activity of drugs, and then optimize drug structures to improve both properties. The method involves: 1) collecting drug molecules with target protein inhibition or in vivo activity, processing the structures, and predicting protein inhibition and disease treatment effectiveness using a multi-task deep learning model. 2) Using reinforcement learning to train drug structures, extract high scoring substructures for protein inhibition and disease treatment, merge them, and optimize the merged structure using a pre-trained encoder-decoder. This generates new drugs with both improved target protein inhibition and disease treatment effectiveness.

Source

A drug-drug interaction prediction model called SSF-plus I that combines sequence and substructure features of drugs to improve accuracy compared to models using just molecular graphs or sequences. The SSF-plus I model captures broader feature information by integrating topological and sequential drug representations. It uses graph neural networks to extract structural features from drug molecules, but also separately encodes drug sequences using transformer-like self-attention. The model combines these sequence and substructure representations for drug interaction prediction. This allows it to leverage both molecular structure and sequence similarity to better capture drug interaction patterns.

Source

Predicting drug-target interactions using only sequence information using a pre-trained natural language processing model. The prediction includes first inputting each sequence of a drug and target protein into a pre-trained natural language processing model and outputting an output embedding; second generating a vector by connecting some tokens among the output embeddings of the natural language processing model; third inputting the vector into an artificial neural network model and outputting the drug-target interaction probability.

Source

Method for identifying drug products using AI engines specialized for specific groups of drugs instead of a single large model. The method involves filtering drug product characteristics to categorize drugs into groups. AI engines are trained for each group. When a drug needs identification, its characteristics are used to select the matching engine for accurate NDC prediction or verification. This improves efficiency by avoiding retraining a large model with many drugs. It also allows smaller entities to train their own engines for the drugs they distribute. A consensus engine checks training consistency across sources to validate new drugs.

Source

Predicting efficacy of drugs for treating diseases using a machine learning model that leverages drug-drug interaction and drug-target interaction data. The model takes as input the chemical structure of a drug and its known interactions with other drugs and targets. It predicts the drug's efficacy for a specific disease based on these interactions. The model is trained using approved drugs and their disease indications. This allows repurposing of existing drugs for new diseases by predicting their efficacy based on known interactions.

Source

A reliable drug-target correlation prediction method using subjective logic to predict drug-target associations and provide reliability estimates. The method involves training a subjective logic model to predict the probability of activity or inactivity for a given drug-target pair. The model uses evidence from known associations to calculate the reliability of the prediction. This allows prioritizing and ranking drug-target pairs based on both predicted correlation and reliability. The reliability estimates can help reduce experimental costs by avoiding waste on unreliable predictions.

Source

Method and device for checking drug interaction that identifies high-risk drug combinations using machine learning on medical records. The method involves generating a drug combination set from medical records, calculating odds ratios for each combination, finding fractions for each drug, and outputting combinations with high odds ratios, high drug fractions, and low quotients. This indicates combinations where the first drug's odds are high, the second drug's odds are low, and the first drug fraction is greater. This suggests interaction between the drugs, not just high-risk individual drugs.

Source

Predicting drug-drug interactions (DDIs) using chemical structure. The prediction includes evaluating a structural similarity between new, unseen drugs of interest and baseline drugs, based on the first substance data element and the relevant baseline drug data element; selecting a first subset of the plurality of baseline drugs, based on the first substance data element and the relevant baseline drug data element; and predicting a DDI between the first substance of interest and the second substance of interest, based on the first substance data element, the second substance data element, and the DDI data structure.

Source

Method for predicting adverse drug reactions between medications using molecular structure plane features to improve accuracy compared to traditional methods based on molecular substructure similarity. The method involves encoding the 2D planar arrangement of atoms in drug molecules instead of just substructure similarity. This captures important positional relationships that influence adverse reactions. The plane feature coding is used to predict adverse drug reactions between medications. The technique reduces the spatial dimension of molecular structure encoding and combines weight information to improve adverse reaction prediction compared to substructure similarity.

