A docking scoring function is the equation that turns a ligand pose in a binding site into a number. That number is used to pick the right pose, to rank compounds against each other, and sometimes to estimate binding affinity. Those are three different jobs, and no single function does all three well. The practical answer is to choose a function for the job at hand, check it against your own target, and combine more than one when the decision is expensive.
This guide covers the four families of scoring functions, what each is good at, where they break, and how an experienced computational team stacks them into a workflow. It is written for biotech and pharma R&D teams deciding how to run a screen, not for a first lecture on docking.
TL;DR
- Scoring functions fall into four families: physics-based, empirical, knowledge-based and machine learning-based. Each makes a different trade between speed, interpretability and accuracy.
- Pose prediction is the more mature problem. Ranking compounds by affinity, especially among close analogs in hit-to-lead, is where scoring functions struggle most.
- No universal scoring function exists. Validate on data relevant to your target before you trust any score.
- Consensus and multi-step (funnel) scoring usually beat a single function for enrichment.
- Benchmark performance can mislead, because dataset bias inflates retrospective results that do not carry into prospective screens.
Table of Contents
- What a scoring function actually does
- The four families of scoring functions
- Where scoring functions fail
- How experienced teams combine scoring functions
- Choosing a function for your project
- Our perspective
- Work with us on your screening campaign
- FAQ
- Sources
What a scoring function actually does
Docking software has two components. A search algorithm generates candidate poses of a ligand in the binding site. A scoring function evaluates each pose. The scoring function can be used to determine the binding mode and site of a ligand, predict binding affinity and help identify drug leads, and the reliability of docking depends heavily on its accuracy.
It helps to separate the tasks the field groups under "scoring":
- Docking power: can the function pick the native-like pose out of a set of generated decoy poses?
- Scoring power: does the score correlate with measured binding affinity?
- Ranking power: can it order a series of ligands for one target correctly?
- Screening power: can it separate true binders from non-binders in a large library?
A function can be excellent at the first and mediocre at the rest. Industry reviews have long made the same point: the scoring problem is distinct from the docking problem, and many algorithms place ligands plausibly yet fail to rank those poses by affinity. Keep that distinction in mind when you read any vendor claim or paper abstract. Ask which power was measured.
The four families of scoring functions
A widely cited overview of scoring functions for protein-ligand interactions sorts them into four types.
Physics-based
These calculate binding energy from force-field terms such as van der Waals and electrostatic interactions, sometimes with solvation corrections. Their strength is that each term has a physical meaning. If a ligand scores well, you can see whether electrostatics or packing is driving it. The cost is compute time and sensitivity to details such as protonation states, partial charges and how water is treated.
Empirical
Empirical functions sum weighted terms (hydrogen bonds, hydrophobic contacts, rotatable-bond penalties, and so on) with weights fitted to experimental binding data. They are fast, which is why many popular docking programs use them. One well-known example is GlideScore, which Schrödinger describes as an empirical function built to separate strong binders from compounds with little or no binding. A review of empirical scoring functions for structure-based virtual screening covers their applications and the critical aspects of using them.
The weakness is the training set. An empirical function is only as general as the complexes it was fitted to, so a target class far from that set can behave poorly.
Knowledge-based
Knowledge-based functions derive statistical potentials from how often atom or residue pairs appear at given distances in known structures. The idea is that frequent contacts are probably favorable. They are fast and capture effects that are hard to write as explicit physics, but they depend on the quality and coverage of the structural database they were built from.
