💻 bioinformatics

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

This study introduces the perturbation-aware Neural ODE (pNODE) framework, which outperforms traditional models by accurately forecasting gut microbiome dynamics and predicting gut-borne bloodstream infections in cancer patients undergoing allogeneic hematopoietic cell transplantation through the integration of microbial abundances and time-resolved antibiotic perturbations.

Original authors: Stamper, I. C., Aeria, B., Xavier, J.

Published 2026-09-10

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Stamper, I. C., Aeria, B., Xavier, J.

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Inside the human gut, a vast and complex community of bacteria works tirelessly to keep the body healthy. These microscopic residents do more than just digest food; they act as a shield against invaders, training the immune system and preventing harmful pathogens from taking hold. However, when a person undergoes aggressive cancer treatment, such as a bone marrow transplant, this delicate ecosystem can be shattered. Powerful drugs and radiation wipe out the immune system's defenses, forcing doctors to use broad-spectrum antibiotics to prevent infections. While these antibiotics save lives, they also strike indiscriminately, destroying the beneficial bacteria that keep the gut stable. In the resulting vacuum, opportunistic germs that usually live in tiny numbers can explode in population, crowding out the good bacteria and eventually spilling into the bloodstream, causing severe, sometimes fatal, infections.

For years, scientists have tried to predict how these microbial communities will react to such disturbances, hoping to find a way to steer the gut back to health before an infection takes hold. Traditional methods for modeling these changes rely on simplified rules that assume bacteria interact in fixed, predictable ways, like gears in a machine. But the real gut is far messier and more dynamic, reacting to antibiotics in complex, non-linear ways that simple rules cannot capture. Researchers needed a new approach, one that could learn directly from the messy reality of patient data rather than forcing that data into a rigid box.

A team of scientists at the Sloan Kettering Institute for Cancer Research and UCLA has developed a new tool called the perturbation-aware Neural ODE, or pNODE, to solve this problem. Instead of relying on pre-set rules, this system uses a type of artificial intelligence designed to learn the laws of motion directly from observation. Imagine watching a river flow: a traditional model might try to guess the water's path based on a static map of the riverbed, but pNODE watches the water move, learns how it swirls around rocks and changes speed with the current, and then predicts where it will go next. In this case, the "river" is the collection of bacteria in a patient's gut, and the "rocks" are the antibiotics being administered. The researchers fed the system data from over 1,000 patients who had received bone marrow transplants, including detailed records of their daily antibiotic treatments and thousands of stool samples tracking the changing bacterial populations.

The results were striking. When tested against synthetic data where the true rules were known, pNODE consistently outperformed the traditional models, even when the data was noisy and incomplete, mimicking the real-world conditions of a hospital. But the true test came when the team applied the model to real patient data. They trained the system to predict how a patient's gut bacteria would change over the next week based solely on an initial stool sample and the planned antibiotic schedule. The model proved remarkably accurate, forecasting the rise and fall of specific bacterial groups better than the older methods. Most importantly, the researchers discovered that the model could predict clinical outcomes it had never been explicitly taught to find. By running the model forward in time, they could generate a predicted path for a specific dangerous bacterium called Enterococcus. If the model predicted that this bacterium would grow to dominate the gut, it was a strong signal that the patient was at high risk of developing a bloodstream infection.

This predictive power held up even when the model was tested on patients it had never seen before. The team found that the model's forecast of Enterococcus growth was a better predictor of future infections than simply looking at the current level of the bacteria or using the older, rule-based models. In some cases, the model's predictions were as accurate as if the researchers had already known the future bacterial counts, a feat that suggests the system had learned the underlying ecological rules governing the gut. The researchers also visualized these dynamics by mapping the gut's bacterial communities onto a two-dimensional landscape. In this view, different antibiotics act like winds, pushing the entire community in specific directions. Some antibiotic regimens push the community toward a state where Enterococcus takes over, while others keep the ecosystem diverse and stable.

The study suggests that this new approach could eventually help doctors make smarter choices about antibiotic use. By simulating how different treatment plans would affect a patient's specific gut ecosystem, clinicians might be able to choose regimens that treat the infection without destroying the protective microbiome. While the model was tested on a specific group of cancer patients and requires further validation in other settings, it represents a significant step forward. It moves the field from simply observing what happens after an infection to understanding the dynamic forces that lead there, offering a potential path to preventing infections before they ever start.

Technical Summary: Perturbation-Aware Neural ODE (pNODE) for Microbiome Dynamics

Problem Statement

Cancer treatments, particularly allogeneic hematopoietic cell transplantation (allo-HCT), severely disrupt the gut microbiota. This disruption, often exacerbated by broad-spectrum antibiotic administration, reduces biodiversity and impairs colonization resistance, allowing opportunistic pathogens like Enterococcus faecium to expand and dominate the gut ("intestinal domination"). This expansion facilitates translocation into the bloodstream, leading to life-threatening bloodstream infections (BSIs).

