A hypothesis-driven theoretical framework for a learnable geometric functional in mutation-conditioned ligand–receptor affinity prediction

Ioan-Matei Rusu ¹, Juliana Margineanu ^{1, 2}, Stefan-Rares Maxim¹, Radu Magop ¹, George Smau ², Razvan Rotaru², Ionel-Bogdan Tamba ¹

¹ Prof. Ostin C. Mungiu Advanced Research and Development Center for Experimental Medicine – CEMEX, Grigore T. Popa University of Medicine and Pharmacy, Iasi, Romania
² Alexandru Ioan Cuza University, Iasi, Romania
* Correspondence to: Bogdan-Ionel Tamba, Prof. Ostin C. Mungiu Advanced Research and Development Center for Experimental Medicine – CEMEX, Grigore T. Popa University of Medicine and Pharmacy, 16 Universitatii Str., 700115, Iasi, Romania. E-mail: bogdan.tamba@umfiasi.ro

Abstract

Predicting ligand–receptor binding affinity remains a cornerstone of modern drug discovery, enabling the identification of potential therapeutic compounds with high potency and selectivity. We introduce a new, interpretable framework that decomposes the total affinity into a geometric component and an energetic correction. The geometric term is defined as a learnable functional over overlapping atomic density fields and orientation-aware kernels, providing a measure of 3D shape complementarity that is explicitly invariant to global rotations and translations. The energetic component is estimated via a lightweight, symmetry-preserving surrogate graph network, capturing effects such as electrostatic interactions and solvation while remaining modular and extensible. A mutation-conditioned message-passing mechanism enables adaptation to receptor variants without recomputing expensive quantum data, addressing challenges in personalized medicine and protein engineering. An active learning loop links model predictions to sparse DFT reference points, balancing physical accuracy and computational efficiency through uncertainty-guided selection. This article is submitted as a hypothesis framework paper: it proposes a conceptual, testable architecture (with a proposed validation roadmap) rather than reporting validated predictive performance.