Combining Ligand- and Structure-Based Methods for More Effective Virtual Screening

In drug discovery, virtual screening is a fast and cost-effective way of narrowing down vast chemical libraries to identify the most promising hits, reducing synthesis and testing requirements while improving research efficiency.

Virtual screening serves two distinct purposes:

• Library enrichment, where very large numbers of diverse compounds are screened to identify a subset with a higher proportion of actives.
• Compound design, involving detailed analysis of smaller series to guide optimization. Here, the focus shifts to greater precision and ideally quantitative prediction of binding affinity.

Advances in computational power and data availability have enhanced the performance and efficiency of these methods, increasing their adoption and furthering their impact in discovery workflows.

Comparing virtual screening methods

Virtual screening methods fall broadly into two categories: ligand- and structure-based.

Ligand-based virtual screening doesn’t require a target protein structure, Instead, it leverages known active ligands to identify hits that show similar structural or pharmacophoric features.

These approaches offer faster and cheaper computation than structure-based methods, excelling at pattern recognition and generalization across diverse chemistries. They can be particularly valuable during the early stages of discovery for prioritizing larger chemical libraries and when no protein structure is available.

At the broadest scale, methods including infiniSee™ (by BioSolveIT) and exaScreen (by Pharmacelera) enable efficient screening of ultra-large synthetically accessible chemical spaces containing tens of billions of compounds. These technologies assess pharmacophoric similarities between library compounds and known active ligands, identifying the potential to form certain types of interactions, but trade off speed in exploring these vast spaces with sensitivity and precision.

For screening smaller libraries (up to thousands or even millions of compounds), other ligand-based methods focus on detailed conformational analysis of individual compounds. Though generally slower and more computationally expensive, they deliver more accurate results. These methods align known active ligands by superimposing 3D structures to maximize similarity across pharmacophoric features, such as shape, electrostatics, and hydrogen bonding interactions. This creates a binding hypothesis to quantify how well virtual compounds align. Traditional pharmacophore methods require users to specify alignment features, whilst modern approaches like eSim^TM (by Optibrium), ROCS®, (by OpenEye Scientific) and FieldAlign (by Cresset) automatically identify relevant similarity criteria with which to rank potentially active compounds.

Advanced methods like Quantitative Surface-field Analysis (QuanSA^TM),¹ (by Optibrium) take this approach further by constructing physically interpretable binding-site models based on ligand structure and affinity data using multiple-instance machine learning. Importantly, most 3D ligand-based methods only provide ranking scores for library enrichment, but 3D quantitative structure-activity relationship methods like QuanSA can predict both ligand binding pose and quantitative affinity, even across chemically diverse compounds. This provides more resolution in predictions with which to guide the efficient design of highly active compounds.

Structure-based virtual screening uses target protein structural information, typically obtained experimentally through X-ray crystallography or cryo-electron microscopy, or through computational methods such as homology modeling.

Structure-based methods provide insights into atomic-level interactions, including hydrogen bonds and hydrophobic contacts. They often provide better enrichment for virtual libraries by incorporating explicit information about the shape and volume of the binding pocket.

The most common approaches involve docking compounds into known binding pockets. Numerous docking methods excel at placing ligands into binding sites in reasonable orientations, but the challenge lies in scoring and ranking the poses. Typically, they cannot accurately predict binding affinities but can eliminate compounds that won’t fit into the binding pocket, thereby increasing the enrichment of high-ranked compounds.

Moving beyond docking, Free Energy Perturbation (FEP)¹ calculations represent state-of-the-art structure-based affinity prediction. Though accurate, they are computationally very demanding and typically limited to small structural modifications around known reference compounds.

The impact of AlphaFold

AlphaFold (by Google DeepMind) has significantly expanded the availability of protein structures. However, important quality considerations remain about their reliability in docking performance.

Models typically predict a single static conformation, potentially missing conformational differences associated with ligand binding. This can yield false negatives or inaccurate binding pose predictions.² Additionally, while AlphaFold’s backbone predictions are reliable, it can struggle with side chain positioning, which is critical to achieving good docking results. Without careful post-modeling refinement, AlphaFold has so far shown limited success in naïve docking experiments.

