Prediction of antimicrobial class specificity using simple machine learning methods applied to an antimicrobial knowledge graph. The knowledge graph is built on ChEMBL, Co-ADD and SPARK. Endpoints are broad terms such as activity against gram-positive or gram-negative bacteria. The best model according to the authors is a Random Forest with MHFP6 fingerprints.
This tool performs PCA, UMAP and tSNE projections taking the Coconut natural products database as a chemical space of reference. The Ersilia Compound Embeddings are used as descriptors. Four PCA components and two UMAP and tSNE components are returned.
This tool performs PCA, UMAP and tSNE projections taking a 100k ChemDiv diversity set as a chemical space of reference. The Ersilia Compound Embeddings are used as descriptors. Four PCA components and two UMAP and tSNE components are returned.
This tool performs PCA, UMAP and tSNE projections taking the DrugBank chemical space as a reference. The Ersilia Compound Embeddings are used as descriptors. Four PCA components and two UMAP and tSNE components are returned.
A set of three binary classifiers (random forest, gradient boosting classifier, and logistic regression) to predict the Blood-Brain Barrier (BBB) permeability of small organic compounds. The best models were applied to natural products of marine origin, able to inhibit kinases associated with neurodegenerative disorders. The training set size was around 300 compounds.
This model evaluates drug-likeness using an unsupervised learning approach, eliminating the need for labeled data and avoiding biases from incomplete negative sets. It extracts features directly from known drug molecules, identifying common characteristics through a recurrent neural network (RNN) language model. By representing molecules as SMILES strings, the model learns the probability distribution of known drugs and assesses new molecules based on their likelihood of appearing in this space.
Uni-Mol offers a simple and effective SE(3) equivariant transformer architecture for pre-training molecular representations that capture 3D information. The model is trained on >200M conformations. The current model outputs a representation embedding.
QupKake is an innovative approach that combines graph neural network (GNN) models with semiempirical quantum mechanical (QM) features to forecast the micro-pKa values of organic molecules. QM has a significant role in both identifying reaction sites and predicting micro-pKa values. Precisely predicting micro-pKa values is vital for comprehending and adjusting the acidity and basicity of organic compounds, This has significant applications in drug discovery, materials science, and environmental c
MProPred predicts the efficacy of compounds against the main protease of SARS-CoV-2, which is a promising drug target since it processes polyproteins of SARS-CoV-2. This model uses PaDEL-Descriptor to calculate molecular descriptors of compounds. It is based on a dataset of 758 compounds that have inhibition efficacy against the Main Protease, as published in peer-reviewed journals between January, 2020 and August, 2021. Input compounds are compared to compounds in the dataset to measure molecul
Prediction of drug-induced cardiotoxicity as a binary classification of cardiotoxicity risk. The probability score depicts risk of the compound being cardiotoxic. Classification is based on the chemical data such as SMILES representations of compounds and a variety of descriptors such as Morgan fingerprints and Mordred physicochemical descriptors that describe the molecular structure of the drug interactions. Biological data is also used including gene expression and cellular paintings after dru
Pharmacokinetics Profiler (PhaKinPro) predicts the pharmacokinetic (PK) properties of drug candidates. It has been built using a manually curated database of 10.000 compounds with information for 12 PK endpoints. Each model provides a multi-classifier output for a single endpoint, along with a confidence estimate of the prediction and whether the query molecule is within the applicability domain of the model.
Search Q-Mug based on WHALES descriptors. Q-Mug is a subset of 600k bioactive molecules from ChEMBL. Three conformers are given for each molecule. WhALES is a simple descriptor useful for scaffold hopping.
REINVENT 4 LibInvent creates new molecules by appending R groups to a given input. If the input SMILES string contains specified attachment points, it is directly processed by LibInvent to generate new molecules. If no attachment points given, the model try to find potential attachment points, and iterates through different combinations of these points. It passes each combination to LibInvent to generate new molecules.
Building on the Chemical Checker bioactivity signatures (available as eos4u6p), the authors use the relation between stereoisomers and bioactivity of over 1M compounds to train stereochemically-aware signaturizers that better describe small molecule bioactivity properties. In this implementation we provide the A1, A2, A3, B1, B4 and C3 signatures
Mol2MolScaffold uses REINVENT4's mol2mol scaffold prior and mol2mol scaffold generic prior to generate around 500 new molecules similar to the provided molecules. The generated molecules will be relatively similar to the input molecules.
The Extended Reduced Graph (ErG) approach uses the description of pharmacophore nodes to encode molecular properties, with the goal of correctly describing pharmacophoric properties, size and shape of molecules. It was benchmarked against Daylight fingerprints and outperformed them in 10 out of 11 cases. ErG descriptors are well suited for scaffold hopping approaches.
Scaled version of the WHALES descriptors (see eos3ae6). WHALES are holistic molecular descriptors useful for scaffold hopping, based on 3D structure to facilitate natural product featurization. The scaling uses sklearn's Robust Scaler trained on a random set of 100K molecules from ChEMBL.
The context discusses a novel notation system called Sequential Attachment-based Fragment Embedding (SAFE) that improves upon traditional molecular string representations like SMILES. SAFE reframes SMILES strings as an unordered sequence of interconnected fragment blocks while maintaining compatibility with existing SMILES parsers. This streamlines complex molecular design tasks by facilitating autoregressive generation under various constraints. The effectiveness of SAFE is demonstrated by trai
The DILI-Predictor predicts 10 features related to DILI toxicity including in-vivo and in-vitro and physicochemical parameters. It has been developed by the Broad Institute using the DILIst dataset (1020 compounds) from the FDA and achieved an accuracy balance of 70% on a test set of 255 compounds held out from the same dataset. The authors show how the model can correctly predict compounds that are not toxic in human despite being toxic in mice.
ADMET AI is a framework for carrying out fast batch predictions for ADMET properties. It is based on ensemble of five Chemprop-RDKit models and has been trained on 41 tasks from the ADMET group in Therapeutics Data Commons (v0.4.1). Out of these 41 tasks, there are 31 classification tasks and 10 regression tasks. In addition to that output also contains 8 physicochemical properties, namely, molecular weight, logP, hydrogen bond acceptors, hydrogen bond doners, Lipinski's Rule of 5, QED, stereo c
The Mol2MolMediumSimilarity leverages REINVENT4's mol2mol medium similarity prior to generate up to 100 unique molecules. The generated molecules will be relatively similar to the input molecule.