Machine learning and deep learning-based
These learn non-linear relationships between interface features and affinity, and they are the fastest-moving family. They also carry the most debate. We return to that below, because it matters for any team deciding how much weight to give an AI-derived score.
| Family | Basis | Speed | Interpretability | Typical risk |
|---|---|---|---|---|
| Physics-based | Force-field terms | Slower | High | Sensitive to setup and charges |
| Empirical | Weighted terms fitted to affinity data | Fast | Moderate to high | Limited by training set |
| Knowledge-based | Statistics from known structures | Fast | Moderate | Database coverage and bias |
| Machine learning | Learned from data | Fast at inference | Low | Overfitting, dataset bias |
Where scoring functions fail
Affinity ranking in hit-to-lead
The core difficulty is accurate prediction of absolute binding affinity, along with the inability of many functions to consistently tell actives from inactives in prospective screens, as discussed in a 2021 study of fast scoring functions. During optimization the problem sharpens. A small structural change, such as moving a methyl group, can change potency sharply, while the score changes by less than the function's own error. A score difference of a fraction of a unit between two analogs should not drive a synthesis decision by itself.
Benchmark bias
Many benchmark sets carry hidden biases. Differences in molecular weight or topology between actives and decoys can let a function look strong in retrospective testing and then fail in a real prospective screen, a pattern the same JCIM analysis describes. A separate benchmark of four virtual screening programs found that how the active/decoy dataset is constructed remains a major determinant of measured performance. So a published enrichment number tells you about the function and the dataset together.
The black-box problem
With physics-based models you can inspect which term drives binding. Deep learning models are harder to read, and a critical review of AI for protein-ligand interactions argues that a lack of large, unbiased datasets means some models may be predicting from data artifacts rather than physical interactions. The same body of work reports that deep learning models often predict affinity in a narrow range around the mean of their training data. That is a poor property when you need to distinguish a 10 nM compound from a 1 µM one.
The picture is not one-sided. Another JCIM paper on simple structure-based methods warns that simple interaction fingerprints can overestimate performance when data are unbiased, and shows target-specific machine-learning models outperforming them. The honest takeaway is that the right choice depends on the target, the data you have, and whether you are screening or optimizing.
Recent AI docking methods
Recent benchmarks on time-split PDBBind data show newer AI-driven tools reaching pose recovery rates, often above 50% of predictions within 2 Å RMSD, as laid out in a 2024 Nature Communications benchmark table. The same body of work stresses that these methods remain sensitive to how datasets are built and to target-specific features. A good pose rate on a public benchmark does not guarantee the same on your kinase, protease or protein-protein interface.
How experienced teams combine scoring functions
The funnel
Because accuracy and speed trade off, workflows are staged. A fast, simpler function filters a library of millions of compounds. A slower, more accurate method then re-ranks the top tier. This funnel logic is described in the industry review noted above. The expensive calculation is spent only on compounds that survived cheap ones.
Consensus scoring
Consensus scoring applies several functions to the same poses, or combines several programs. When functions built on different principles, say one empirical and one knowledge-based, agree on a top-ranked compound, the chance it is a true binder is higher than when one function alone ranks it. One study combining AutoDock, DOCK and Vina reported individual pose-prediction success of 55 to 64%, rising above 82% with a consensus approach.
Consensus is not free. Functions that share the same terms and the same training data make correlated errors, so agreement between two near-duplicates adds little. Choose functions that differ in their underlying logic.
A practical sequence
- Prepare the target honestly. Check protonation states, waters, and whether the binding site conformation matches the ligand class. For flexible pockets, consider ensemble docking with a small set of pocket-focused conformers.
- Validate the protocol before screening. Redock known ligands, then test whether known actives are enriched over decoys. Our guide to validating docking protocol selection with EF1% checks walks through this step.
- Screen with a fast function. Keep the cut generous enough that you do not discard real binders.
- Re-score the survivors with a second, different function or a physics-based refinement.
- Inspect poses by eye for sensible interactions, strain and unsatisfied polar groups.
- Send a diverse subset to the lab, not only the top ten by score. Experimental results then tell you how well the workflow ranks on this target.
For the wider workflow around these steps, see our virtual screening drug discovery workflow guide, and for what to do when hit lists fill with artifacts, reducing false positives in virtual screening and validating virtual screening results.