Source

Drug interaction prediction method and system based on substructure to improve accuracy compared to existing substructure-based methods. The technique utilizes correlations between substructure interactions to learn harder-to-learn interactions. It uses a zigzag iterative network that learns clearer substructure interactions first and then leverages those to capture fuzzier interactions. This helps capture more substructure interactions for drug interaction prediction. The method involves obtaining the substructure diagrams of drugs, feeding them into the iterative network to learn interactions, and using the network to predict drug interactions based on substructure overlap.

Source

Predicting interactions between chemicals and protein targets using deep learning models. The method involves encoding chemical structures and protein sequences using neural networks to generate fixed-length representations. These representations are fed into a deep learning model with an attention mechanism to predict interaction between chemicals and targets. The model learns to directly map chemical structures to interaction properties without requiring known ligand-target pairs.

Source

Predicting synergistic drug combinations for specific diseases like cancer using deep learning on protein interaction networks. The method involves a Graph Convolutional Network (GCN) to learn the topological relationships between drugs, diseases, and proteins in a protein-protein interaction (PPI) network. The GCN uses an attention mechanism to allocate weights to proteins based on their importance for drug synergy prediction. The weights reveal key proteins that contribute to the connectivity and mechanism of the PPI network. This allows identifying synergistic drug combinations for specific diseases by leveraging the protein interactions.

Source

A method for predicting adverse drug reactions between drug pairs based on drug dependence. The method involves building a model to predict adverse reactions between drug pairs by considering drug dependence during treatment. It uses known drug pair adverse reaction data, drug structure data, and side effect data to train the model. To predict adverse reactions for an unknown drug pair, the model takes the predicted structure vectors of the drugs in the pair, their predicted side effect vectors, and the known structure and side effect correlation matrix as input to quickly predict the adverse reactions between the drugs.

Source

Method to repurpose drugs by using deep learning models to predict the efficacy of drugs on new diseases based on their molecular structure and cellular genomics. The method involves encoding the drug molecule and cell genomics into vector representations, then using multi-head attention to capture the interaction between the vectors. This interaction representation is used to predict drug response to new diseases. The multi-head attention extracts the relationship between drug structure and cell genomics. This improves prediction accuracy for drug repurposing compared to just using molecular or genomic data alone.

Source

Drug target affinity prediction method using digital twin and distillation BERT to accurately and reliably predict drug-target interactions with biological interpretability. The method involves leveraging 3D protein structure information from AlphaFold2 to simulate the spatial bioreaction between drugs and targets. It constructs a bipartite graph with 3D structures, adds context from sequences, and uses bi-directional attention to predict affinity. The method achieves higher accuracy and robustness compared to sequence-based methods, as it considers 3D structure interactions.

Source

Method for predicting drug-target protein interactions using deep learning techniques. It involves extracting latent features from preprocessed drug and protein sequences using variational autoencoders. These features are then screened using an attention mechanism to obtain more precise representations. Finally, a multi-layer perceptron is used to predict the interaction between the drug and protein based on the screened features. This improves accuracy compared to prior methods.

Source

Protein-drug interaction prediction model generation, protein-drug interaction prediction, and uncertainty determination. The method involves generating protein and drug phenotypes using transfer learning and graph structures, respectively. A Bayesian neural network is trained using these phenotypes and interactions. During prediction, dropout is applied multiple times to determine uncertainty, averaged prediction is the interaction, and uncertainty is calculated from the distribution.

Source

A method for predicting protein-drug interactions and determining the uncertainty of those predictions using a Bayesian neural network. The method involves generating protein and drug phenotype data from the original protein and drug molecule data. This phenotype data is then used to train a Bayesian neural network to predict protein-drug interactions. The network outputs both the interaction prediction and an uncertainty score. This allows identifying uncertain predictions and potentially removing low-quality data from the training set to improve model reliability.

Source

Using neural networks to accurately predict drug side effects when multiple drugs are administered concurrently. The technique involves training a multimodal cell complex neural network (MCXN) on a dataset containing drugs, proteins, interactions, and side effects. The MCXN can predict side effect probabilities, severities, and frequencies for combinations of 3 or more drugs not in the training set. It can also monitor changes in side effects over time due to factors like new drugs or lifestyle. This allows personalized treatment optimization and proactive side effect management.