The authors tested the dataset of 39312 compounds used to train the antibiotics-ai model (eos18ie) against several cytotoxicity endpoints; human liver carcinoma cells (HepG2), human primary skeletal muscle cells (HSkMCs) and human lung fibroblast cells (IMR-90). Cellular viability was measured after 20133 days of treatment with each compound at 10 μM and activities were binarized using a 90% cell viability cut-off. 341 (8.5%), 490 (3.8%) and 447 (8.8%) compounds classified as cytotoxic for HepG2
This model has been trained on a publicly available collection of 5 datasets manually curated for drug-induced-liver-injury (DILI). DILI outcome has been binarised, and ECFP descriptors, together with physicochemical properties have been used to train a random forest classifier which achieves AUROC > 0.9
The authors use a mid-size dataset (more than 30k compounds) to train an explainable graph-based model to identify potential antibiotics with low cytotoxicity. The model uses a substructure-based approach to explore the chemical space. Using this method, they were able to screen 283 compounds and identify a candidate active against methicillin-resistant S. aureus (MRSA) and vancomycin-resistant enterococci.
The context discusses a novel notation system called Sequential Attachment-based Fragment Embedding (SAFE) that improves upon traditional molecular string representations like SMILES. SAFE reframes SMILES strings as an unordered sequence of interconnected fragment blocks while maintaining compatibility with existing SMILES parsers. This streamlines complex molecular design tasks by facilitating autoregressive generation under various constraints. The effectiveness of SAFE is demonstrated by trai
The authors curated a dataset of 282 compounds from ChEMBL, of which 160 (56.7%) were labeled as active N. gonorrhoeae inhibitor compounds. They used this dataset to build a naïve Bayesian model and used it to screen a commercial library. With this method, they identified and validated two hits. We have used the dataset to build a model using LazyQSAR with Ersilia Compound Embeddings as molecular descriptors. LazyQSAR is an AutoML Ersilia-developed library.
The model predicts the inhibitory potential of small molecules against Histone deacetylase 3 (HDAC3), a relevant human target for cancer, inflammation, neurodegenerative diseases and diabetes. The authors have used a dataset of 1098 compounds from ChEMBL and validated the model using the benchmark MUBD-HDAC3.
The authors use a two-step approach to build a model that accurately predicts the lipophilicity (LogP) of small molecules. First, they train the model on a large amount of low accuracy predicted LogP values and then they fine-tune the network using a small, accurate dataset of 244 druglike compounds. The model achieves an average root mean squared error of 0.988 and 0.715 against druglike molecules from Reaxys and PHYSPROP.
This model leverages the ChemProp network (D-MPNN) to build a predictor of hERG-mediated cardiotoxicity. The model has been trained using a published dataset which contains 7889 molecules with several cut-offs for hERG blocking activity. The authors select a 10 uM cut-off. This implementation of the model does not use any specific featurizer, though the authors suggest the moe206 descriptors (closed-source) improve performance even further.
Prediction of antimicrobial potential using a dataset of 29537 compounds screened against the antibiotic resistant pathogen Burkholderia cenocepacia. The model uses the Chemprop Direct Message Passing Neural Network (D-MPNN) abd has an AUC score of 0.823 for the test set. It has been used to virtually screen the FDA approved drugs as well as a collection of natural product list (>200k compounds) with hit rates of 26% and 12% respectively.
Based on a molecule's pharmacophore, this model generates new molecules de-novo to match the pharmacophore. Internally, pharmacophore hypotheses are generated for a given ligand. A graph neural network encodes spatially distributed chemical features and a transformer decoder generates molecules.
The Morgan Fingerprints are one of the most widely used molecular representations. They are circular representations (from an atom,search the atoms around with a radius n) and can have thousands of features. This implementation uses the RDKit package and is done with radius 3 and 2048 dimensions, providing a binary vector as output. For Morgan counts, see model eos5axz.
The pharmacophore mapper (pmapper) identifies common 3D pharmacophores of active compounds against a specific target and uniquely encodes them with hashes suitable for fast identification of identical pharmacophores. The obtained signatures are amenable for downstream ML tasks.
USRCAT is a real-time ultrafast molecular shape recognition with pharmacophoric constraints. It integrates atom type to the traditional Ultrafast Shape Recognition (USR) descriptor to improve the performance of shape-based virtual screening, being able to discriminate between compounds with similar shape but distinct pharmacophoric features.
Prediction of the activity of small molecules against Mycobacterium tuberculosis. This model has been developed by Ersilia thanks to the data provided by the Seattle Children's (Dr. Tanya Parish research group). In vitro activity against M. tuberculosis was measured i na single point inhibition assay (10000 molecules) and selected compounds (259) were assayed in MIC50 and MIC90 assays. Cut-offs have been determined according to the researcher's guidance.
Using self-supervised learning, the authors pre-trained a large model using one millon unlabelled molecules from ChEMBL. This model can subsequently be fine-tuned for various QSAR tasks. Here, we provide the encodings for the molecular structures using the pre-trained model, not the fine-tuned QSAR models.
MoLeR is a graph-based generative model that combines fragment-based and atom-by-atom generation of new molecules with scaffold-constrained optimization. It does not depend on generation history and therefore MoLeR is able to complete arbitrary scaffolds. The model has been trained on the GuacaMol dataset. Here we sample the 300k building blocks library from Enamine.
Small World is an index of chemical space containing more than 230B molecular substructures. Here we use the Small World API to post a query to the SmallWorld server. We sample 100 molecules within a distance of 10 specifically for the Wuxi map, not the entire SmallWorld domain. Please check other small-world models available in our hub.
Small World is an index of chemical space containing more than 230B molecular substructures. Here we use the Small World API to post a query to the SmallWorld server. We sample 100 molecules within a distance of 10 specifically for the ZINC map, not the entire SmallWorld domain. Please check other small-world models available in our hub.
Small World is an index of chemical space containing more than 230B molecular substructures. Here we use the Small World API to post a query to the SmallWorld server. We sample 100 molecules within a distance of 10 specifically for the Enamine REAL map, not the entire SmallWorld domain. Please check other small-world models available in our hub.
