Title: Defining and Measuring Fairness in Deep Reinforcement Learning for Drug Discovery: A Rapid Evidence Review

Abstract: Although deep reinforcement learning (DRL) is gaining traction in de novo molecular design, variations in data selection, reward structures, and evaluation methods often lead to disparate performance across different chemotypes and disease areas. Currently, however, there is a lack of comprehensive synthesis regarding how fairness is conceptualized, quantified, and validated within DRL-based drug discovery frameworks. This rapid evidence review addresses this gap by consolidating fairness definitions and metrics specifically for DRL-driven molecule generation in the healthcare sector. The study investigates three primary inquiries: (i) the impact of dataset composition and splitting methodologies—particularly the contrast between scaffold-based and random splits—on evaluation integrity and distribution shift; (ii) the role of reward engineering (such as QED, docking scores, toxicity assessments, and synthetic accessibility metrics) in either exacerbating or alleviating bias, with a specific focus on cancer targets; and (iii) the identification of optimal measurable metrics for capturing fairness. These metrics encompass parity between cancer and non-cancer indications, as well as among various cancer subtypes. Additionally, they account for distributional equilibrium in critical physicochemical descriptors, scaffold and chemotype diversity, group-level validity, toxicity profiles, and synthetic accessibility. Conducting a search from 2017 onwards across major biomedical, computer science, and engineering databases, alongside arXiv for horizon scanning, we employed PRISMA-style screening protocols. The resulting records were analyzed through content coding to correlate reported parity outcomes with specific dataset and reward configurations. This review delivers a streamlined collection of fairness definitions and metrics for DRL molecule generation, offering actionable guidance for reporting both distribution and outcome parity. Furthermore, it elucidates the relationship between dataset and reward selections and observed parity effects, while highlighting critical gaps that must be addressed to ensure trustworthy, cancer-relevant DRL generation.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC