Proxy-Objective Misspecification in AI-Guided Design of Targeted Nanocarriers: Why Optimising Predicted Affinity Can Reduce Therapeutic Selectivity

Abstract

Machine-learning and generative approaches to nanocarrier design optimise a predicted quantity (binding affinity, cellular uptake, or avidity for a target receptor) as a stand-in for the therapeutic objective, which is selective delivery to target rather than non-target tissue. This theoretical–analytical paper shows that the two are not monotonically related, and that the region in which they diverge is one an optimiser will actively seek out. The study is non-empirical: a minimal thermodynamic model of multivalent binding is stated with its assumptions declared, therapeutic selectivity is defined on it as the ratio of target to non-target accumulation including a non-specific uptake floor, and the behaviour of selectivity with respect to ligand number is derived analytically and then mapped numerically across the receptor-contrast and adsorption parameter space; sources were identified by structured searches of the biomedical corpus and verified individually against the publisher record, with screening for retractions, corrections and expressions of concern. Three regimes emerge. At low valency the selectivity exponent equals unity exactly and selectivity simply equals the receptor-density ratio, so added affinity buys nothing. In the superselective regime, where the product of ligand number and binding opportunity exceeds unity while binding remains below saturation, selectivity is amplified well beyond the receptor-density ratio. Once target binding approaches saturation, further avidity cannot raise target accumulation but continues to recruit non-target accumulation, and selectivity falls. Selectivity therefore has an interior maximum, located near the valency that just saturates target binding, and numerical mapping shows the full rise-and-fall structure in ten of twelve parameter combinations examined. The cost of overshooting that maximum is greatest precisely where the receptor-density contrast is smallest: at a density ratio of two, tripling the optimal valency returned selectivity to unity, the value corresponding to no targeting at all, whereas at a ratio of fifty the same overshoot cost little. Since receptors exploited for targeting are characteristically overexpressed rather than exclusive, the realistic design problem sits in the unforgiving region. Critically, predicted avidity was non-decreasing in ligand number throughout, while selectivity was negative in its derivative over two-thirds of the range examined, so an optimiser maximising the proxy reports monotone improvement at every step, including every step that reduces selectivity. This is a misspecification rather than an inaccuracy: the surrogate may be predicting avidity perfectly, and no gain in predictive accuracy, training-set size or validation rigour can correct it. The paper specifies the form a selectivity-aware objective must take, noting that it must reference non-target receptor density and non-specific uptake, quantities absent from any affinity surrogate, and proposes six objective-specification and reporting requirements for AI-guided targeting design, arguing that optimal rather than maximal ligand density is the appropriate design target.