Drug discovery with explainable artificial intelligence

José Jiménez-Luna; Francesca Grisoni; Gisbert Schneider

Source-linked AI summary

Drug discovery with explainable artificial intelligence

José Jiménez-Luna, Francesca Grisoni, Gisbert Schneider

arXiv:2007.00523v2Jul 1, 2020cs.AIcs.LGstat.ML

TL;DR

Deep-learning models can perform strongly in drug discovery, yet their mathematical reasoning is often difficult for humans to interpret. This review surveys explainable artificial-intelligence methods, applications, limitations, and future opportunities for drug discovery.

Problem
Deep-learning models can capture complex drug-discovery relationships but often remain difficult for humans to interpret.
Method
The review organizes prominent explainable-artificial-intelligence methods, their drug-discovery applications, limitations, and future opportunities.
Results
The review identifies contemporary XAI approaches and discusses their potential applications and methodological improvements for pharmaceutical research.
Takeaways & Limitations
XAI could support interpretation of complex chemical data and formulation of pharmacological hypotheses while fostering collaboration across drug-discovery disciplines.
Takeaways & Limitations
Most deep-learning drug-discovery models do not address applicability domains, risking high-confidence wrong predictions and unnecessary extrapolation.

Abstract

from arXiv · show

Deep learning bears promise for drug discovery, including advanced image analysis, prediction of molecular structure and function, and automated generation of innovative chemical entities with bespoke properties. Despite the growing number of successful prospective applications, the underlying mathematical models often remain elusive to interpretation by the human mind. There is a demand for 'explainable' deep learning methods to address the need for a new narrative of the machine language of the molecular sciences. This review summarizes the most prominent algorithmic concepts of explainable artificial intelligence, and dares a forecast of the future opportunities, potential applications, and remaining challenges.

1 Introduction

Deep learning is expanding computer-assisted drug discovery through nonlinear modeling, pattern recognition, and broader applications, but its models often remain difficult for humans to interpret. Explainable artificial intelligence seeks to make these systems more transparent, reliable, and useful across drug discovery and medicinal chemistry.

Deep learning algorithms model complex nonlinear input-output relationships and extract patterns and features from low-level data representations.
Some deep learning models match or exceed existing machine-learning and QSAR methods while broadening computer-assisted discovery applications.
Deep neural networks remain difficult for the human mind to understand, motivating efforts toward explainable artificial intelligence.
XAI aims to clarify AI decision-making, prevent wrong-reason predictions and unfair biases, and bridge machine learning with other scientific disciplines.In medicinal chemistry, it supports mechanistic interpretation of drug action, drug safety enhancement, and organic synthesis planning.
The review surveys recent XAI methods, applications, benefits, limitations, and future opportunities for drug discovery.

2 State of the art

The state of the art in explainable AI spans feature attribution, instance-based, self-explaining, prototype-based, and concept-learning approaches, with reported applications and prospective uses in drug discovery. Their interpretability depends on the explanatory method and the choice of molecular representation, while some approaches remain largely unexplored in drug discovery.

Feature attribution: Feature attribution methods comprise gradient-based, surrogate-model, and perturbation-based approaches that assign input-feature relevance for individual predictions.Surrogate methods approximate the original model with an interpretable model, whereas global surrogates can describe the underlying function more generally.
Feature attribution: Gradient-based methods can partially reconstruct original features, risking misinterpretation, and attribution interpretability is further constrained by complex molecular descriptors.The review recommends representations with interpretable meaning when constructing drug-discovery models.
Instance-based explanations: Anchor algorithms generate model-agnostic if–then rules that sufficiently guarantee a class prediction while explicitly modeling explanation coverage.They have been applied to image recognition, text classification, and visual question answering.
Instance-based explanations: Contrastive explanation methods provide instance-based classifier interpretability through pertinent positive and pertinent negative feature sets.They obtain these sets by perturbing an instance to change its predicted class and identifying critical original features.
Drug-discovery applications: Instance-based approaches had not yet been applied to drug discovery but were proposed for activity-cliff prediction, fragment-based virtual screening, and hit-to-lead optimization.These uses target structural variations, minimal activity-associated atom subsets, and optimization guidance, respectively.
Beyond post-hoc explanations: Self-explaining methods aim to provide human-interpretable explanations alongside predictions, while prototype-based reasoning and human-interpretable concept learning support example-based forecasting and human-like generalization.An active-learning application reported efficient cycle guidance in the low-data regime by selecting molecules with the largest estimated epistemic uncertainty for melting-point and aqueous-solubility tasks.

3 Available software

Several software tools have been developed to facilitate interpretation of deep-learning models, including Captum and Alibi. Captum supports most feature-attribution techniques described in the work, while Alibi provides instance-specific explanations for certain models.

Available software: Captum extends PyTorch and provides support for most feature-attribution techniques described in this work.It is built on PyTorch’s deep-learning and automatic-differentiation package.
Available software: Alibi provides instance-specific explanations for certain models.

4 Conclusions and outlook

XAI is expected to support collaboration, chemical-data interpretation, and pharmacological hypothesis generation in drug discovery, while requiring application-specific methods and domain expertise. Key challenges include molecular representation, applicability domains, open collaborative platforms, and the need for comprehensible, affordable models.

Opportunities: XAI could foster collaboration among medicinal chemists, chemoinformaticians, and data scientists while supporting chemical-data interpretation and pharmacological hypothesis formulation.The passage also notes that XAI may help avoid human biases and that drug-discovery challenges could drive application-tailored methods.
Limitations: Full comprehensibility may be difficult because chemical language, molecular pathology, and nonlinear structure–activity relationships incompletely capture model decision spaces.The passage emphasizes both opportunities and limitations in established chemical representations and acknowledges incomplete molecular-level understanding of human pathology.
Technical challenges: XAI explanations must be tailored to each application and be non-trivial, non-artificial, and sufficiently informative for the relevant scientific community.Selecting meaningful explanations requires technical methods and profound problem-domain knowledge, with collaboration across deep learning, chemistry, biology, and data science.
Model foundations: Interpretable low-level molecular representations, established descriptors, and applicability-domain restrictions are important directions for explainable drug-discovery models.Examples include SMILES strings, amino acid sequences, spatial 3D-voxelized representations, hashed binary fingerprints, and topochemical and geometrical descriptors; applicability domains help address confident wrong predictions.
Collaboration and outlook: Opencommunity platforms for sharing code, interpretations, and training data, together with initiatives such as MELLODDY, could support validation and acceptance of XAI approaches.MELLODDY is described as enabling decentralized, federated model development and secure data handling across pharmaceutical companies.
Outlook: Users should cautiously inspect deep-learning predictions, while continued development of more comprehensible and computationally affordable models remains important.The passage frames this responsibility especially in time- and cost-sensitive drug-discovery scenarios.