This model predicts the growth of the fungus M. mycetomatis, causal agent of Mycetoma, in presence of small drugs. It has been developed using the data from MycetOS, an opemn source initiative aiming at finding new patent-free drugs. The model has been trained using the LazyQSAR package (MorganBinaryClassifier) from Ersilia.
Prediction of the inhibition of the Human Histone Deacetylase 1 to revert HIV latency. The dataset is composed of all available pIC50 values from ChEMBL target 325, and the model has been developed using Ersilia's LazyQsar package (MorganBinaryClassifier)
A simple sampler of the ChEMBL database using their API. It looks for similar molecules to the input molecule and returns a list of 100 molecules by default. This model has been developed by Ersilia. It posts queries to an online server.
This model predicts the toxicity of small molecules in HepG2 cells. It has been developed by Ersilia thanks to data provided by MMV. We have used two cut-offs to define activity (5 and 10 uM respectively) with a dataset of 1335 molecules. 5-fold crossvalidation showed an AUROC of 0.8 and 0.77 respectively
Prediction of the in vitro antimalarial potential of small molecules. This model has been developed by Ersilia thanks to experimental data provided by MMV. The model provides the probability of inhibition of the malaria parasite (NF54) measured both as percentage of inhibition (with luminescence and LDH) and IC50. 5-fold crossvalidation of the models shows AUROC>0.75 in all models.
Prediction of the activity of small molecules against the schistosoma parasite. This model has been developed by Ersilia thanks to the data provided by the Swiss TPH. In vitro activity against newly transformed schistosoma (nts) and adult worms was measured (% of inhibition of activity and IC50, respectively)
This model is a Chemprop neural network trained with a growth inhibition dataset. Authors screened ~7,500 molecules for those that inhibited the growth of A. baumannii in vitro. They discovered abaucin, an antibacterial compound with narrow-spectrum activity against A. baumannii.
A simple sampler of the PubChem database using their API. It looks for similar molecules to the input molecule and returns a list of 100 molecules by default. This model has been developed by Ersilia and posts queries to an online server.
The STONED sampler uses small modifications to molecules represented as SELFIES to perform a search of the chemical space and generate new molecules. The use of string modifications in the SELFIES molecular representation bypasses the need for large amounts of data while maintaining a performance comparable to deep generative models.
The Smiles Pair Encoding method generates smiles substring tokens based on high-frequency token pairs from large chemical datasets. This method is well-suited for both QSAR activities as well as generative models. The model provided here has been pretrained using ChEMBL.
This model predicts the antimalarial potential of small molecules in vitro. We have collected the data available from the Open Source Malaria Series 4 molecules and used two cut-offs to define activity, 1 uM and 2.5 uM. The training has been done with the LazyQSAR package (Morgan Binary Classifier) and shows an AUROC >0.8 in a 5-fold cross-validation on 20% of the data held out as test. These models have been used to generate new series 4 candidates by Ersilia.
FasmiFra is a molecular generator based on (deep)SMILES fragments. The authors use Deep SMILES to ensure the generated molecules are syntactically valid, and by working on string operations they are able to obtain high performance (>340,000 molecule/s). Here, we use 100k compounds from ChEMBL to sample fragments. Only assembled molecules containing one of the fragments of the input molecule are retained.
Prediction of the antimalarial potential of small molecules using data from various chemical libraries that were screened against the asexual and sexual (gametocyte) stages of the parasite. Several compounds' molecular fingerprints were used to train machine learning models to recognize stage-specific active and inactive compounds.
Analysis of metabolic stability, determining the inhibition of CYP3A4 activity and whether the compounds are a substrate for the CYP3A$ enzyme. The data to build these models has been publicly available at PubChem (AID1645840, AID1645841, AID1645842) by ADME@NCATS.
Analysis of metabolic stability, determining the inhibition of CYP2D6 activity and whether the compounds are a substrate for the CYP2D6 enzyme. The data to build these models has been publicly available at PubChem (AID1645840, AID1645841, AID1645842) by ADME@NCATS
Analysis of metabolic stability, determining the inhibition of CYP2C9 activity and whether the compounds are a substrate for the CYP2C9 enzyme. The data to build these models has been publicly available at PubChem (AID1645840, AID1645841, AID1645842) by ADME@NCATS
Molecular representation of small molecules via ingerprint-based molecular maps (images). Typically, the goal is to use these images as inputs for an image-based deep learning model such as a convolutional neural network. The authors have demonstrated high performance of MolMap out-of-the-box with a broad range of tasks from MoleculeNet.
This panel of models provides predictions for the H3D virtual screening cascade. It leverages the Ersilia Compound Embedding and FLAML. The H3D virtual screening cascade contains models for Mycobacterium tuberculosis and Plasmodium falciparum IC50 predictions, as well as ADME, cytotoxicity and solubility assays
Bioactivity-aware chemical embeddings for small molecules. Using transfer learning, we have created a fast network that produces embeddings of 1024 features condensing physicochemical as well as bioactivity information The training of the network has been done using the FS-Mol and ChEMBL datasets, and Grover, Mordred and ECFP descriptors
ChemGPT (4.7M params) is a language-based transformer model for generative molecular modeling, which was pretrained on the PubChem10M dataset. Pre-trained ChemGPT models are also robust, self-supervised representation learners that generalize to previously unseen regions of chemical space and enable embedding-based nearest-neighbor search.
Electrotopological state (Estate) indices are numerical values computed for each atom in a molecule, and which encode information about both the topological environment of that atom and the electronic interactions due to all other atoms in the molecule
Parallel Artificial Membrane Permeability is an in vitro surrogate to determine the permeability of drugs across cellular membranes. PAMPA at pH 7.4 was experimentally determined in a dataset of 5,473 unique compounds by the NIH-NCATS. 50% of the dataset was used to train a classifier (SVM) to predict the permeability of new compounds, and validated on the remaining 50% of the data, rendering an AUC = 0.88. The Peff was converted to logarithmic, log Peff value lower than 2.0 were considered to h
Analysis of metabolic stability, determining the inhibition of CYP450 activity and whether the compounds are a substrate for the CYP450 enzymes. The data to build these models is publicly available at PubChem, AID1645840, AID1645841, AID1645842. The tested cyps include CYP2C9, CYP2D6 and CYP3A4.
