Probing the toxicity of nanoparticles: a unified in silico machine learning model based on perturbation theory

Nanoparticles (NPs) are part of our daily life, having a wide range of applications in engineering, physics, chemistry, and biomedicine. However, there are serious concerns regarding the harmful effects that NPs can cause to the different biological systems and their ecosystems. Toxicity testing is an essential step for assessing the potential risks of the NPs, but the experimental assays are often very expensive and usually too slow to flag the number of NPs that may cause adverse effects. In silico models centered on quantitative structure–activity/toxicity relationships (QSAR/QSTR) are alternative tools that have become valuable supports to risk assessment, rationalizing the search for safer NPs. In this work, we develop a unified QSTR-perturbation model based on artificial neural networks, aimed at simultaneously predicting general toxicity profiles of NPs under diverse experimental conditions. The model is derived from 54,371 NP-NP pair cases generated by applying the perturbation theory to a set of 260 unique NPs, and showed an accuracy higher than 97% in both training and validation sets. Physicochemical interpretation of the different descriptors in the model are additionally provided. The QSTR-perturbation model is then employed to predict the toxic effects of several NPs not included in the original dataset. The theoretical results obtained for this independent set are strongly consistent with the experimental evidence found in the literature, suggesting that the present QSTR-perturbation model can be viewed as a promising and reliable computational tool for probing the toxicity of NPs.

纳米颗粒（Nanoparticles, NPs）已融入日常生活，在工程学、物理学、化学与生物医学等领域拥有广泛应用。然而，学界对纳米颗粒可能对各类生物系统及其生态系统造成的有害影响存在诸多严重关切。毒性测试是评估纳米颗粒潜在风险的必要环节，但传统实验检测往往成本高昂且耗时良久，难以快速甄别出可能产生不良效应的纳米颗粒数量。以定量构效/毒性关系（quantitative structure–activity/toxicity relationships, QSAR/QSTR）为核心的计算机模拟模型，已成为风险评估的重要辅助工具，可合理优化安全纳米颗粒的研发筛选流程。本研究构建了一种基于人工神经网络的统一QSTR扰动模型，旨在同时预测不同实验条件下纳米颗粒的综合毒性谱。该模型通过对260种独特纳米颗粒应用扰动理论，生成54371组纳米颗粒-纳米颗粒对案例进行训练，在训练集与验证集上的准确率均超过97%。本文还对模型中所用的各类描述符开展了理化层面的阐释。随后，利用该QSTR扰动模型对原始数据集未覆盖的多种纳米颗粒的毒性效应进行预测，针对该独立测试集得到的理论结果与文献记载的实验证据高度一致，表明本研究所提出的QSTR扰动模型可作为一种可靠且极具应用前景的计算工具，用于探究纳米颗粒的毒性。