This work is licensed under Creative Commons Attribution–NonCommercial International License
(CC BY-NC 4.0).
Background:Periodontal infection involves complex interactions among bacteria, immune cells,
cytokines, and tissue, which can be challenging to model and predict.
Objectives:To develop a minimal mechanistic model of periodontal infection incorporating key
components (bacteria, neutrophils, macrophages, cytokine, tissue), generate a synthetic cohort, and train a
Long Short-Term Memory (LSTM) neural network to forecast tissue integrity and evaluate antibiotic
intervention schedules.
Results:A 100-subject synthetic cohort was generated using the mechanistic model. The LSTM network
was trained on five days of multi-omic data to predict one-day-ahead tissue integrity, achieving a mean
squared error of approximately 690 on a hold-out set. Virtual antibiotic treatment simulations
demonstrated a preservation of about 15% more tissue by day 30 compared to untreated controls.
Conclusion:The combined mechanistic and neural modeling framework effectively captured disease
progression and enabled virtual evaluation of therapy schedules, highlighting its potential for optimizing
periodontal treatment strategies.