The cytochrome bcc complex (QcrB) is a subunit of the mycobacterial cyt-bcc-aa3 oxidoreductase in the electron transport chain (ETC), and it has been suggested as a good M.tb target due to the bacteria's dependence on oxidative phosphorylation for its growth. The authors use a dataset of 352 molecules, of which 277 are classified as active (QIM < 1 uM), 58 as moderately active ( 1 > QIM < 20 uM) and 78 as inactive (QIM > 20). Qim refers to quantification of intracellular mycobacteria.
RXNFP uses a pre-trained BERT Language Model to transform a reaction represented as smiles into a fingerprint amenable for downstream applications. The authors show how the RXN-fps can be used to identify nearest neighbors on reaction datasets, or map the reaction space without knowing the reaction centers.
The Human Liver Microsomal assay takes into account the liver-mediated drug metabolism to assess the stability of a compound in the human body. The NIH-NCATS group took a proprietary dataset of 4300 compounds with its associated HLM (in vitro half-life; unstable ≤ 30 min, stable >30 min) and used it to train a classifier.
The model uses Word2Vec, a natural language processing technique to represent SMILES strings. The model was trained on over <2000 small molecules with associated experimental HepG2 cytotoxicity data (IC50) to classify compounds as HepG2 toxic (IC50 <= 30 uM) or non-toxic. Data was gathered from the public repository ChEMBL.
HobPre predicts the oral bioavailability of small molecules in humans. It has been trained using public data on ~1200 molecules (Falcón-Cano et al, 2020, complemented with other literature and ChEMBL compounds). The molecules were labeled according to two cut-offs: HOB > 20% and HOB > 50%, due to ongoing discussions as to which would be a more appropriate cut-off.
Predictor of several endpoints related to Sars-CoV-2. It provides predictions for Live Virus Infectivity, Viral Entry, Viral Replication, In Vitro Infectivity and Human Cell Toxicity using a combination of three models. Consensus results are obtained by averaging the prediction for the three different models for each activity and toxicity models. The models have been built using NCATS COVID19 data. Further details on result interpretations can be found here: https://drugcentral.org/Redial
The model uses Word2Vec, a natural language processing technique to represent SMILES strings. The model was trained on over <4000 small molecules with associated experimental HBV inhibition data (IC50) to classify compounds as HBV inhibitors (IC50 <= 1 uM) or non-inhibitors. Data was gathered from the public repository ChEMBL.
The human liver cytosol stability model is used for predicting the stability of a drug in the cytosol of human liver cells, which is beneficial for identifying potential drug candidates early during the drug discovery process. If a drug compound is quickly absorbed, it may not reach the intended target in the body or become toxic. On the other hand, if a drug compound is too stable, it could accumulate and cause detrimental effects. The authors use an NCATS dataset of 1450 compounds screened in
IDL-PPB aims to obtain the plasma protein binding (PPB) values of a compound. Based on an interpretable deep learning model and using the algorithm fingerprinting (AFP) this model predicts the binding affinity of the plasma protein with the compound.
Kinetic aqueous solubility (μg/mL) was experimentally determined using the same SOP in over 200 NCATS drug discovery projects. A final dataset of 11780 non-redundant molecules and their associated solubility was used to train a SVM classifier. Approximately half of the dataset has poor solubility (< 10 μg/mL), and two-thirds of these low soluble molecules report values of < 1 μg/mL. A subset of the data used is available at PubChem (AID 1645848).
Parallel Artificial Membrane Permeability is an in vitro surrogate to determine the permeability of drugs across cellular membranes. PAMPA at pH 5 was experimentally determined in a dataset of 5,473 unique compounds by the NIH-NCATS. 50% of the dataset was used to train a classifier (SVM) to predict the permeability of new compounds, and validated on the remaining 50% of the data, rendering an AUC = 0.88. The Peff was converted to logarithmic, log Peff value lower than 2.0 were considered to hav
ImageMol is a Representation Learning Framework that utilizes molecule images for encoding molecular inputs as machine readable vectors for downstream tasks such as bio-activity prediction, drug metabolism analysis, or drug toxicity prediction. The approach utilizes transfer learning, that is, pre-training the model on massive unlabeled datasets to help it in generalizing feature extraction and then fine tuning on specific tasks. This model is fine tuned on 10 GPCR assays with the largest number
Using the Datamol package, the model receives a SMILE as input, then goes through a process of sanitizing and standardization of the molecule to generate four outputs: Canonical SMILES, SELFIES, InChI and InChIKey
Representation Learning Framework that utilizes molecule images for encoding molecular inputs as machine readable vectors for downstream tasks such as bio-activity prediction, drug metabolism analysis, or drug toxicity prediction. The approach utilizes transfer learning, that is, pre-training the model on massive unlabeled datasets to help it in generalizing feature extraction and then fine tuning on specific tasks.
ImageMol is a Representation Learning Framework that utilizes molecule images for encoding molecular inputs as machine readable vectors for downstream tasks such as bio-activity prediction, drug metabolism analysis, or drug toxicity prediction. The approach utilizes transfer learning, that is, pre-training the model on massive unlabeled datasets to help it in generalizing feature extraction and then fine tuning on specific tasks. This model is fine tuned on 13 assays concerned with a number of t
This model has been developed using ImageMol, a deep learning model pretrained on 10 million unlabelled small molecules and fine-tuned in a second step to predict the binding of inhibitors to the human beta secretase 1 (BACE-1) protein. The BACE-1 dataset from MoleculeNet contains 1522 compounds with their associated pIC50. A compound with pIC50 => 7 is considered a BACE-1 inhibitor.
TThis model has been developed using ImageMol, a deep learning model pretrained on 10 million unlabelled small molecules and fine-tuned in a second step to predict the inhibition of the human immunodeficiency virus (HIV). The HIV dataset is from MoleculeNet and contains 43850 small molecules and their in vitro activity against HIV (CA - Confirmed active, CM - Confirmed moderately active, CI - Confirmed inactive). The classification was based on EC50 values and expert knowledge.
