Quantifying Analogue Suitability for SAR-Based Read-Across Toxicological Assessment

Abstract

Structure activity relationship (SAR)-based read-across often is an integral part of toxicological safety assessment, and justification of the prediction presents the most challenging aspect of the approach. It has been established that structural consideration alone is inadequate for selecting analogues and justifying their use, and biological relevance must be incorporated. Here we introduce an approach for considering biological and toxicological related features quantitatively to compute a similarity score that is concordant with suitability for a read-across prediction for systemic toxicity. Fingerprint keys for comparing metabolism, reactivity, and physical chemical properties are presented and used to compare these attributes for 14 case study chemicals each with a list of potential analogues. Within each case study, the sum of these nonstructural similarity scores is consistent with suitability for read-across established using an approach based on expert judgment. Machine learning is applied to determine the contributions from each of the similarity attributes revealing their importance for each structure class. This approach is used to quantify and communicate the differences between a target and a potential analogue as well as rank analogue quality when more than one is relevant. A numerical score with easily interpreted fingerprints increases transparency and consistency among experts, facilitates implementation by others, and ultimately increases chances for regulatory acceptance.

Introduction to the Special Issue: AI Meets Toxicology

“…The methods described in this special issue cover a wide range of AI methods ranging from expert systems, ,, over similarity measures including read-across methods, − to classical machine learning such as random forests (RF), support vector machines (SVM), and artificial neural networks (ANN) − ,, to deep learning (DL) methods ,,− , including equivariant neural networks, deep generative models, and even large language models . In addition to models relying purely on the chemical structure, there is a notable trend of bringing in additional modalities to improve or inform predictive models. , In the following, we provide an overview of the AI approaches used in the publications contained in the SI.…”

The Nitrosamine “Saga”: Lessons Learned from Five Years of Scrutiny

“…The "Tanimoto" similarity is a term which includes a variety of potential similarity metrics; the Tanimoto comparison is a specific method for the similarity comparison of any pair of structural fingerprints rather than a universal number, and there are also other similarity measures available. Comparison of whole-molecule fingerprints, by Tanimoto or other similarity metrics mentioned in Table 2, is often used as a measure of similarity; however, it may not be as relevant for NAs (or perhaps other classes, as reported by Lester et al 58 ). This is due to the high dependence on local reactivity rather than pharmacophoric similarity for the potency of NAs; the two carbons adjacent to the nitrosamine are, respectively, the sites of metabolic activation and formed diazonium.…”

“…A framework for automating, by quantifying, this for the general read-across case has recently been published. 58 The selection of a specific analogue for a given NA may not follow the exact same metrics described by Lester et al, but the general principle may well be useful. An important point to make is that the difference in molecular weight can and should be used to scale weight-based limits (e.g., the 26.5 ng/day limit applied to NDEA (MW = 102.14) and those compounds read across to it) based on the number of molecules per mole present in vivo as a worst-case scenario, simply because NA mutagenicity is mechanistically linked to the molar amount and not to the mass.…”

“…Selection of an analogue, given the properties listed above, becomes a multiparameter optimization problem, where the most suitable analogue in one category may not necessarily be the most similar in another. A framework for automating, by quantifying, this for the general read-across case has recently been published . The selection of a specific analogue for a given NA may not follow the exact same metrics described by Lester et al, but the general principle may well be useful.…”

High Throughput Read-Across for Screening a Large Inventory of Related Structures by Balancing Artificial Intelligence/Machine Learning and Human Knowledge

“…Consideration of evidence from these diverse domains enables more holistic comparisons of chemical substances from the perspective of their fate and behaviors in biological systems. Also reported are the methods for quantification of the qualities of analogues and their study data, , which enables assessment of read-across reliabilities, outcomes (final resulting end point values, for example, an estimated point of departure (POD) for the target), and corresponding uncertainties. To support the needs, several public and commercial tools have become available in the past several years, including OECD QSAR Toolbox, VEGA HUB (including ToxREAD), ToxMatch, Generalized Read-Across GenRA , from the United States (US) Environmental Protection Agency (EPA) and ChemTunes·ToxGPS. , Several organizations use the latter platform as their internal system to support in silico safety/risk assessment, for example, the chemoinformatics group in the US Food and Drug Administration (FDA) Center for Food Safety and Applied Nutrition (CFSAN) and Cosmetics Europe.…”