Inductive Bias and the Reference-Class Problem in Machine-Learning Prediction of Nanoparticle Tumour Accumulation: Enhanced Permeability and Retention as an Assumption Embedded in Training Data

Abstract

Machine-learning models that predict nanoparticle tumour accumulation and biodistribution are trained predominantly on murine data, in which enhanced permeability and retention is pronounced and comparatively consistent. This theoretical paper argues that the effect therefore enters such models not as a hypothesis under test but as an inductive bias embedded in the training distribution, with consequences that differ from ordinary covariate shift. The study is non-empirical and conceptual: the sampling structure of such training corpora is characterised, contrasted with the established framings of dataset shift and applicability domain, and its consequences for validation and interpretation derived; sources were identified by structured searches of the biomedical corpus, deliberately retaining work on both sides of the dispute about the effect's magnitude in human disease, and each was verified against the publisher record with screening for retractions, corrections and expressions of concern. Because the assumption is carried by the sampling of the training data rather than by any identifiable feature, no function of the available features can represent the dependence of the outcome on it, and the model cannot represent its own inapplicability. It follows that validation conducted within the murine distribution, however rigorous, and including held-out, cross-study and cross-tumour-model designs, cannot detect the bias, because a bias shared by every stratum is invisible to comparisons among strata; reported validation performance can therefore be excellent and entirely uninformative about transferability. The paper formulates the difficulty as one of reference class: a predicted accumulation value is well defined only relative to a population, and the population implicit in the training data is not the population at deployment. Its central practical contribution separates the components of such a prediction that depend on the contested magnitude of the effect from those that depend only on the weaker and widely accepted claim of heterogeneous, tumour-dependent permeability, showing that ranking particles within a tumour model and generating hypotheses about structure–accumulation relationships remain supportable, whereas predicting absolute accumulation, delivered fraction, or therapeutic sufficiency in human tumours does not. This decomposition allows the modelling question to be settled without adjudicating the biological dispute, and the argument is deliberately constructed so as not to require adjudication. Three routes to licensing interspecies transfer are specified: encoding permeability-relevant physiology per case, coupling learned components to mechanistic structures whose parameters carry species-specific values, and direct paired cross-species evidence. Reference-class declaration is proposed as a reporting requirement, expressed in the context-of-use idiom of current regulatory guidance.