Hepatic metabolic stability is key to ensure the drug attains the desired concentration in the body. The Rat Liver Microsomal (RLM) stability is a good approximation of a compound’s stability in the human body, and NCATS has collected a proprietary dataset of 20216 compounds with its associated RLM (in vitro half-life; unstable ≤30 min, stable >30 min) and used it to train a classifier based on an ensemble of several ML approaches (random forest, deep neural networks, graph convolutional neural
Small molecules are represented by a variety of machine-readable strings (SMILES, InChi, SMARTS, among others). On the contrary, IUPAC (International Union of Pure and Applied Chemistry) names are devised for human readers. The authors trained a language translator model treating the SMILES and IUPAC as two different languages. 81 million SMILES were downloaded from PubChem and converted to SELFIES for model training. The corresponding IUPAC names for the 81 million SMILES were obtained with Che
DrugTax takes SMILES inputs and classifies the molecule according to their taxonomy, organic or inorganic kingdom and their subclasses, using a 0/1 binary classification for each one. It generates a vector of 163 features including the taxonomy classification and other key information such as number of carbons, nitrogens… These vectors can be used for subsequent molecular representation in chemoinformatic pipelines.
Small molecules are metabolized by the liver in what is known as phase I and phase II reactions. Those can lead to reduced drug efficacy and generation of toxic metabolites, causing serious side effects. This model predicts the human metabolites of small molecules using a molecular transformer pr-trained on general chemical reactions and fine tuned to human metabolism. It provides up to 10 metabolites for each input molecule.
CReM (chemically reasonable mutations) is a fragment-based generative model that takes as input a small molecule, breaks it down into fragments and iteratively replaces them with other fragments from a database. It has three implementations (MUTATE: arbitrarily replaces one fragment with another one); GROW (arbitrarily replaces an hydrogen with another fragment) and LINK (replaces hydrogen atoms in two molecules to link them with a fragment). Here, we use a MUTATE and GROWTH approach, which prov
MoLeR is a graph-based generative model that combines fragment-based and atom-by-atom generation of new molecules with scaffold-constrained optimization. It does not depend on generation history and therefore MoLeR is able to complete arbitrary scaffolds. The model has been trained on the GuacaMol dataset. Here we sample a fragment library from Enamine.
MolT5 (Molecular T5) is a self-supervised learning framework pretrained on unlabeled natural language text and molecule strings with two end goals: molecular captioning (given a molecule, generate its description) and text-based de novo molecular generation (given a description, propose a molecule that matches it). This implementation is focused on molecular captioning.
To define drug-likeness, a set of 2136 approved drugs from DrugBank was taken as drug-like, and three negative datasets were selected from ZINC15 (19M), the Network of Organic Chemistry (6M) and ligands from the Protein Data Bank (13k), respectively. The drug dataset was combined with an equal subsampling of the negative dataset for each experiment, using five different molecular representations (Mold2, RDKit, MCS, EXFP4, Mol2Vec). We have re-trained it following the author’s specifications.
This model uses a Bloom filter to query the ZINC20 database to identify if a molecule is purchasable. A bloom filter is a space-efficient probabilistic data structure to identify whether an element is in a given set. Due to the nature of bloom filters, false negatives are not possible (i.e if the model returns False, the molecule is not purchasable). As stated by the author, if the model returns True the molecule is purchasable with an error rate of 0.0003 (according to the ZINC20 catalog).
Drug function, defined as Medical Subject Headings (MeSH) “therapeutic use” is predicted based on the chemical structure. 6955 non-redundant molecules, pertaining to one of the twelve therapeutic use classes selected, were downloaded from PubChem and used to train a binary classifier. The model provides the probability that a molecule has one of the following therapeutic uses: antineoplastic, cardiovascular, central nervous system (CNS), anti-infective, gastrointestinal, anti-inflammatory, derma
ADMETLab2 is the improved version of ADMETLab, a suite of models for systematic evaluation of ADMET properties. ADMETLab2 provides predictions on 17 physicochemical properties, 13 medicinal chemistry properties, 23 ADME properties, 27 toxicity endpoints and 8 toxicophore rules. The code and training data are not released, using this model posts predictions to the ADMETLab2 online server. The Ersilia Model Hub also offers ADMETLab (v1) as a downloadable package for IP-sensitive queries.
Carcinogenicity is a result of several potential effects on cells. This model predicts the carcinogenic potential of a small molecule based on their potential to induce cellular proliferation, genomic instability, oxidative stress, anti-apoptotic responses and epigenetic alterations. Metabokiller uses the Chemical Checker signaturizer to featurize the molecules, and the Lime package to provide interpretable results. Using Metabokiller, the authors screened a panel of human metabolites and exper
Molecular representation of small molecules via descriptor-based molecular maps (images). The fingerprint-based molecular maps are available at eos59rr. These images can be used as inputs for an image-based deep learning model such as a convolutional neural network. The authors have demonstrated high performance of MolMap out-of-the-box with a broad range of tasks from MoleculeNet.
Prediction of the antimalarial potential of small molecules. This model is an ensemble of smaller QSAR models trained on proprietary data from various sources, up to a total of >7M compounds. The training sets belong to Evotec, Johns Hopkins, MRCT, MMV - St. Jude, AZ, GSK, and St. Jude Vendor Library. The code and training data are not released, using this model posts predictions to the MAIP online server. The Ersilia Model Hub also offers MAIP-surrogate as a downloadable package for IP-sensitiv
Given a molecule, this model looks for its 100 nearest neighbors in the ChEMBL database, according to ECFP4 Tanimoto similarity. Due to size constraints, the model redirects queries to the ChEMBL server, so when using this model predictions are posted online.
