Perturbation-Aware Neural ODE (pNODE) Learns Microbiome Dynamics from Clinical Data and Predicts Gut-Borne Bloodstream Infections in Patients Receiving Cancer Treatment

Disruption of the gut microbiota during cancer treatment, particularly in allogeneic hematopoietic cell transplantation (allo-HCT), is a major contributor to adverse clinical outcomes, including gut-borne bloodstream infections. Accurately forecasting microbial population dynamics under clinical perturbations, such as antibiotic administration, could inform treatment strategies that reduce the risk of infection. However, traditional models like the Generalized Lotka-Volterra (gLV), which consider only pairwise interactions of constant sign and magnitude, are limited in their ability to capture the real-world nonlinear dynamics of multispecies microbiomes following ecosystem disturbances. Here, we introduce a perturbation-augmented Neural Ordinary Differential Equation (pNODE) framework that flexibly models microbial population dynamics in continuous time, integrating both microbial abundances and time-resolved antibiotic perturbations. Using both synthetic and real clinical data from over 1,000 allo-HCT patients, we demonstrate that pNODEs outperform gLV in predictive accuracy, robustness to noise, and the ability to forecast critical events, such as the intestinal expansion of an opportunistic pathogen. Notably, we demonstrate that running a pre-trained pNODE in generative mode to simulate prospective Bacilli abundance trajectories from an initial sample and antibiotic timeline yields scores that accurately predict subsequent Enterococcus infections in held-out cohorts, outperforming baseline and ground-truth-based predictors. Because pNODEs are purely data-driven dynamical systems, they can generate artifacts, such as small negative abundance values, when run in generative mode; nevertheless, our results show that these artifacts do not impair the model's ability to encode clinically relevant ecological structure and may even enhance it. Our findings demonstrate the potential of pNODEs as a next-generation tool for modeling clinical microbiome dynamics, with applications for predicting infections in immunocompromised patients hospitalized for cancer treatment.