Explaining molecular properties with natural language

Abstract

We present a model-agnostic method that gives natural language explanations of molecular structure property predictions. Machine learning models are now common for molecular property prediction and chemical design. They typically are black boxes -- having no explanation for predictions. We show how to use surrogate models to attribute predictions to chemical descriptors and molecular substructures, independent of the black box model inputs. The method generates explanations consistent with chemical reasoning, like connecting existence of a functional group or molecular polarity. We see in a genuine test like blood brain barrier permeation, our descriptor explanations match biologically observed SARs with mechanistic support. We show these quantitative explanations can be further translated to natural language.

A Perspective on Explanations of Molecular Prediction Models

“…28 An extrinsic method is one that can be applied post-training to any model. 33 Posthoc methods found in the literature focus on interpreting models through (1) training data 34 and feature attribution, 35 (2) surrogate models 10 and, (3) counterfactual 9 or contrastive explanations. 36 Often, what is a "good" explanation and what are the required components of an explanation are debated.…”

“…Model interpretation is certainly not new and a common step in ML in chemistry, but XAI for DL models is becoming more important. ,− ,,,, Here, we illustrate some practical examples drawn from our published work on how model-agnostic XAI can be utilized to interpret black-box models and connect the explanations to structure–property relationships. The methods are “Molecular Model Agnostic Counterfactual Explanations” (MMACE) and “Explaining molecular properties with natural language” . Then, we demonstrate how counterfactuals and descriptor explanations can propose structure–property relationships in the domain of molecular scent …”

“…Deep learning (DL) is advancing the boundaries of computational chemistry because it can accurately model nonlinear structure–function relationships. − Applications of DL can be found in a broad spectrum spanning from quantum computing , to drug discovery − to materials design. , DL models can contribute to scientific discovery in three “dimensions”: (1) as a “computational microscope” to gain insight which is not attainable through experiments, (2) as a “resource of inspiration” to motivate scientific thinking; (3) as an “agent of understanding” to uncover new observations . However, the rationale of a DL prediction is not always apparent due to the model architecture consisting of a large parameter count. , DL models are thus often termed “black-box” models.…”

“…In the remainder of this Review, we review recent approaches for XAI of chemical property prediction while drawing specific examples from our recent XAI work. ,, We show how in various systems these methods yield explanations that are consistent with known mechanisms in structure–property relationships.…”

Explaining compound activity predictions with a substructure-aware loss for graph neural networks

“…The popularity of GNNs has also been accompanied by an increasing need for explainability, [5][6][7][8][9][10][11][12][13] as these models have been notoriously known for their black-box character. Towards this goal, explainable artificial intelligence techniques, such as feature attribution analyses, have become relevant tools.…”