Given a molecule, this model for its 100 nearest neighbors, according to ECFP4 Tanimoto similarity, in the medicinal chemistry database ChEMBL17_DrugBank17_UNPD17. This combined database contains all the compounds from the three collections (DrugBank, ChEMBL22 and Universal natural product directory (UNPD)) with up to 17 heavy atoms. It features a total of 128k compounds. The whole ChEMBL17_DrugBank17_UNPD17 database is not downloaded with the model, by using it you post queries to an online ser
The model looks for 100 nearest neighbors of a given molecule, according to ECFP4 Tanimoto similarity, in the GDBMedChem database. GDBMedChem is a 10M molecule-sampling from GDB17, a database containing all the enumerated molecules of up to 17 atoms heavy atoms (166.4B molecules). GDBMedChem compounds have reduced complexity and better synthetic accessibility than GDB17 but retain high sp3 carbon fraction and natural product likeness, providing a database of diverse molecules for drug design. Th
The model looks for 100 nearest neighbors of a given molecule, according to ECFP4 Tanimoto similarity, in the GDBChEMBL database. GDBChEMBL is a 10M molecule-sampling from GDB17, a database containing all the enumerated molecules of up to 17 atoms heavy atoms (166.4B molecules). GDBChEMBL compounds were selected using a ChEMBL-likeness score, with the objective of having a collection with higher synthetic accessibility and high bioactivity while maintaining continuous coverage of the GDB17 chemi
This variational autoencoder (VAE) for chemistry uses an encoder-decoder-predictor framework to predict new small molecules. The input SMILES molecule is converted into a continuous vector, and the decoder converts this molecular representation back to a discrete SMILES. These continuous molecular representations allow for simple operations to generate new chemical matter. The decoder is constrained to produce valid molecules. In addition, a predictor estimates the chemical properties of the mol
The Fréchet ChemNet distance is a metric to evaluate generative models. It unifies, in a single score, whether the generated molecules are valid according to chemical and biological properties as well as their diversity from the training set. The score measures the Fréchet Inception Distance between molecules represented by ChemNet, a deep neural network trained to predict biological and chemical properties of small molecules.
BayeshERG is a predictor of small molecule-induced blockade of the hERG ion channel. To increase its predictive power, the authors pretrained a bayesian graph neural network with 300,000 molecules as a transfer learning exercise. The pretraining set was obtained from Du et al, 2015, and the fine tuning dataset is a collection of 14,322 molecules from public databases (8488 positives and 5834 negatives). The model was validated on external datasets and experimentally, from 12 selected compounds (
Utilizes a Weisfeiler-Lehman network (attentive mechanism) to predict the products of an organic reaction given the reactants. The model identifies the reaction centers (set of atoms/bonds that change from reactant to product) and obtains the products directly from a graph-based neural network.
DeepSMILES converts a SMILES string to a more accurate syntax for molecule representation, taking into account both the branches (closed parenthesis in the SMILES strings) and rings (using a single symbol at ring closure that also indicates ring size). This syntax is particularly suitable in generative models, when the output is a SMILES string. With DeepSMILES, scientists can train a network using this new syntax, generate new molecules represented as DeepSMILES and then decode them back to nor
A series of models for the systematic ADMET evaluation of drug candidate molecules. Models include blood-brain barrier penetration; inhibition and substrate affinity for CYP1A2, CYP2C9, CYP2C19, CYP2D6, CYP3A4, and pgp; F 20% and F 30% bioavailability; human intestinal absorption; Ames mutagenicity; skin sensitization; plasma protein binding; volume distribution; LD50 of acute toxicity; human hepatotoxicity; hERG blocking; clearance; half-life; Papp (caco-2 permeability); LogD distribution coeff
This model used a multitask deep neural network (DNN) to predict the probability that a molecule is a hERG blocker. It was trained using 7889 compounds with experimental data available (IC50). The checkpoints of the pretrained model were not available, therefore we re-trained the model using the same method but without mol2vec featuriztion. Molecule featurization was instead done with Morgan fingerprints. Six models were tested, with several thresholds for negative decoys (10, 20, 40, 60, 80 and
A tool for planning retrosynthesis of a target molecule based on template reactions and a stock of precursors. The algorithm breaks down the input molecule into purchasable blocks until it has been completely solved.
String representation of small molecules that is more robust than SMILES, since, by design, all SELFIES strings are valid molecules. It is particularly helpful when applied in generative models, as all the SELFIES proposed are valid molecules. The authors also found that on generative models, SELFIES produces more diverse molecules than compared to SMILES.
This model employs transfer learning with graph neural networks in order to predict micro-state pKa values of small molecules. The model enumerates the molecule's protonation states and predicts its pKa values. It was trained in two phases, first, using a large ChEMBL dataset and then fine-tuning the model for a small training set of molecules with available pKa values. The model in this repository is the pkasolver-light, which does not require an Epik license and is limited to monoprotic molecu
Prediction of the electronic spectra and excited state energy of small molecules. The training set is the QM8 from Molecule Net, where the electronic properties have been calculated by multiple quantum mechanic methods. This model has been trained using the GROVER transformer (see eos7w6n or grover-embedding for a detail of the molecular featurization step with GROVER)
The model predicts the atomization energy of a molecule. It has been trained using the QM7 dataset from MoleculeNet, a subset of GDB13 containing all molecules up to 23 atoms (7 heavy atoms + C, S, O, N). This dataset contains the computed atomization energy of 7165 molecules. This model has been trained using the GROVER transformer (see eos7w6n or grover-embedding for a detail of the molecular featurization step with GROVER)
Prediction of octanol/water distribution coefficient (logD at pH 7.4) trained using the Lipophilicity Molecule Net dataset. This model has been trained using the GROVER transformer (see eos7w6n or grover-embedding for a detail of the molecular featurization step with GROVER)
Prediction of water solubility data (log solubility in mols per litre) for common organic small molecules. trained using the Molecule Net ESOL dataset.
Model based on experimental and calculated hydration free energy of small molecules in water, the FreeSolv dataset from MoleculeNet. Hydration free energies are relevant to understand the binding interaction between a molecule (in solution) into its binding site. This model has been trained using the GROVER transformer (see eos7w6n or grover-embedding for a detail of the molecular featurization step with GROVER).
Prediction across the ToxCast toxicity panel, containing hundreds of toxicity outcomes, as part of the MoleculeNet benchmark. This model has been trained using the GROVER transformer (see eos7w6n or grover-embedding for a detail of the molecular featurization step with GROVER)
Prediction of Beta-secretase 1 (BACE-1) inhibition. BACE-1 is expressed mainly in neurons and has been involved in the development of Alzheimer's disease. This model has been trained on the BACE dataset from MoleculeNet using the GROVER transformer (see eos7w6n or grover-embedding for a detail of the molecular featurization step with GROVER).
