How can I predict binding affinity without a target protein structure?

Author

Himani Tandon

If you don’t have a structure for your target protein, you might think accurate binding affinity prediction is not possible. After all, the most widely used methods in computational drug discovery, like free energy perturbation (FEP), require high-quality structures to work.

The good news is that you can still make reliable predictions. There are well-validated approaches that work without any structural information at all.

In addition, these methods also require much fewer computational resources, so they are significantly faster and cheaper to run, and have a much wider window of applicability, so they can be used to explore novel chemical space.

Why traditional ligand-based QSAR methods have fallen short

You might be sceptical about ligand-based QSAR approaches; many researchers are. We discussed why there has been a lack of confidence in traditional QSAR methods to predict binding affinity in our recent Bio-IT World article.

One factor is that these statistical models typically use inputs that bear little or no connection to the underlying physics of binding. They also often struggle to predict activities for structurally novel molecules that differ from the training set, which is precisely what we want to do in drug discovery.

QuanSA™: A physics-aware ligand-based alternative that overcomes traditional QSAR limitations

QuanSA (Quantitative Surface Field Analysis) was developed specifically to overcome these limitations. Rather than treating affinity prediction as a purely statistical problem, QuanSA builds virtual binding pockets directly from your ligand data. They represent the physical reality of protein-ligand interactions without requiring the actual protein structure.

What makes this possible is multiple-instance machine learning. QuanSA automatically aligns your ligands in 3D space, with the model evolving as it learns from the ligand poses. Unlike traditional QSAR, where alignment is fixed at the start, QuanSA dynamically identifies and refines optimal poses as its parameters develop, learning both structure and physical interactions simultaneously.

This enables QuanSA to account for what actually drives binding:

Steric complementarity
Hydrogen bonding geometry and energetics
Detailed electrostatic interactions
Ligand strain (the energetic cost when a molecule adopts its bound conformation)

By modelling these physical interactions rather than relying on statistical correlations alone, QuanSA can generalise to structurally diverse molecules, including scaffold hops, which is exactly what you need when exploring new chemical space.

What this means for you: Accurate predictions at a fraction of the computational cost

QuanSA delivers comparable accuracy to FEP whilst being roughly 1,000 times less computationally expensive. Once you’ve built a model (which takes hours, not weeks), scoring new compounds happens in seconds. The model has already learned the essential physics of binding from your training data, so evaluating how well a new molecule fits into the learned pocket field, accounting for shape, hydrogen bonding, electrostatics, and strain, is computationally straightforward.

In addition to being able to make accurate binding affinity predictions without waiting for a target structure, this enables you to:

Screen thousands of compounds in the time FEP would take to evaluate a handful.
Explore chemical space broadly in early-stage projects, not just make small modifications around known scaffolds.
Build and run models on standard hardware, without the need for specialised computing infrastructure.
Iterate quickly, through rebuilding models as new data comes in and quickly screening your next design ideas.

Proven performance in peer-reviewed projects

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

This project focused on identifying 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 and FEP+ affinity predictions.

Each individual method showed similar levels of high accuracy in predicting pKi. Furthermore, a hybrid model that averaged predictions from both approaches performed even better. QuanSA and FEP+ make uncorrelated errors, so this work shows how you can take advantage of both when structures are available

Visualise the interactions driving binding affinity

The QuanSA plugin for PyMOL makes these predictions more widely accessible. The plugin’s clear visualisations identify the key interactions that drive molecular affinity, providing essential insights to design even more potent compounds.

An example of the visual output provided by the QuanSA PyMOL plugin. m32 is approximately 50 times more potent than the structurally similar m01, despite having near identical patterns of hydrogen bonding (red and blue cones). The difference is explained by additional steric contributions indicated by the surface patches highlighted by the black arrows.

Getting started with QuanSA

The absence of a protein structure doesn’t have to stop your affinity prediction work. Whilst structure-based methods like FEP are powerful when you have high-quality structures, QuanSA provides a reliable alternative that’s fast enough for high-throughput screening, accurate enough to guide design decisions, and transparent enough to build confidence in predictions.

Want to try it out for yourself? QuanSA is part of our BioPharmics platform, which provides industry-leading 3D methods for structure and ligand-based design. Fill the form below to learn more.

Ready to trial BioPharmics?

Simply complete the form now for a free demo and see how BioPharmics fits your unique research needs.

Himani Tandon, PhD

Himani is a Principal Scientist at Optibrium, specialising in computational drug discovery. With a PhD in Bioinformatics and Computational Biology from the Indian Institute of Science, she has extensive experience in AI/ML techniques for modeling drug candidates, including small molecules and peptides. She is also involved in 3D ligand-based and structure-based drug design.

Before joining Optibrium, Himani completed a postdoctoral fellowship at the MRC Laboratory of Molecular Biology, focusing on cell reprogramming and transcriptomics.