Artificial intelligence uncovers carcinogenic human metabolites

From Black Boxes to Actionable Insights: A Perspective on Explainable Artificial Intelligence for Scientific Discovery

The application of Explainable Artificial Intelligence (XAI) in the field of chemistry has garnered growing interest for its potential to justify the prediction of black-box machine learning models and provide actionable insights. We first survey a range of XAI techniques adapted for chemical applications and categorize them based on the technical details of each methodology. We then present a few case studies to illustrate the practical utility of XAI, such as identifying carcinogenic molecules and guiding molecular optimizations, in order to provide chemists with concrete examples of ways to take full advantage of XAI-augmented machine learning for chemistry. Despite the initial success of XAI in chemistry, we still face the challenges of developing more reliable explanations, assuring robustness against adversarial actions, and customizing the explanation for different applications and needs of the diverse scientific community. Finally, we discuss the emerging role of large language models like GPT in generating natural language explanations and discusses the specific challenges associated with them. We advocate that addressing the aforementioned challenges and actively embracing new techniques may contribute to establishing machine learning as an indispensable technique for chemistry in this digital era.

Advancing chemical carcinogenicity prediction modeling: opportunities and challenges

Carcinogenicity assessment of any compound is a laborious and expensive exercise with several associated ethical and practical concerns. While artificial intelligence (AI) offers promising solutions, unfortunately, it is contingent on several challenges concerning the inadequacy of available experimentally validated (non)carcinogen datasets and variabilities within bioassays, which contribute to the compromised model training. Existing AI solutions that leverage classical chemistry-driven descriptors do not provide adequate biological interpretability involved in imparting carcinogenicity. This highlights the urgency to devise alternative AI strategies. We propose multiple strategies, including implementing data-driven (integrated databases) and known carcinogen-characteristic-derived features to overcome these apparent shortcomings. In summary, these next-generation approaches will continue facilitating robust chemical carcinogenicity prediction, concomitant with deeper mechanistic insights.

Chemistry-intuitive explanation of graph neural networks for molecular property prediction with substructure masking

Graph neural networks (GNNs) have been widely used in molecular property prediction, but explaining their black-box predictions is still a challenge. Most existing explanation methods for GNNs in chemistry focus on attributing model predictions to individual nodes, edges or fragments that are not necessarily derived from a chemically meaningful segmentation of molecules. To address this challenge, we propose a method named substructure mask explanation (SME). SME is based on well-established molecular segmentation methods and provides an interpretation that aligns with the understanding of chemists. We apply SME to elucidate how GNNs learn to predict aqueous solubility, genotoxicity, cardiotoxicity and blood-brain barrier permeation for small molecules. SME provides interpretation that is consistent with the understanding of chemists, alerts them to unreliable performance, and guides them in structural optimization for target properties. Hence, we believe that SME empowers chemists to confidently mine structure-activity relationship (SAR) from reliable GNNs through a transparent inspection on how GNNs pick up useful signals when learning from data.

ALipSol: An Attention-Driven Mixture-of-Experts Model for Lipophilicity and Solubility Prediction

Lipophilicity (logD) and aqueous solubility (logS_w) play a central role in drug development. The accurate prediction of these properties remains to be solved due to data scarcity. Current methodologies neglect the intrinsic relationships between physicochemical properties and usually ignore the ionization effects. Here, we propose an attention-driven mixture-of-experts (MoE) model named ALipSol, which explicitly reproduces the hierarchy of task relationships. We adopt the principle of divide-and-conquer by breaking down the complex end point (logD or logS_w) into simpler ones (acidic pK_a, basic pK_a, and logP) and allocating a specific expert network for each subproblem. Subsequently, we implement transfer learning to extract knowledge from related tasks, thus alleviating the dilemma of limited data. Additionally, we substitute the gating network with an attention mechanism to better capture the dynamic task relationships on a per-example basis. We adopt local fine-tuning and consensus prediction to further boost model performance. Extensive evaluation experiments verify the success of the ALipSol model, which achieves RMSE improvement of 8.04%, 2.49%, 8.57%, 12.8%, and 8.60% on the Lipop, ESOL, AqSolDB, external logD, and external logS data sets, respectively, compared with Attentive FP and the state-of-the-art in silico tools. In particular, our model yields more significant advantages (Welch's t-test) for small training data, implying its high robustness and generalizability. The interpretability analysis proves that the atom contributions learned by ALipSol are more reasonable compared with the vanilla Attentive FP, and the substitution effects in benzene derivatives agreed well with empirical constants, revealing the potential of our model to extract useful patterns from data and provide guidance for lead optimization.