Using the Molecule Net dataset ClinTox, the authors trained a classification model to predict the likelihood of failure in clinical trials due to toxicity. The dataset has been built using FDA approved drugs (non-toxic) and a set of drugs that have failed at advanced clinical trial stages. This model has been trained using the GROVER transformer (see eos7w6n or grover-embedding for a detail of the molecular featurization step with GROVER).
Predicts activity of compounds in the Tox21 toxicity panel, comprising of 12 toxicity pathways, as part of the MoleculeNet benchmark datasets. This model has been trained using the GROVER transformer (see eos7w6n or grover-embedding for a detail of the molecular featurization step with GROVER)
Estimation of synthetic accessibility score (SAScore) of drug-like molecules based on molecular complexity and fragment contributions. The fragment contributions are based on a 1M sample from PubChem and the molecular complexity is based on the presence/absence of non-standard structural features. It has been validated comparing the SAScore and the estimates of medicinal chemist experts for 40 molecules (r2 = 0.89). The SAScore has been contributed to the RDKit Package.
Identification of active molecules against Mycobacterium tuberculosis using an ensemble of data from ChEMBL25 (Target IDs 360, 2111188 and 2366634). The final model is a stacking model integrating four algorithms, including support vector machine, random forest, extreme gradient boosting and deep neural networks.
Toxicity prediction across the Tox21 panel from MoleculeNet, comprising 12 toxicity pathways. The model uses the Mean Teacher Semi-Supervised Learning (MT-SSL) approach to overcome the low number of data points experimentally annotated for toxicity tasks. For the MT-SSL, Tox21 (831 compounds and 12 different endpoints) was used as labeled data and a selection of 50K compounds from other MoleculeNet datasets was used as unlabeled data.
CoPriNet has been trained on 2D graph representations of small molecules with their associated price in the Mcule catalog. The predicted price provides a better overview of the compound availability than standard synthetic accessibility scores or retrosynthesis tools. The Mcule catalog is proprietary but the trained model as well as the test dataset (100K) are publicly available.
A deep neural network was trained to predict the LogP value of small molecules and fluorescent probes using an experimentally annotated dataset of >13k molecules (OPERA). This dataset was complemented with fluorescent probes to improve the model accuracy in this space. Probes predicted impermeant to cell membranes consistently showed experimental LogP <1.
Using Coarse Grained (CG) models, where several atoms are aggregated into a single bead, the authors obtain a set of 500,000 compounds with their simulated permeability across a single-component DOPC lipid bilayer. With this approach, the authors are able to cover a large and representative portion of the chemical space. We have used the data generated in this publication to train a simple regression model to predict compound permeability.
The authors provide a dataset of 200 small molecules and their experimentally measured permeability in a PAMPA assay. Using this data, we have trained a model that predicts the logarithm of the effective permeability coefficient.
The model uses a combination of two multilayer perceptron networks (baseline and auxiliar) and an autoencoder-like network to extract natural-product specific fingerprints that outperform traditional methods for molecular representation. The training sets correspond to the coconut database (NP) and the Zinc database (synthetic).
Prediction of the antimalarial potential of small molecules. This model was originally trained on proprietary data from various sources, up to a total of >7M compounds. The training sets belong to Evotec, Johns Hopkins, MRCT, MMV - St. Jude, AZ, GSK, and St. Jude Vendor Library. In this implementation, we have used a teacher-student approach to train a surrogate model based on ChEMBL data (2M molecules) to provide a lite downloadable version of the original MAIP
SYBA uses a fragment-based approach to classify whether a molecule is easy or hard to synthesize, and it can also be used to analyze the contribution of individual fragments to the total synthetic accessibility. The easy-to-synthesize dataset is an extract of the ZINC purchasable compounds, and the hard-to-synthesize dataset is generated using a Nonpher approach (introducing small molecular perturbations to transform molecules into more complex compounds). The fragments are calculated with ECFP8
A simple score to distinguish between natural products (-like) and synthetic compounds. The score was calculated using an analysis of the structural features that distinguish natural products (NP) from synthetic molecules. NP structures were obtained from the CRC Dictionary of Natural products and synthetic molecules belong to an in-house collection. This method has been contributed to the RDKit package, Ersilia is simply implementing the RDKit NP_Score.
The model is a derivation of the natural product fingerprint (eos6tg8). In addition to generating specific natural product fingerprints, the activation value of the neuron that predicts if a molecule is a natural product or not can be used as a NP-likeness score. The method outperforms the NP_Score implemented in RDKit.
Retrosynthetic accessibility score based on the computer aided synthesis planning tool AiZynthfinder. The authors have selected a ChEMBL subset of 200.000 molecules, and checked whether AiZinthFinder could identify a synthetic route or not. This data has been trained to create a classifier that computes 4500 times faster than the underlying AiZynthFinder. Molecules outside the applicability domain, such as the GBD database, need to be fine tuned to their use case.
Fast aqueous solubility prediction based on the Molecule Attention Transformer (MAT). The authors used AqSolDB to fine-tune the MAT network to solubility prediction, achieving competitive scores in the Second Challenge to Predict Aqueous Solubility (SC2).
By combining a Message-Passing Graph Neural Network (MPGNN) and a Forward fully connected Neural Network (FNN) with an integrated gradients explainable artificial intelligence (XAI) method, the authors developed MolGrad and tested it on a number of ADME predictive tasks. MolGrad incorporates explainable features to facilitate interpretation of the predictions. In this model, they train MolGrad with data from a Plasma-protein binding assay (PPB) to predict the fraction bound in plasma of small mo
By combining a Message-Passing Graph Neural Network (MPGNN) and a Forward fully connected Neural Network (FNN) with an integrated gradients explainable artificial intelligence (XAI) method, the authors developed MolGrad and tested it on a number of ADME predictive tasks. MolGrad incorporates explainable features to facilitate interpretation of the predictions.In this model, they train MolGrad with a dataset of hERG channel blockers/non-blockers to predict the cardiotoxicity of small molecules (I
By combining a Message-Passing Graph Neural Network (MPGNN) and a Forward fully connected Neural Network (FNN) with an integrated gradients explainable artificial intelligence (XAI) method, the authors developed MolGrad and tested it on a number of ADME predictive tasks. MolGrad incorporates explainable features to facilitate interpretation of the predictions. This model has been trained using experimental data on the permeability of molecules across Caco2 cell membranes (Papp, cm s-1)
A robust predictor for hERG channel blockade based on an ensemble of five deep learning models. The authors have collected a dataset from public sources, such as BindingDB and ChEMBL on hERG blockers and non-blockers. The cut-off for hERG blockade was set at IC50 < 10 uM for the classifier.
