Accelerating Multi-Property Molecular Design via Entropic-Risk-Based Counterfactual Explanations

Pasan Dissanayake, Po-Yen Chen, Sanghamitra Dutta

Keywords: multi-target counterfactual explanations, molecular inverse design, biopolymer nanocomposites, optimization, entropic risk

TL;DR: We propose FINDER, a multi-target counterfactual generating method based on entropic risk measure for biopolymer nanocomposite inverse design.

Abstract:

Inverse design of composite molecular structures is computationally expensive because of vast search spaces. The challenge is further exacerbated if the desired molecule is expected to simultaneously satisfy multiple properties. For instance, biopolymer nanocomposites offer significant promise as sustainable plastic alternatives but the candidates are required to meet multiple
performance criteria simultaneously, such as mechanical strength, biodegradability, optical transparency, etc. Existing techniques often focus on only one property at a time, and rely upon computationally expensive genetic algorithms or perturbation-based optimization techniques. In this paper, we propose FINDER (Fast INverse Design via Entropic-Risk-based counterfactual explanations), a unified framework for composite molecular design that can cater to multiple target properties efficiently using a novel iterative counterfactual generation mechanism. Counterfactual explanations, a term from explainable AI, typically refer to the smallest possible changes to an input that can lead a machine learning model to give a different desired output. FINDER brings in several new innovations: (i) proposing a new constrained optimization for finding counterfactual explanations that satisfy multiple target properties; (ii) introducing a flexible tuning knob via entropic risk that balances different properties rather than a worst-case multi-property optimization (min-max); (iii) incorporating iterative projected gradient descent that is much faster; and (iv) integrating with strings like SMILES to modify functional groups and ultimately arrive at realistic and synthetically-achievable molecules. Our experiments on the high-entropy alloys benchmark successfully finds candidates with minimal composition changes that satisfy multiple properties simultaneously. We further validate FINDER on a real-world lab-generated dataset of biopolymer nanocomposites, finding entirely new composite molecules with not just adjusted ratios but modified functional groups altogether.

Submission Track: Full Paper

Submission Category: AI-Guided Design

Submission Number: 56