Source

Predicting side effects of multiple drugs administered concurrently using a neural network called multimodal cell complex neural network (MCXN). The MCXN can handle higher-order interactions between drugs and proteins beyond pairwise relations, enabling more accurate prediction of complex side effect profiles. The MCXN models drugs, proteins, interactions, and side effects using a multimodal cell complex representation. It predicts side effect severity and frequency for multiple drug combinations, ranks them based on overall side effects, monitors changes in side effects over time, and predicts new drug side effects without retraining.

Source

A network pharmacology method using a deep learning model to accurately predict drug-target interactions for network pharmacology analysis and drug discovery. The method involves training a deep learning model called DTI (Drug-Target Interaction) to predict drug-target interactions based on molecular structures. This addresses the incomplete data issue in traditional network pharmacology by improving prediction accuracy. The DTI model can be used in network pharmacology analysis to enhance accuracy and prediction power.

Source

Predicting GPCR drugs and targeted pathways using a deep belief network model based on a heterogeneous network of drug-pathway interactions. The method involves calculating similarities between drug substructures, disease phenotypes, and protein sequences to construct a drug-pathway network. A deep belief network is then used to predict potential drug-pathway relationships from the network. The model leverages drug-pathway similarities to predict interactions for unlisted GPCR drugs and target pathways.

Source

AI-based drug molecule screening method to efficiently find valuable drug candidates from large compound libraries. The method involves steps like preprocessing compounds, homology modeling, activity prediction, and molecular docking scoring using AI models. It leverages AI techniques like image encoding, text transformers, and neural networks to accurately represent molecules and proteins for screening. This allows efficient identification of target drug molecules for specific proteins by fusing their representations, predicting activity, and docking.

Source

Method for predicting drug interactions using structural information of drugs, as well as screening drug combinations with low probability of adverse interactions. The method involves calculating structural similarity profiles of drugs, inputting them into a trained model, predicting interactions using the model, and determining combinations with low interaction potential when no interaction is predicted. The interaction predictions are output as sentences describing the interaction mechanisms.

Source

Using artificial intelligence (AI) to efficiently discover novel drug compounds by leveraging protein dialects to generate sequences with desired activities. The AI model iteratively trains subsets of layers to identify sequences for each dialect's specific grammar. It generates shared portions of sequences that satisfy secondary objectives. Final layers receive primary objectives to determine remaining portions. Sharing subsets saves computation by avoiding recomputing shared portions for each dialect.

Source

Anticancer drug synergy prediction method using hypergraph neural networks to improve accuracy and efficiency of finding novel drug combinations for cancer treatment. The method involves extracting drug structures and cell line gene expressions, encoding them into node features, constructing a hypergraph with drug-drug-cell line triples, embedding the hypergraph nodes, and using a predictor to determine synergy. This leverages deep learning to extract high-level associations from the hypergraph structure, along with biochemical features, to predict synergy.

Source

Automatically designing molecules with high target specificity and high target selectivity for novel targets using machine learning models trained on molecular databases. The method involves generating molecules using a binding affinity model trained on a first molecular database and an embedding of a second molecular database, instead of just optimizing directly on the high-dimensional input space or a low-dimensional representation learned using a latent variable model. This allows accounting for context like protein sequences to improve target selectivity. The binding affinity model predicts target affinity and off-target selectivity. The generated molecules are then screened for properties like drug-likeliness and toxicity.

Source

Predicting medicinal effects of new compounds using deep learning by generating three types of feature data from medicinal substance data, training a neural network model using the features, and applying new compound data to the model to predict effects. The features are latent knowledge, molecular interactions, and chemical properties. The deep learning approach allows precise prediction of medicinal effects even when limited, heterogeneous, and incomplete information is available on the compounds.

Source

A system to provide personalized drug interaction and efficacy information to patients, caregivers, and healthcare professionals. The system uses a database of drug side effects, interactions, and management strategies. It allows users to input their medication lists and provides customized information on potential side effects, warnings, and recommendations for optimizing therapy. The system leverages knowledge mining, AI, and user contributions to improve drug safety and outcomes for people with comorbidities.

