Build diverse, drug-like compound libraries for high-throughput screening campaigns.

Screening Decks

A well-designed screening library balances chemical diversity, drug-likeness, and synthetic accessibility to maximize the probability of identifying active compounds against a target. This workflow guides the assembly of screening decks from multiple sources including generative AI, vendor catalogs, and computational enumeration.

Overview

Screening deck quality directly impacts hit rates. Random or biased libraries waste screening resources on compounds unlikely to become drugs. This workflow uses GenMol for AI-generated novel scaffolds, REINVENT for targeted analog generation around known pharmacophores, and RDKit for property calculation, diversity analysis, and filtering. The result is a curated library optimized for the target class with verified drug-likeness and maximal structural diversity.

Tool Chain

GenMol -- AI-driven de novo molecular generation
REINVENT -- Reinforcement learning-based molecular design
RDKit -- Cheminformatics property calculation and filtering

Prerequisites

Defined target class or therapeutic area for the screen
Property criteria (molecular weight range, LogP limits, acceptable functional groups)
Desired library size (typically 1,000-100,000 compounds)
Diversity requirements (scaffold diversity, property space coverage)

Step-by-Step Guide

Define the library design criteria

Establish property ranges based on your target class and screening format. For a general drug discovery screen, apply Lipinski's Rule of Five: molecular weight below 500 Da, LogP below 5, hydrogen bond donors fewer than 5, hydrogen bond acceptors fewer than 10. For CNS targets, apply tighter constraints (MW below 400, LogP 1-3, PSA below 90). Document acceptable and excluded functional groups (exclude reactive warheads, PAINS substructures, and aggregator-prone compounds).

Generate novel scaffolds with GenMol

Use GenMol to generate 5,000-20,000 novel molecular scaffolds that satisfy your property criteria. Configure the generative model with diversity maximization to ensure broad coverage of chemical space. GenMol can produce entirely novel structures not found in existing vendor catalogs, expanding your screening beyond commercially available chemical space.

Generate targeted analogs with REINVENT

For known active scaffolds or pharmacophoric patterns relevant to your target, use REINVENT in scaffold-constrained mode to generate focused analog libraries. Set the scoring function to optimize diversity within a defined property window while maintaining the core pharmacophore. This generates a focused subset of compounds with higher expected hit rates against the specific target class.

Calculate and filter properties with RDKit

Process all generated compounds through RDKit for property calculation: molecular weight, LogP (Wildman-Crippen), hydrogen bond donors and acceptors, topological polar surface area, rotatable bond count, aromatic ring count, and fraction of sp3 carbons. Apply hard filters to remove compounds violating your design criteria and flag PAINS, Brenk, and other structural alerts.

Assess and maximize diversity

Calculate pairwise Tanimoto similarities using Morgan fingerprints (radius 2, 2048 bits) and cluster the library using Butina clustering or k-means. Remove redundant compounds (Tanimoto similarity above 0.85 to any other library member) to maximize diversity. Verify that the final library covers the target property space uniformly by inspecting molecular weight, LogP, and PSA distributions.

Validate purchasability

For compounds intended for physical screening, cross-reference with ZINC to identify which library members are commercially available or have close purchasable analogs. Partition the library into "buy" (available from vendors) and "make" (requires custom synthesis) subsets. For virtual screening, purchasability filtering is optional but useful for downstream hit follow-up.

Finalize and export the screening deck

Compile the final library with SMILES, calculated properties, diversity cluster assignments, purchasability status, and any structural alert flags. Export as SMILES file for docking, CSV for analysis, or SDF for structure-based tools. Document the library composition including scaffold diversity metrics, property distributions, and filtering statistics.

Expected Outputs

GenMol-generated novel scaffold library with property profiles
REINVENT-generated focused analog sets around known pharmacophores
RDKit-calculated property matrices with drug-likeness assessments
Diversity analysis report with clustering and similarity statistics
Purchasability annotations from a ZINC cross-reference
Final screening deck in SMILES, CSV, or SDF format

Tips and Best Practices

Aim for scaffold diversity, not just compound count. A library of 5,000 diverse scaffolds outperforms 50,000 compounds from three scaffolds for novel target classes.
Apply PAINS and structural alert filters aggressively. Promiscuous compounds generate false positives that waste experimental resources.
Include 5-10% of the library as known actives or tool compounds as positive controls for your screening campaign.
For phenotypic screens with no structural target information, prioritize broad diversity. For target-based screens with structural data, bias the library toward complementary shapes and pharmacophores.
Track the provenance of each library member (GenMol-generated, REINVENT-generated, vendor-sourced) to analyze which generation strategy produces the highest hit rates.
Update screening decks quarterly by adding newly generated compounds and removing those that have been tested across multiple campaigns without activity.

Screening Decks

On this page