Ultra-large virtual screening

1. Introduction & Background

In the early stages of drug discovery, the size and diversity of chemical libraries directly influence the quality and success rate of identifying hit compounds. With the recent expansion of make-on-demand chemical spaces (such as the Enamine REAL Space and ZINC databases) into tens of billions of molecules, conventional Structure-Based Virtual Screening (SBVS) has encountered severe computational bottlenecks. Evaluating multi-billion-compound libraries sequentially using traditional physics-based molecular docking is both computationally cost-prohibitive and time-consuming.

To overcome these challenges, recent research has leveraged AI-driven screening algorithms, active learning, and deep contrastive learning. These advancements have increased screening throughput by thousands to hundreds of thousands of times, pushing ULVS into a new era characterized by automation, flexible precision, and genome-wide coverage.

2. Core Technological Paradigms in ULVS

The three papers highlight three major technological paradigms driving the ULVS landscape:

A. Active Learning & Physics-Based Precision (OpenVS / RosettaVS)

Core Concept: Combines high-accuracy physics-based docking with active learning surrogate models.

Key Innovations:

RosettaGenFF-VS Force Field: Incorporates updated torsional potentials and entropy estimation models (Default and Simple Entropy) to improve the accurate ranking of different ligands binding to the same target.

Dual Screening Modes (VSX & VSH): VSX (Express) enables rapid initial screening with rigid protein side chains, while VSH (High Precision) models full side-chain and partial backbone receptor flexibility for refined final ranking.

OpenVS Platform: Employs a fingerprint-based feed-forward neural network active learning model. By docking only ~0.1% of the total library across iterative rounds, it rapidly converges to isolate top binders.

B. Multidimensional Grid Partitioning & Cloud-Based Scaling (AdaptiveFlow)

Core Concept: Utilizes chemical space partitioning and target-guided prescreening to drastically reduce the required search space.

Key Innovations:

18-Dimensional Property Grid: Maps the Enamine REAL Space (69 billion compounds) into an 18-dimensional grid based on 28 molecular properties (e.g., molecular weight, logP, hydrogen bond donors/acceptors).

Adaptive Target-Guided Screening (ATG-VS): Prescreens only 1–10 representative molecules per grid cell (tranche) to construct a target-specific activity heat map. Machine learning models then identify and prioritize active tranches, achieving up to a 10,000-fold speedup.

Linear Cloud Scaling: Features near-perfect linear scaling up to 5.6 million CPU cores, allowing ultra-large screenings to be conducted routinely and cost-effectively.

C. Deep Contrastive Learning & Dense Retrieval (DrugCLIP)

Core Concept: Formulates virtual screening as a dense retrieval task rather than affinity regression.

Key Innovations:

Contrastive Learning Framework: Align representations from separate molecule and pocket encoders into a shared vector embedding space, using cosine similarity to score binding probabilities.

ProFSA & GenPack Strategies: Uses Protein Fragment-Surrounding Alignment (ProFSA) for pretraining on synthetic data. For AlphaFold2-predicted targets with imprecise side chains, the GenPack strategy uses generative AI to induce pocket conformations into realistic bound states.

Ultra-Fast Efficiency: Screens 100,000 times faster than traditional docking and 100 times faster than current machine learning models, enabling genome-wide screening campaigns in hours.

3. Real-World Applications and Case Studies

Case Study 1: E3 Ubiquitin Ligase KLHDC2 (OpenVS / RosettaVS)

Target Background: KLHDC2 is an important substrate receptor subunit for targeted protein degradation (PROTACs), but lacked known small-molecule binders.

Screening Campaign: OpenVS was applied to screen the 5.5-billion-compound Enamine REAL library using active learning, followed by VSH flexible re-docking of the top 50,000 molecules.

Experimental Validation:

Initial Hits: Out of 29 synthesized compounds, C29 was identified as a hit with single-digit micromolar affinity (IC50=2.9 μM).

Focused Expansion: Substructure searching in ZINC22 (4.1 billion compounds) yielded 6 additional hits, including C2.8 (IC50=1.1 μM).

