When a binding affinity prediction pipeline includes both a docking run and a machine learning scoring model, there is a temptation to treat the two numbers as independent votes to be averaged or combined by weighted sum. That works only if you understand what each is actually measuring, because the two approaches fail in completely different ways and agree for completely different reasons.
What docking is actually doing
Classical docking tools perform a conformational search within a receptor binding site. The algorithm generates a large number of candidate poses for the ligand, scores each pose with an empirical energy function that approximates non-bonded interactions, solvation penalties, and entropic costs, and returns the pose with the lowest estimated binding free energy.
What docking is measuring is binding geometry compatibility: does the molecule fit into the pocket in a physically plausible conformation, and does it form the expected interactions, hydrogen bonds, hydrophobic contacts, electrostatic complementarity, with the pocket residues?
The number coming out, typically expressed in kcal/mol, is not a directly calibrated binding affinity. It is a relative ranking metric that is most reliable when comparing structurally similar compounds against the same receptor structure. Docking scores are notably sensitive to the quality of the protein structure used: an apo crystal structure, a co-crystal structure from a different ligand class, and a homology model can produce substantially different rank orderings for the same compound set.
What ML scoring is doing differently
Graph neural network based scoring models and related QSAR-derived approaches are not running a physics simulation. They learn a mapping from molecular structure, represented as a molecular graph or fingerprint, to a predicted property value, typically measured biological activity from assay data.
These models have no explicit representation of the receptor. They learn, from statistical patterns in their training data, which structural features correlate with measured activity against the target class. A GNN trained on kinase inhibitor data has absorbed the SAR logic implicit in thousands of IC50 measurements: which hinge-binding motifs work, which gatekeeper interactions matter, which substituent patterns tend to reduce potency.
What ML scoring is measuring is binding likelihood from structural precedent. A high ML score for a new compound means: compounds that look like this, in the feature space the model understands, tend to be active. It does not specify how the compound binds or whether the binding mode matches precedent compounds.
Where each method fails
Docking fails when the binding mode is genuinely novel relative to the receptor structure used, when the binding site is highly flexible and not captured by a single static conformation, or when the ligand itself has conformational flexibility the search algorithm does not sample adequately. It also systematically struggles with charged groups, where the empirical energy function poorly approximates desolvation penalties at the interface.
ML scoring fails when the compound is structurally dissimilar from the training set, what is often called activity cliff prediction failure. A compound with an unusual scaffold, even one that docking suggests fits the pocket well, may score low by ML simply because the model has no structural precedent to draw from. ML also cannot reliably distinguish binding at the orthosteric site from allosteric activity, since it learns from overall activity data that conflates both contributions.
The failure modes are largely orthogonal. This is what makes them complementary rather than redundant, but it also means that disagreement between the two is informative, not just noise to be averaged away.
When they agree and when they disagree
High docking score plus high ML score is the most confident signal. The compound fits the pocket physically and resembles precedent active compounds. This combination is the strongest basis for synthesis prioritization in Alkira's ranked output.
High docking, low ML is the interesting case. The molecule may be binding in a way that the training data does not capture well, or it may have a structural novelty that the ML model penalizes but which is genuinely advantageous for selectivity. These compounds are worth a second structural look, not automatic deprioritization.
Low docking, high ML is a situation where the ML model is extrapolating from activity precedent but the compound is not geometrically fitting the pocket well with the available receptor structure. This can indicate a receptor structure quality issue, or a compound that would require an induced-fit binding mode the rigid docking did not find.
Low scores on both is generally a reliable signal to deprioritize: the two failure modes would need to coincide in an unlikely way to produce a false negative on both simultaneously.
The Tanimoto distance problem in ML prediction
One underappreciated aspect of ML scoring reliability is how rapidly performance degrades with structural distance from the training data. Tanimoto similarity on Morgan fingerprints is the standard measure of structural coverage: compounds with Tanimoto similarity below 0.4 to the nearest training compound are in genuinely extrapolated territory for most ML models.
This matters during scaffold hopping. When you move from a well-characterized scaffold to a novel chemotype, the ML binding score for the new scaffold may be based on training compounds that are structurally distant from your candidate. The confidence band widens, but the point estimate can still look plausible and mislead you into treating it as a reliable prediction.
The clearest signal that you are in low-Tanimoto territory is a combination of wide confidence band and a structural class that you know is sparsely represented in public bioactivity data. In those cases, treat the ML score as a weak prior and give more weight to the docking geometry compatibility, since docking is not limited by training data coverage in the same way.
How Alkira handles the combination
Rather than averaging docking and ML scores into a single number, the binding affinity output integrates them as separate evidence layers. The reported binding score reflects the ML prediction as the primary signal, with receptor geometry compatibility used to flag cases where the ML confidence and the docking compatibility are significantly misaligned.
When you see a high binding score in your Alkira output, it is primarily a statement about structural precedent and learned SAR. The confidence band on that score widens when the two signals diverge. A compound with strong ML prediction but poor docking compatibility shows a wider band, signaling that the prediction is less settled than the point estimate suggests.
The compounds in the middle of a confidence band spread, where ML and docking are partially disagreeing, are often the most interesting for a computational chemist to examine manually. They are not reliably good or bad candidates; they are the cases where the two methods are seeing something different in the same molecule, and that discrepancy is usually worth understanding before synthesis.