By combining a Message-Passing Graph Neural Network (MPGNN) and a Forward fully connected Neural Network (FNN) with an integrated gradients explainable artificial intelligence (XAI) method, the authors developed MolGrad and tested it on a number of ADME predictive tasks. MolGrad incorporates explainable features to facilitate interpretation of the predictions. This model has been trained using a ChEMBL dataset of CYP450 3A4 inhibitors (0) and non-inhibitors (1).
MycPermCheck predicts potential to permeate the Mycobacterium tuberculosis cell membrane based on physicochemical properties.
PaDEL is a commonly used molecular descriptor. It calculates 1875 molecular descriptors (1444 1D and 2D descriptors, 431 3D descriptors) and 12 types of fingerprints for small molecule representation. Originally developed in Java, here we provide PaDDELPy, its python implementation.
Molecular embedding based on natural language processing. It converts SMILES into fingerprints using an unsupervised model pre-trained on a very large SMILES dataset from ChEMBL. The transformer is particularly well-suited for low-data drug discovery.
A set of ca 1,800 chemical descriptors, including both RDKit and original modules. It is comparable to the well known PaDEL-Descriptors (see eos7asg), but has shorter calculation times and can process larger molecules.
Path-based fingerprints calculated with the RDKit package Chem.RDKFingerprint. It is inspired in the Daylight fingerprint. As explained in the RDKit Book, the fingerprinting algorithm identifies all subgraphs in the molecule within a particular range of sizes, hashes each subgraph to generate a raw bit ID, mods that raw bit ID to fit in the assigned fingerprint size, and then sets the corresponding bit.
Molecular representation using the BERT language Transformer. The model has been pre-trained on the GuacaMol dataset (~1.6M molecules from ChEMBL), and can be fine-tuned to the desired QSAR tasks. It has been benchmarked in MoleculeNet.
A set of 200 physicochemical descriptors available from the RDKIT, including molecular weight, solubility and druggability parameters. We have used the DescriptaStorus selection of RDKit descriptors for simplicity.
Avalon is a path-based substructure key fingerprint (1024 bits), developed for substructure screen-out when searching. It is part of the Avalon Chemoinformatics Toolkit and has also been implemented as an external RDKit tool.
The model is simply an implementation of the function Descriptors.MolWt of the chemoinformatics package RDKIT. It takes as input a small molecule (SMILES) and calculates its molecular weight in g/mol.
The Morgan Fingerprints, or extended connectivity fingerprints (ECFP4) are one of the most widely used molecular representations. They are circular representations (from an atom, search the atoms around with a radius n) and can have thousands of features. This implementation uses the RDKit package and is done with radius 3 and 2048 dimensions.
Weighted Holistic Atom Localization and Entity Shape (WHALES) is a descriptors based on 3D structure to facilitate natural product featurization. It is aimed at scaffold hopping exercises from natural products to synthetic compounds
GROVER is a self-supervised Graph Neural Network for molecular representation pretrained with 10 million unlabelled molecules from ChEMBL and ZINC15. The model provided has been pre-trained on 10 million molecules (GROVERlarge). GROVER has then been fine-tuned to predict several activities from the MoleculeNet benchmark, consistently outperforming other state-of-the-art methods for serveral benchmark datasets.
A set of 25 Chemical Checker bioactivity signatures (including 2D & 3D fingerprints, scaffold, binding, crystals, side effects, cell bioassays, etc) to capture properties of compounds beyond their structures. Each signature has a length of 128 dimensions. In total, there are 3200 dimensions. The signaturizer is periodically updated. We use the 2020-02 version of the signaturizer.
Low dimension continuous descriptor based on a neural machine translation model. This model has been trained by inputting a IUPAC molecular representation to obtain its SMILES. The intermediate continuous vector representation encoded by when reading the IUPAC name is a representation of the molecule, containing all the information to generate the output sequence (SMILES). This model has been pretrained on a large dataset combining ChEMBL and ZINC.
The model predicts the putative adverse drug reactions (ADR) of a molecule, using the SIDER database (MoleculeNet) that contains pairs of marketed drugs and their described ADRs. This model has been trained using the GROVER transformer (see eos7w6n or grover-embedding for a detail of the molecular featurization step with GROVER).
This model predicts the Blood-Brain Barrier (BBB) penetration potential of small molecules using as training data the curated MoleculeNet benchmark containing 2000 experimental data points. It has been trained using the GROVER transformer (see eos7w6n or grover-embedding for a detail of the molecular featurization step with GROVER).
This is a ligand-based target prediction model developed by the ChEMBL team. They trained the model using pairs of small molecules and their protein targets, and produced a multitask predictor. The thresholds of activity where determined by protein families (kinases: <= 30nM, GPCRs: <= 100nM, Nuclear Receptors: <= 100nM, Ion Channels: <= 10μM, Non-IDG Family Targets: <= 1μM). Here we provide the model trained on ChEMBL_28, which showed an accuracy of 85%.
The eToxPred tool has been developed to predict, on one hand, the synthetic accessibility (SA) score, or how easy it is to make the molecule in the laboratory, and, on the other hand, the toxicity (Tox) score, or the probability of the molecule of being toxic to humans. The authors trained and cross-validated both predictors on a large number of datasets, and demonstrated the method usefulness in building virtual custom libraries.
This model was developed to support the early efforts in the identification of novel drugs against SARS-CoV2. It predicts the probability that a small molecule inhibits SARS-3CLpro-mediated peptide cleavage. It was developed using a high-throughput screening against the 3CL protease of SARS-CoV1, as no data was yet available for the new virus (SARS-CoV2) causing the COVID-19 pandemic. It uses the ChemProp model.
Based on a simple E.coli growth inhibition assay, the authors trained a model capable of identifying antibiotic potential in compounds structurally divergent from conventional antibiotic drugs. One of the predicted active molecules, Halicin (SU3327), was experimentally validated in vitro and in vivo. Halicin is a drug under development as a treatment for diabetes.