MolScreen

MolScreen, developed by Molsoft L.L.C., is a comprehensive panel of over 2,500 high-quality 2D fingerprint and 3D pharmacophore models covering a broad range of pharmacology and toxicology targets. Designed for drug discovery professionals, it supports lead identification, target identification, compound profiling, and drug repurposing workflows. The models leverage MolSoft's proprietary 2D QSAR/Fingerprint and 3D Atomic Property Fields (APF) methods, spanning approximately 2,500 models across 1,200 targets. MolScreen can be run directly using MolSoft's ICM-Pro + VLS software, or compounds can be screened through MolSoft's contract research services.

MolScreen employs multiple machine learning and computational chemistry approaches — including kernel regression, neural networks, 3D-QSAR, and 4D docking — to deliver robust, validated predictions for both activity and ADMET properties.

Key Applications

• Target Identification: Search chemicals against a broad set of protein targets to identify likely binding partners.
• Lead Identification: Identify candidate chemicals that can bind to a specific protein target of interest.
• Profiling: Screen multiple chemicals against multiple protein targets simultaneously.
• Drug Repurposing: Discover new protein targets for existing approved drugs using the full model panel.

Available Model Categories

• All Targets — covering the full panel of approximately 2,500 models.
• Approved Drug Targets — focused on clinically relevant targets.
• Drug Toxicity Targets — models relevant to toxicological assessment.
• Endocrine Targets — models for endocrine system-related proteins.
• Non-Mammalian Targets — models extending beyond mammalian biology.

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

• Currently 999 mammalian models trained on ChEMBL Ki, IC50, and EC50 data.
• Median training set size of 370 ligands; median external test set AUC of 96% and Q2 of 0.5.
• Extremely fast screening — thousands of compounds processed in minutes.
• Training pipeline clusters actives by fingerprint, adds 40,000 ChEMBL decoys, applies kernel functions to generate probability scores (kcc/MolClass Score), and uses Partial Least Squares and Kernel Regression for activity prediction (kca/MolpKd Score).
• Final MolScore combines MolpKd and MolSimilarity to known binders.

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

• Currently 504 mammalian models built using Atomic Property Fields with Pocketome ligands or custom alignments as APF templates.
• Validated against ChEMBL compounds; median AUC of 92% (139 compounds vs. decoy).
• Relatively fast screening at approximately 5 seconds per compound per single template cluster.

Pocket Docking 3D-QSAR Models (dpc)

• Currently 343 mammalian models with AUC greater than 80%.
• Trained on ChEMBL Ki, IC50, EC50, and DrugBank assignments; median ligand set size of 307.
• Median external AUC of 95% and external Q2 of 0.53.
• Training uses Pocketome pocket clustering, 4D docking with co-crystallized ligand as APF template, and 3D-QSAR for activity prediction, combined into a final MolScore.

Hybrid 4D/2D Models (dfa)

• Currently 612 mammalian models with AUC greater than 80%.
• Trained on ChEMBL Ki, IC50, EC50, and DrugBank data; median ligand set size of 270.
• Median external AUC of 96% and external Q2 of 0.65 — among the highest-performing model types in the panel.
• Training combines 4D docking with ligand APF templates, 3D pose clustering, APF scoring, and 3D-QSAR per compound cluster, with a final MolScore integrating MolpKd and MolSimilarity.

Neural Network Chemical Fingerprint Classifier (ncc)

• Covers 6 target families, each containing 12 to 234 targets and 3,000 to 144,000 ligands.
• All models validated with 25% of data held out as an external test set; median external AUC of 99.5%.
• Architecture uses ECFP input, fully connected neural networks with 2–3 hidden layers, and multitask prediction across all targets in a family simultaneously.

ADMET Prediction Models (mcp)

• Currently 38 models, primarily derived from PubChem data, all validated on a 20% external test set.
• Regression models (mean external Q2: 0.7) cover CACO2, PAMPA permeability, LD50 (mg/kg), and half-life (hours).
• Classification models (median external AUC: 84%) include hERG, P-glycoprotein inhibitor/substrate, PAINS, Cytochrome P450 isoforms (1A2, 2C19, 2C9, 2D6, 3A4), and 25 Tox21 classifiers covering Estrogen Agonist/Antagonist, Genotoxicity, Aromatase, and more.
• Includes fully connected neural network models alongside regression and classification approaches.

MolScreen is accessible either as a standalone screening tool within MolSoft's ICM-Pro + VLS software environment or through MolSoft's contract research services, making it suitable for both in-house computational teams and organizations seeking outsourced screening support.