Source

Deep learning method for predicting drug interactions using molecular graph networks. The method involves training a model with two graph networks: an atomic-level network representing the internal structure of drug molecules, and a molecular-level network representing interactions between drugs. The model learns representations for both networks, then uses contrastive learning to compare atomic-level representations from the molecular-level network. This leverages the atomic-level structure for drug interaction prediction. The contrastive learning involves comparing the atomic-level representations of a fixed anchor drug molecule with its first-order neighbors vs non-first-order neighbors in the molecular-level network.

Source

Predicting the potential of new drugs using machine learning and reinforcement learning. The method involves predicting the binding force between a protein and a compound using learning from preprocessed compound and protein data. This predicted binding force is then used to determine the drug potential for a compound against a disease based on the protein involved. Reinforcement learning is also used to generate novel compounds by modifying existing ones and then predicting their drug potential.

Source

A method for predicting potential inhibitors of target proteins when no specific inhibitor dataset is available. The method involves leveraging similar proteins with known inhibitors to create a curated target-specific dataset. A transfer learning technique is applied to a pre-trained generative model using this dataset to generate target-specific molecules. Reinforcement learning is used to optimize the molecules' affinity for the target protein. This allows designing new inhibitors against novel targets with limited data.

Source

Artificial intelligence engine for generating candidate drugs using experimental validation and peptide drug optimization. The AI engine receives signals from proxy organisms after administering candidate drugs. It analyzes the signals to detect biomarkers indicating drug effectiveness. This allows preclinical validation of drug candidates before clinical trials. The AI engine uses neural networks for drug discovery, design, and testing in a larger design space than conventional methods. It enlarges the drug space using encoding techniques and reduces complexity with causal inference.

Source

Virtual screening of drugs based on receptors and ligands using a combination of receptor-based and ligand-based drug screening techniques to improve accuracy and efficiency. The method involves obtaining activity data for target receptors, ligands, and their molecular fingerprints. It molecularly docks the receptors with ligands, performs molecular dynamics simulations, and decomposes the binding energy. It then fuses the fingerprints and energy decomposition values to train a machine learning model for virtual drug screening. This complementary approach leverages the strengths of receptor-based and ligand-based methods to efficiently characterize active ligand sets and train a higher accuracy drug screening model.

Source

Method to minimize side effects of drugs by predicting interactions based on chemical structure. The method involves training a classification model using binary drug interaction tasks. The drugs are represented by their labels, adverse reactions, and disease associations. By separating drugs based on chemical structure and adverse effects, the model can predict interactions and identify potentially harmful combinations. This allows minimizing side effects when prescribing multiple drugs for chronic diseases.

Source

Method for predicting drug interactions using structural information of drugs, allowing prediction of interaction mechanisms and identifying drug combinations with lower probability of adverse interactions. The method involves calculating structural similarity profiles for two drugs, inputting them into a trained model, predicting interactions using the profiles, and interpreting the output sentences to determine interaction mechanisms. When no interaction is predicted, it indicates lower probability of adverse interactions.

Source

Artificial intelligence engine for generating candidate drugs using causal inference and machine learning techniques. The engine enlarges the design space for drug discovery by encoding drug information like structure, activity, and semantics. It uses machine learning models like GANs and RNNs to efficiently search the enlarged space. The engine also uses benchmarking and reinforcement learning to iteratively improve the drug generation process. The encoded drug representations are aggregated from databases and translated into the AI engine's format. The engine generates, tests, and refines candidate drugs based on the encoded representations.

Source

Predicting the affinity between a drug and a target using topological analysis of the drug-target complex. The method involves calculating the topological structure diagram to identify spatial characteristics of atoms in the complex. This diagram is then used to determine the affinity between the drug and target using a trained deep learning model. This allows accurate and reliable prediction of binding affinity by leveraging the complex topology instead of just molecular descriptors.

Source

Predicting the properties of a new compound affecting the nervous system using artificial intelligence and neuronal activity data. The method involves training an AI model on known compounds by analyzing their effects on neural activity. When a new compound's neural activity is inputted, the model predicts its properties. The AI learns the relationship between neural activity and compound properties by analyzing burst patterns in raster plots of neural activity recorded with microelectrodes.

Source