NiMARN: Negative Memory Replay to Training Collapse and Exploitation in Bifunctional Antimicrobial Peptide Generation and Optimization
1 Graduate School in the Program of Pharmaceutical Sciences, Faculty of Pharmaceutical Sciences, Khon Kaen University, Khon Kaen, Thailand.
2 Division of Pharmaceutical Chemistry, Faculty of Pharmaceutical Sciences, Khon Kaen University, Khon Kaen, Thailand.
* Corresponding author: Tarapong Srisongkram (tarasri@kku.ac.th)
Source codes and supplementary information for the manuscript are available in this repository.
Designing dual-function antimicrobial peptides with simultaneous anti-MRSA and antioxidant activities is severely constrained by extreme data scarcity, while multi-objective reinforcement learning policies optimizing against surrogate predictive models frequently suffer from out-of-distribution exploitation, sequence diversity collapse, and severe reward hacking. In this paper, we propose the Nightmare-Informed Memory, Awareness and Replay Network (NiMARN), a memory replay framework that utilizes historical failed sequence trajectories as an internal policy regularizer during generative multi-objective optimization. We establish empirical benchmark evaluations across five replay fraction configurations (eta in {0.0, 0.2, 0.5, 0.8, 1.0}) over 2,000 training steps, supported by a compact anti-MRSA Transformer predictor achieving an accuracy of 0.9849, Matthews correlation coefficient of 0.9405, and AUROC of 0.9964. Our theoretical and experimental results demonstrate that bounded negative memory replay (eta in [0.5, 0.8]) completely suppresses flagged reward-hacking candidates from 22.8% (91/400) in the baseline (eta = 0.0) down to 0.0% (0/400), reducing early-to-late top-4-mer motif overlap from 100% to 20.0% (eta = 0.8), while expanding the proportion of high-reward, high-novelty candidates to 33.1% (eta = 0.5), offering a new principled approach for data-efficient, safe peptide drug design.
Torch: 2.11.0; TensorFlow: 2.16.2; Transformers: 5.5.4; umap-learn: 0.5.12; XGBoost: 3.2.0; Scikit-learn: 1.7.2; SciPy: 1.11.4; NumPy: 1.26.4; Pandas: 2.1.4; Matplotlib: 3.7.5.
Copyright (c) [2026] [Dr. Tarapong Srisongkram]