| |

MolScreen


MolScreen Contains a Panel of >2500 High Quality 2D and 3D Models.

MolScreen is a set of high quality 2D fingerprint and 3D pharmacophore models for a broad range of pharmacology and toxicology targets. The models can be used for lead discovery or counter screening. The models use MolSoft's 2D QSAR/Fingerprint and 3D Atomic Property Fields ( Totrov 2008) methods. There are currently approximately 2500 models for 1200 targets.

The models can be screened directly using MolSoft's ICM-Pro + VLS software. Alternatively we can screen a set of chemicals for you via our contract research services. Please contact us for more information about how to use MolScreen.

MolScreen Applications

MolScreen can be used for:

Available Models

You can download and view the available models using the links below (updated (3/8/2019). Each model has a name which starts with the 3 letter abbreviation of the model type as described below followed by the gene name.

Model Types

There are two categories of models:

1. ADMET ( mcp)

2. Different types of Activity Models

About the Models

Machine Learning Models - Hybrid 2D QSAR/Fingerprint Models kcc(+kca)

kcc(+kca): Kernel regression Chemical fingerprint Classification/Activity prediction

  • Currently: 999 mammalian models
  • Training set: ChEMBL Ki, IC50, EC50
  • Report kcc(Classification) score and kca(pKd regression) score
  • Median training set: 370 ligands
  • Median external test set AUC: 96%
  • Median external test set Q2: 0.5
  • Extremely fast (thousands of cpds in min)

Training:

  • Cluster Actives by fingerprint
  • Add 40k ChEMBL actives decoy
  • Kernel function to each cluster -> probability score (kcc/MolClass Score)
  • Partial Least Square Regression for each cluster + Kernel Regression (kca/MolpKd Score)
  • MolScore: combine MolpKd and MolSimilarity to known binders

Ligand Field Docking Models - 3D Atomic Property Field Models (dfz)

dfz: Docking to ligand Field Z-score prediction model
  • Built using Atomic Property Fields
  • Currently: 504 mammalian models
  • Pocketome ligands/custom alignment as APF template
  • ChEMBL cpds for validation
  • Median AUC: 92%, 139 cpds vs decoy
  • Fast-ish (single template cluster ~5 sec per cpd)

Pocket Docking 3D QSAR Models

dpc: Docking to Pocket Classification/Activity
  • Currently: 343 mammalian models w/ AUC> 80%
  • Training set: ChEMBL Ki, IC50, EC50, Drugbank assignment
  • Median size: 307 ligands
  • Median external Q2: 0.53
  • Median external AUC: 95%

Training:

  • Pocketome -> Clustering of pocket residues
  • 4D Docking w/ co-crystallized ligand as APF template
  • Docking Score -> Probability score (dpc/MolClass score)
  • 3D QSAR training of Activity-> (dpa/MolpKd)
  • MolScore: combine MolpKd and MolSimilarity to known binders

Hybrid 4D/2D - Hybrid Models (dfa)

dfa: Docking to ligand Field Activity prediction
  • Currently: 612 mammalian models w/ AUC > 80%
  • Training set: ChEMBL Ki, IC50, EC50, Drugbank assignment
  • Median size: 270 ligands
  • Median external Q2: 0.65
  • Median external AUC: 96%

Training:

  • Also from Pocketome -> 4D Docking + Ligand APF template
  • Cpd align to ligand template -> cluster by 3D poses
  • APF Score -> Probability Score (dfc/MolClass score)
  • 3D-QSAR training for each cpd cluster (dfa/MolpKd score)
  • MolScore: combine MolpKd and MolSimilarity to known binders


ADMET Models

mcp: Miscellaneous Chemical Property Models