Choosing a function for your project
Match the function to the decision you need to make.
| Your decision | What matters most | Suggested approach |
|---|---|---|
| Screen a large library for new chemotypes | Speed, screening power | Fast empirical or knowledge-based first pass, then re-scoring |
| Pick the binding mode for a known series | Docking power | Test several functions on redocking of your own crystal ligands |
| Rank close analogs in hit-to-lead | Ranking power | Treat scores as weak evidence; add refinement and consensus, and rely on assay data |
| Covalent or unusual chemistry | Correct modeling of the bond and geometry | Use a method designed for it; see our note on covalent docking considerations |
| Peptide ligands | Flexibility, large interfaces | Dedicated peptide workflow; see peptide binding affinity prediction |
Protein-protein docking has its own scoring landscape. A survey of scoring functions for protein docking models is a good starting point if your target is an interface rather than a small-molecule pocket.
A note for teams working with peptide therapeutics: the receptor biology matters as much as the score. If you want pharmacology background on how single, dual and triple receptor agonist peptides differ, this comparison of RT-3 against single and dual receptor agonists is one reference for framing the target question before you model it.
Our perspective
We get asked often which scoring function is "the best." We think the question hides the real one: best for which decision, on which target, with which data? Over a project, the single most useful habit we see is refusing to let one number carry a decision. A docking score is a hypothesis generator. It narrows millions of compounds to hundreds, and it is not a measurement.
We design projects so that each stage has a check. Before a screen, we test whether the chosen protocol recovers known actives on the specific target. During the screen, we favor combinations of functions that disagree in useful ways. After it, we want the first assay results back quickly, because experimental data on a handful of compounds teaches more about a scoring function's behavior on your target than any benchmark paper does.
We are also cautious with black-box scores in lead optimization. When a medicinal chemist asks why compound A outranks compound B, an answer rooted in interactions that can be inspected is far more useful than a number. That is why we pair learned models with interpretable analysis rather than replacing one with the other.
The last piece is communication. A scoring result with no stated uncertainty is easy to over-read, so we report what was validated, what was not, and how much confidence each ranking deserves.
— Hooman
Work with us on your screening campaign
If you are planning a structure-based or ligand-based campaign and want the scoring strategy built around your target rather than a default setting, our team supports projects from first consultation through delivery, with confidential workflows and regular updates. See our virtual screening and hit-to-lead services, or if your ligands are peptides, our peptide design service. For a broader introduction, what is virtual screening in drug discovery and our protein-ligand docking practical guide are good places to start, and how to build protein-ligand docking models that work covers model setup.
FAQ
What is the most accurate docking scoring function?
No universal winner exists. Recent benchmarking emphasizes that performance depends on the dataset and the target, so practitioners are advised to validate on data relevant to their own project. A function that excels at pose prediction may rank affinities poorly, so test the specific power you need.
Are machine learning scoring functions better than classical ones?
Sometimes, but not reliably. Learned models can improve pose recovery and, with target-specific training, screening performance. Their risks are dataset bias, overfitting and low interpretability, so check them on prospective or truly held-out data before relying on them in lead optimization.
Why do high docking scores fail in the lab?
Scoring functions approximate binding free energy and miss effects such as entropy, desolvation details and protein flexibility. Benchmark bias can also make a method look better than it is in real screens. Treat top scores as candidates for testing, and send a diverse set to the assay.
What is consensus scoring?
Consensus scoring combines the results of several scoring functions or docking programs on the same compounds. Agreement between functions built on different principles makes a true binder more likely than a high rank from one function alone. It works best when the functions make different kinds of errors.
Sources
- An Overview of Scoring Functions Used for Protein-Ligand Interactions in Molecular Docking
- True Accuracy of Fast Scoring Functions to Predict High-Throughput Screening Data from Docking Poses
- Empirical Scoring Functions for Structure-Based Virtual Screening
- Predicting or Pretending: Artificial Intelligence for Protein-Ligand Interactions
- Benchmark of four popular virtual screening programs
- Beware of Simple Methods for Structure-Based Virtual Screening
- Time-based PDBBind docking accuracy benchmark, Nature Communications
- Docking and virtual screening review, Nature Reviews Drug Discovery
- Consensus holistic virtual screening for drug discovery
- A comprehensive survey of scoring functions for protein docking models