Co-folding methods that generate ligand-bound protein structures, like Boltz-2 (by MIT and Recursion) and AlphaFold3, have recently become available. Despite their promise, questions remain about their generalizability. Performance has been shown to falter when predicting structures that differ from the training set³ or when predicting allosteric binding sites.⁴ Though their utility has been limited, they can still serve as valuable ways to develop experimental hypotheses.

The power of a hybrid approach

Structure- and ligand-based approaches are highly complementary, often yielding more reliable results when used together.

Sequential integration first employs rapid ligand-based filtering of large compound libraries, followed by structure-based refinement of the most promising subset. For example, an initial ligand-based screen can identify novel scaffolds early, offering chemically diverse starting points that can then be analyzed through docking experiments to confirm binding interactions.

This approach conserves computationally expensive calculations to only a small set of compounds likely to succeed, increasing efficiency while improving precision over using a single method.

Parallel screening involves running both ligand- and structure-based screening independently but simultaneously on the same compound library. Each method generates its own ranking of compounds, and results can be compared or combined using consensus scoring frameworks.

• Parallel scoring selects top candidates from both approaches without requiring consensus. This increases the likelihood of recovering potential actives and helps mitigate the limitations inherent in each approach.
• Hybrid (consensus) scoring creates a single unified ranking through multiplicative or averaging strategies. By favoring compounds ranking highly across both methods, this approach reduces the number of candidates while increasing confidence in selecting true positives.

The choice between strategies depends on your objectives: use parallel approaches for broader hit identification and preventing any missed opportunities when you can afford to test more compounds, and consensus methods when you need higher confidence in your selections.

Case study

In collaboration with Bristol Myers Squibb, we found improved affinity prediction in LFA-1 inhibitor lead optimization.

Compounds in this work¹ were generated to identify orally available small molecules targeting the LFA-1/ICAM-1 interaction, which modulates immune responses. Structure-activity data from these compounds were split into chronological training and test datasets for QuanSA (ligand-based) and FEP+ (structure-based), (by Schrödinger) affinity predictions.

Each individual method showed similar levels of high accuracy in predicting pKi. However, the hybrid model averaging predictions from both approaches performed better than either method alone. Through partial cancellation of errors, the mean unsigned error (MUE) dropped significantly, achieving high correlation between experimental and predicted affinities.

Progressing virtual screening results

Binding affinity alone does not make a promising therapeutic candidate. Multi-parameter optimization (MPO)⁵ helps to prioritize hits from virtual screening by identifying compounds with the best overall drug-like properties and the highest probability of clinical success. MPO methods incorporate multiple objectives, including potency, selectivity, ADME, and safety profiles.

Summary

Your virtual screening strategy will depend on your success criteria, the computational time and resources you are willing to invest, and the available data. Ligand-based methods provide a faster and less costly alternative, valuable for filtering very large, chemically diverse libraries, or when structural data is limited.  Structure-based approaches work when high-quality protein structures are available. They often provide better library enrichment but are more computationally expensive.

Evidence strongly supports hybrid approaches that combine atomic-level insights from structure-based methods with pattern recognition capabilities of ligand-based approaches. Whether through sequential workflows or parallel consensus scoring, integrated strategies can outperform individual methods by reducing prediction errors and increasing hit identification confidence.

Tamsin Mansley, PhD, is the president of Optibrium Inc. and global head of application science.

References

1. Cleves, A. E., Johnson, S. R. & Jain, A. N. Synergy and Complementarity between Focused Machine Learning and Physics-Based Simulation in Affinity Prediction. J. Chem. Inf. Model. 2021; 61, 5948–5966.
2. Scardino, V., Di Filippo, J. I. & Cavasotto, C. N. How good are AlphaFold models for docking-based virtual screening? iScience. 2023; 26, 105920.
3. Škrinjar, P., Eberhardt, J., Durairaj, J. & Schwede, T. Have protein-ligand co-folding methods moved beyond memorisation? bioRxiv preprint 2025;https://doi.org/10.1101/2025.02.03.636309.
4. Nittinger, E., Yoluk, Ö., Tibo, A., Olanders, G. & Tyrchan, C. Co-folding, the future of docking – prediction of allosteric and orthosteric ligands. Artif. Intell. Life Sci. 2025; 8, 100136.
5. Segall, M. D. Multi-Parameter Optimization: Identifying High Quality Compounds with a Balance of Properties. Curr. Pharm. Des. 2012;18, 1292–1310.