Use of a GNN to estimate synthesizability of compounds

Uses invertible flow-based model for generating unique molecules with desired properties. The model generates molecular graphs

Applies semantic constraints to variational autoencoders. Compared to grammar VAE and character VAE, these generate more valid and novel molecules.

Novel hydrogen-count based graph-based architecture for generating novel molecules tested on ChemBL, MOSES, and ZINC 250K datasets

Contains several models (LASSO regression, decision tree, and CNN-SMILES) to predict binding to an RNA hairpin target against tuberculosis

NN-based predictor of druglike small molecule lipophilicity, using transfer learning

Baysian Optimization of small molecules with synthesis planning

Needs to be retrained

Template-free retrosynthesis planner utilizing a transformer-based architecture

Prediction of the toxicity features across the Tox21 dataset using Binet (https://github.com/bioinf-jku/binet)

Drug target interaction

Drug target interaction

Prediction of ADRs

Drug Target Interaction

Drug Target Interaction

Needs training for each target protein

Predicting MIC for Klebisella from genomic data

Needs to be updated to PY3

Drug Drug interaction prediction

protein-specific ligand activity prediction

We can implement the search from smiles, but not the other way around from proteins

Predicts gene expression profiles given a chemical compound

Generative model

Chemical Structuere-based inference of biological alterations

Activity prediction EGFR Inhibitors

Drug-likeness measure

Toxicological classification of chemical compouds

Prediction of drug functions MeSH

Prediction of hERG toxicity

It is developed in py2.7, it would be good to test if it works in py3

Prediction of ADR using gene expression profiles

Uses Genomic data as input

Retrosynthetic pathway prediction using MACCS keys

Predicts the activity coefficient (measures the deviation of a mix of chemical substances from its ideal behaviour)

Conversion from IUPAC to Structure

Not interesting as the authors do not provide the Struct2IUPAC, which is the one more difficult. STOUT does show both capabilities

Predicts the enzymatic reaction of a small molecule

Optimizes molecules based on chemist intuition Transformer trained model

The prediction models for activity, ADME etc are not provided openly

A model that uses SHAP to explain metabolic stability

No pretrained models

Generation of CYP inhibitors

Prediction of hERG Cardiotoxicity

Checkpoints not provided not dataset, we can only redirect to their online server

This paper explores performance of non deep learning techniques for the task of chemical space exploration. It presents two algorithms, namely, a graph based genetic algorith, and a graph based generative model with monte carlo tree search for this task and positions them as strong competitors to previous RNN based approaches with respect to algorithm runtime.

Histone deacetylase 3 (HDAC3) Finder trained model help to screen(identify) for HDAC3 inhibitor in compounds. Histone deacetylase 3 (HDAC3) is a prospective drug target for the treatment of human diseases such as cancer.

Molecular Interaction Transformer (MolTrans) gives more accurate results as compared to other baselines by incorporating pattern mining algorithm and interaction modeling module which uses sub-structural pattern for more accurate and interpretable DTI prediction. Secondly, it incorporates an augmented transformer encoder to better extract and capture the semantic relations among substructures extracted from massive unlabeled biomedical data while existing methods focuses on limited labeled data and ignored massive unlabelled molecular data. MolTrans is evaluated on real world data and it showed improved DTI prediction performance compared to state-of-the-art baselines. MoITrans is basically a classfication model that determine whether a pair of drug and target protein will interact.

Atom mapping on valid reaction SMILES. Given a SMILES, the model returns the mapped reactions and confidence scores.

SumGNN is a trainer for Drug-Drug Interaction(DDI) prediction which incorporates knowledge summarization graph neural network an improvement from tradition KG(Knowledge Graph) network used in present base trainer. Models produced by SumGNN produced better pharmacological effect prediction score from other trainer by 5.57%. DDI prediction is critical in determining side effects of drugs in people with pre-existing conditions.pharmacological effect prediction score is used to measure the adversity of interaction of two or more drugs. The DDI model can be used to rule out some medicines for patients or suggest change in drug's chemical composition to suite a patient. This will accelerate drug discovery by offering faster and accurate reaction feedback

DeepDrug is a deep learning framework, using residual graph convolutional networks (RGCNs) and convolutional networks (CNNs) to learn the comprehensive structural and sequential representations of drugs and proteins in order to boost the drug-drug interactions(DDIs) and drug-target interactions(DTIs) prediction accuracy.

Terpenes are a wide range family of naturally occurring substances with different types of chemical and biological properties. Many of these molecules have already found use in pharmaceuticals. Characterisation of these wide range of molecules with classical approaches has proved to be a daunting task. This model provides more insight to identifying types of terpenes by using a natural product database, COCONUT to extract information about 60,000 terpenes. For clustering approach to this dataset, PCA, FastICA, Kernel PCA, t-SNE and UMAP were used as benchmark. For classification approach, Light gradient boosting machine, k-nearest neighbors, random forests, Gaussian naiive Bayes and Multilayer perceptron were used. The best performing algorithms yielded accuracy, F1 score, precision and other metrics all over 0.9. Input- Terpene features Output- Chemical subclass Programming Language- Python

More information of the model are provided in this pdf below: 2110.15047.pdf

Prediction of hERG Blocker

Drug pair scoring is a machine learning task that involves a set of drugs and the task of predicting the behavior of drug pairs. It is used for predicting the effectiveness and safety of drug combinations.

Multimodal, needs to be broken into pieces

The above Paper discusses the use of machine learning models to accurately and quickly predict molecular properties in drug discovery and material design. However, the vast chemical space and limited availability of property labels make supervised learning challenging. To address this, the authors present MoLFormer. The MOLFORMER's design is based to learn about a model trained on a small molecules which are represented as SMILES string. The Model architecture has an efficient linear attention mechanism and relative positional embeddings with the goal of learning a meaningful and compressed representation of chemical molecules.

Drug-Repurposing for COVID-19

Anatomical Therapeutic Chemical (ATC) classification for compounds/drugs plays an important role in drug development and basic research. However, previous methods depend on interactions extracted from STITCH dataset which may make it depend on lab experiments. ATC_CNN presents a pilot study to explore the possibility of conducting the ATC prediction solely based on the molecular structures. The motivation is to eliminate the reliance on the costly lab experiments so that the characteristics of a drug can be pre-assessed for better decision-making and effort-saving before the actual development

Framework for establishing antibiotic property predictions. It consists of four components: (1) molecular representation, (2) feature down-selection, (3) ML algorithm selection, and (4) molecular descriptor importance analysis

That paper works on exploring the interaction between different drugs which is called drug-drug interactions (DDIs), The model takes as an input a DDI tuple (Gx , Gy , r) and predicts the probability of a pair of drugs (Gx , Gy ) having an interaction r.

Needs re training

Zero-shot prediction of therapeutic use with geometric deep learning and clinician centered design

This paper implements and builds an anti-TB inhibitor prediction model

Very simple model, data not available

COATI: multi-modal contrastive pre-training for representing and traversing chemical space

BioGPT embeddings from biomedical text

This model is useful for the GRADIENT project

BioBERT embeddings from biomedical text

GlyLES: Grammar-based Parsing of Glycans from IUPAC-condensed to SMILES

Antimalarial prediction based on BERT

Lipophilicity

hERG inhibition

Toxicity endpoints

Transformer-based translation of SMILES text into embedding