Structural Confirmation: A 2.0 Å X-ray crystal structure (PDB: 8UXS) confirmed that the experimental binding pose aligned closely with the predicted docking pose.

Case Study 2: Human NaV1.7 Channel Antagonists (OpenVS / RosettaVS)

Target Background: Voltage-gated sodium channel NaV1.7 (VSD4) is a key target for pain therapeutics.

Screening Campaign: Screened ~4.1 billion compounds from ZINC22, filtering out known arylsulfonamide motifs to discover novel chemotypes.

Experimental Validation:

4 out of 9 synthesized compounds exhibited IC50<10 μM (44.4% hit rate).

Lead compound Z8739902234 achieved an IC50 of 1.33 μM with inactivated-state dependence and selectivity over NaV1.5 and hERG channels.

Case Study 3: Oncology Targets PARP1 and FSP1 (AdaptiveFlow)

Target Background: Tested on PARP1 (established benchmark) and FSP1 (ferroptosis suppressor protein 1, an emerging cancer target with a complex co-factor binding site).

Screening Campaign: Screened against the 69-billion-molecule Enamine REAL Space grid.

Experimental Validation: Identified potent inhibitors with binding affinities comparable to pharmaceutically relevant benchmark drugs for both targets.

Case Study 4: Psychiatric Targets 5-HT2AR and NET (DrugCLIP)

5-HT2AR (Serotonin Receptor 2A):

Screened 1.65 million compounds from ChemDiv. Identified V008-4481 with a binding affinity of 21.0 nM (EC50=60.3 nM in β-arrestin2 recruitment assays).

NET (Norepinephrine Transporter):

Evaluated under a zero-shot setting on a newly solved structure not present in the training set.

Achieved a 15% hit rate (15 out of 104 compounds showed >60% inhibition). Compounds 0086-0043 (IC50=1.14 μM) and Y510-9709 (IC50=0.31 μM) outperformed the benchmark antidepressant bupropion (IC50=1.5 μM) and were validated via Cryo-EM.

4. Genome-Wide Scale Virtual Screening Paradigm

DrugCLIP demonstrated the feasibility of performing virtual screening across the druggable human genome:

Target Coverage: Covered ~10,000 AlphaFold2-predicted human proteins (~20,000 pockets), expanding available target coverage compared to traditional databases like ChEMBL (~4,810 targets).

Computational Scale: Computed binding scores across 500 million molecules (ZINC and Enamine REAL), completing over 10 trillion scoring operations.

Execution Speed: The entire genome-wide screening run was executed in just 24 hours using a single computing node with 8 Nvidia A100 GPUs.

Open Data Resource: Created an open-access repository containing over 2 million high-confidence candidate complexes and docking poses (drug-the-whole-genome.yanyanlan.com).

5. Comparative Overview of the Three Approaches

References & Formal Citations

Zhou, G., Rusnac, D. V., Park, H., Canzani, D., Nguyen, H. M., Stewart, L., Bush, M. F., Nguyen, P. T., Wulff, H., Yarov-Yarovoy, V., Zheng, N., & DiMaio, F. (2024). An artificial intelligence accelerated virtual screening platform for drug discovery. Nature Communications, 15(1), 7761. https://doi.org/10.1038/s41467-024-52061-7

Cecchini, D., Gorgulla, C., et al. (2026). AI-enhanced adaptive virtual screening of large libraries for ligand discovery. Nature Biotechnology. https://doi.org/10.1038/s41587-026-03217-x (Reported via St. Jude Children's Research Hospital / Phys.org).

Jia, Y., Gao, B., Tan, J., Hong, X., Zhu, W., Tan, H., Xiao, Y., Huang, Y., Jin, Y., Yuan, Y., Tian, J., Ma, W., Zhang, Y., Yan, C., Zhang, W., & Lan, Y. (2024). Deep contrastive learning enables genome-wide virtual screening. bioRxiv preprint. https://doi.org/10.1101/2024.09.02.610777