Beyond the Ligand Applicability Domain: Generalization and Local Resolution in Structure-Aware Foundation Models for Affinity Prediction
6 affiliations
- 1Avicenna Biosciences Inc., 101 W Chapel Hill St, Durham, North Carolina, 27701
- 2College of Pharmacy
- 3Medical University of South Carolina
- 4Department of Computer Science and Department of Mathematics and Statistics
- 5Department of Biochemistry, Genetics and Microbiology
- 6University of Pretoria
Abstract
Structure-aware foundation models that jointly represent proteins and ligands promise to extend affinity prediction beyond the chemical applicability domains of conventional ligand-based models, but whether they provide sufficient resolution to guide fine-grained medicinal-chemistry decisions remains unclear. We evaluatedBoltz-2across12kinasetargetsusinghistoricalbioactivityreconstruction,binding-pocket randomization, cross-target selectivity analysis, and chemical-series evaluation. Across chemically diverse compounds in temporally reconstructed evaluation sets, Boltz-2 and target-specific ligand-based machine learningshowedcomparablemeanglobalrankingacrosstargets,withsubstantialtarget-dependent differences (mean Kendall’s τ = 0.40 vs 0.31). Randomizing the binding pocket reduced mean τ from 0.33 to 0.14, while predictions were centered near −7.64 kcal/mol and retained correlated compound profiles acrossproteins.Together,theseresultssuggestastrongligand-conditionedsignal,withadditional predictive information contributed by the protein pocket. Despite this protein sensitivity, compound-level selectivitypredictionremainedweak(meanτ=0.17),andtheexperimentallystrongesttargetwas recovered for only 18.9% of compounds. Ranking also weakened within 72 temporally classified chemical series and varied substantially among series (protein-balanced mean τ = 0.18 for Boltz-2 and 0.09 for ligand-based models). These results show that structure-aware models can recover affinity information beyond relationshipscaptured by conventionalligand-based approaches whilerevealingaresolution boundary between broad prioritization and the local energetic discrimination required for selectivity and analogue ranking. At their current demonstrated resolution, such models appear best suited to broad compound or series prioritization and complementary hypothesis generation. Extending their use to fine analogueranking,selectivityoptimization,andinterpretationofsmallaffinitydifferenceswillrequire prospective evidence of reliable local resolution.
Related
- Reliable Structure-Processing-Performance Learning in Organic Photovoltaics: Chemical Generalization, Processing-Aware Prediction and Uncertainty under Domain Shift
- FLOWR.ROOT – A flow matching-based foundation model for joint multi-purpose structure-aware 3D ligand generation and affinity prediction
- Applicability-Domain-Aware Evaluation of Machine-Learning NMR Chemical-Shift Prediction in Small drug Molecules: GT-NMR for 13C and CASCADE for 1H
- Spatial Applicability of Fictitious Point Feedback for Frequency Response Prediction in Unidirectional Carbon Fiber-Reinforced Polymer Plates
- What Random Splits Measure in Polymer Glass-Transition Prediction: Chemistry Holdouts, Applicability Domains, and Prior Integration
BibTeX
@misc{burger2026beyond,
title = {{Beyond the Ligand Applicability Domain: Generalization and Local Resolution in Structure-Aware Foundation Models for Affinity Prediction}},
author = {Pieter B. Burger and Thomas M. Kaiser and Parker Dryja and Morné Muller and MacLean Levin and Sandro Belvedere and Ilya Balabin and Yuri K. Peterson and Rian E. Pierneef and Fourie Joubert},
howpublished = {ChemRxiv},
year = {2026},
month = oct,
doi = {10.26434/chemrxiv.15010040/v1},
url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15010040/v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - Beyond the Ligand Applicability Domain: Generalization and Local Resolution in Structure-Aware Foundation Models for Affinity Prediction
AU - Pieter B. Burger
AU - Thomas M. Kaiser
AU - Parker Dryja
AU - Morné Muller
AU - MacLean Levin
AU - Sandro Belvedere
AU - Ilya Balabin
AU - Yuri K. Peterson
AU - Rian E. Pierneef
AU - Fourie Joubert
PY - 2026
DA - 2026/10/07
PB - ChemRxiv
DO - 10.26434/chemrxiv.15010040/v1
UR - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15010040/v1
AB - Structure-aware foundation models that jointly represent proteins and ligands promise to extend affinity prediction beyond the chemical applicability domains of conventional ligand-based models, but whether they provide sufficient resolution to guide fine-grained medicinal-chemistry decisions remains unclear. We evaluatedBoltz-2across12kinasetargetsusinghistoricalbioactivityreconstruction,binding-pocket randomization, cross-target selectivity analysis, and chemical-series evaluation. Across chemically diverse compounds in temporally reconstructed evaluation sets, Boltz-2 and target-specific ligand-based machine learningshowedcomparablemeanglobalrankingacrosstargets,withsubstantialtarget-dependent differences (mean Kendall’s τ = 0.40 vs 0.31). Randomizing the binding pocket reduced mean τ from 0.33 to 0.14, while predictions were centered near −7.64 kcal/mol and retained correlated compound profiles acrossproteins.Together,theseresultssuggestastrongligand-conditionedsignal,withadditional predictive information contributed by the protein pocket. Despite this protein sensitivity, compound-level selectivitypredictionremainedweak(meanτ=0.17),andtheexperimentallystrongesttargetwas recovered for only 18.9% of compounds. Ranking also weakened within 72 temporally classified chemical series and varied substantially among series (protein-balanced mean τ = 0.18 for Boltz-2 and 0.09 for ligand-based models). These results show that structure-aware models can recover affinity information beyond relationshipscaptured by conventionalligand-based approaches whilerevealingaresolution boundary between broad prioritization and the local energetic discrimination required for selectivity and analogue ranking. At their current demonstrated resolution, such models appear best suited to broad compound or series prioritization and complementary hypothesis generation. Extending their use to fine analogueranking,selectivityoptimization,andinterpretationofsmallaffinitydifferenceswillrequire prospective evidence of reliable local resolution.
ER -