Abstract
Ultra-large ‘building block’ catalogs provide inexpensive access to billions of synthesis-on-demand molecules, but the combinatorial scale renders conventional virtual screening impractical. We present Vector Virtual Screen (VVS), a score-function-agnostic machine learning framework for efficient navigation of combinatorial libraries and rapid identification of promising molecules for experimental validation. VVS comprises four key innovations: (i) the Embedding Decomposer, which factors molecules into building blocks in latent space; (ii) ChemRank, a correlation-based loss that improves retrieval precision; (iii) BBKNN, an algorithm for nearest-neighbor search directly in building block space; and (iv) a multi-scale hill-climbing algorithm for gradient-based navigation of molecular embedding vector databases. Across diverse scoring functions, VVS consistently outperforms existing methods in retrieving high-scoring molecules while evaluating only a fraction of the library, achieving orders-of-magnitude runtime improvements. By turning ultra-large libraries into tractable search spaces, VVS enables virtual screening to keep pace with the rapid expansion of chemical space and adapt seamlessly to future advances in scoring functions.
Introduction
Modern "synthesis-on-demand" catalogs, such as Enamine REAL, have expanded to include billions of molecules that can be reliably synthesized from a set of standardized building blocks. However, this combinatorial scale creates a massive computational bottleneck. Traditional virtual screening requires "docking" or scoring every single molecule in a library to find a hit. At the scale of $10^{10}$ or $10^{12}$ compounds, this exhaustive approach is physically and financially impossible.
Existing workarounds, like Active Learning or similarity-based retrieval, often still require full library enumeration or are restricted to specific, rigid rules. The Vector Virtual Screen (VVS) framework addresses this by performing "latent retrosynthesis"—decomposing molecular embeddings into building block embeddings—allowing researchers to navigate gigascale libraries without ever fully enumerating them.
Methods
VVS relies on four core technical innovations to make ultra-large libraries searchable on a single compute node:
- Embedding Decomposer: An MLP-based model that performs latent retrosynthesis. It takes the embedding of a full molecule (from a Chemical Language Model) and predicts the embeddings of the constituent building blocks.
- ChemRank Loss: A novel correlation-based loss function. Instead of forcing the model to match exact vector coordinates (like MSE), it optimizes the model to preserve the relative similarity between molecules, significantly improving retrieval precision.
- BBKNN (Building Block K-Nearest Neighbors): Instead of searching a database of billions of finished molecules, BBKNN searches a much smaller database of building blocks. It retrieves the top $k$ blocks for each predicted part, cross-combines them using SMARTS reaction rules, and re-ranks the resulting molecules.
- Multi-scale Hill-Climbing: VVS uses the scores of retrieved molecules to estimate a gradient in the embedding space, iteratively "walking" toward better-scoring regions of the latent space.
Key Insights and Results
A critical discovery during development was that correlation-based optimization (ChemRank) is far more effective for retrieval tasks than standard Mean Squared Error (MSE). By maximizing the Pearson correlation between predicted and target similarity profiles, the model better preserves the structure of the chemical manifold. Additionally, the authors found that sorting building blocks by SMILES string length during training provided a ~25% performance boost, likely by helping the model disambiguate building block roles in a reaction.
In benchmarking, VVS demonstrated significant advantages:
- Speed: VVS completed searches 1.4x to 17x faster than state-of-the-art methods like SyntheMol or RxnFlow.
- Efficiency: The system identified high-scoring molecules by evaluating only 50,000 candidates out of a possible 10 billion ($10^{10}$).
- Generalization: BBKNN recovered the exact query molecule in 97.5% of in-distribution cases and maintained high structural similarity even for out-of-distribution natural products.
Conclusion
VVS transforms ultra-large libraries into tractable search spaces by navigating the chemical manifold intelligently rather than exhaustively. While the system's performance is currently tied to the quality of the underlying Chemical Language Model and specific reaction types, it offers a scalable path forward for virtual screening. Future work will focus on integrating VVS into closed-loop "design-synthesis-test" cycles and extending the framework to handle multi-step chemical assembly.
Summary updated