Current predictive models for microbiome dynamics, such as the Generalized Lotka-Volterra (gLV) model, rely on simplifying assumptions: linear, pairwise, time-invariant interactions in a homogeneous environment. These models struggle to capture the nonlinear, context-dependent dynamics of the intestinal microbiome under complex clinical perturbations like antibiotic regimens. Furthermore, clinical data is characterized by sparse, irregular sampling and high noise, conditions under which traditional parametric models often fail to generalize. There is a critical need for a flexible, data-driven framework that can model continuous-time microbial dynamics while explicitly integrating time-resolved external perturbations (antibiotics) to forecast ecological shifts and clinical outcomes.

Methodology

The pNODE Framework

The authors introduce the Perturbation-Aware Neural Ordinary Differential Equation (pNODE). Unlike standard Neural ODEs (NODEs) that model autonomous dynamics, pNODE explicitly incorporates external perturbations into the differential equation governing the system's state.

Architecture: The system state $x(t)$ represents microbial abundances. The dynamics are governed by a latent state $z(t)$ integrated by a neural network $f_\theta$ :
$\frac{dz}{dt} = f_\theta(x(t), p(t), t)$
where $x(t)$ is the current species abundance vector, $p(t)$ is a time-dependent perturbation vector encoding antibiotic exposure, and $t$ is the absolute time (days relative to transplant).
Perturbation Encoding: Antibiotic administration is encoded as a 15-dimensional binary vector $p(t)$ , where each dimension corresponds to a specific antibiotic category (e.g., carbapenems, fluoroquinolones). $p_i(t) = 1$ if the patient is receiving an antibiotic from category $i$ at time $t$ , and $0$ otherwise.
Compositional Output: To ensure predictions remain valid relative abundance vectors (summing to 1), the latent state is mapped to the observation space via a softmax function: $x(t) = \text{softmax}(z(t))$ .
Training: The model is trained using a mean absolute error (MAE) loss between predicted and observed relative abundances over short-horizon windows (7 days). The ODE solver (Euler method via torchdiffeq) differentiates through the integration process to update weights.

Data and Experimental Design

Clinical Cohort: The study utilized a longitudinal dataset of over 1,000 allo-HCT patients from Memorial Sloan Kettering Cancer Center, comprising >10,000 fecal microbiota samples (16S rRNA sequencing), detailed antibiotic records, and clinical outcomes.
Taxonomic Resolutions: Two model configurations were trained:
1. Class-level (N=14): Aggregated at the bacterial class level, with Enterococcus explicitly separated from the Bacilli class. This subset was restricted to patients with qPCR absolute abundance data to enable direct comparison with gLV.
2. Genus-level (N=136): A higher-resolution model trained on the full cohort, retaining genera with >5% prevalence.
Synthetic Benchmark: To validate the framework under controlled conditions, the authors generated synthetic 5-species ecosystems governed by a perturbation-augmented gLV model. These datasets included known ground-truth parameters, irregular sampling, and Gaussian noise to mimic clinical data characteristics.
Generative Mode for Prediction: Once trained, the pNODE operates in generative mode. Starting from an initial fecal sample and a patient's specific antibiotic timeline, the model integrates forward to predict future microbial trajectories without requiring intermediate observations.

Key Results

1. Superior Performance over gLV in Synthetic and Clinical Settings

Synthetic Data: Despite the ground truth being generated by a gLV model, pNODE consistently outperformed gLV inference across a grid of conditions (varying noise levels, sampling densities, and training set sizes). The performance gap was most pronounced in sparse, noisy, and data-poor regimes typical of clinical studies.
Clinical Data (qPCR Subset): On the subset of patients with absolute abundance data, pNODE achieved a lower 7-day prediction MAE ( $\approx 0.047$ ) compared to gLV ( $\approx 0.049$ ) and a constant-composition baseline. This advantage widened at longer horizons (30 and 60 days).
Clinical Data (Full Cohort): The genus-level pNODE achieved a 7-day MAE of $\approx 0.007$ on held-out patients, successfully capturing major ecological transitions, including progression to Enterococcus domination and recovery of diverse communities.

2. Prediction of Bloodstream Infections (BSI)

The study demonstrated that pNODE-generated trajectories can predict Enterococcus BSIs without the model ever being trained on infection outcomes.

Method: For each patient, the model generated a predicted Enterococcus abundance trajectory over a future window $W$ . The mean predicted abundance served as a risk score.
Performance: The pNODE-based risk score outperformed both the initial Enterococcus abundance baseline and the gLV generative predictions.
- At a 30-day horizon, pNODE achieved an AUC of $0.83 \pm 0.01$ , surpassing gLV ($0.70$) and matching or exceeding an interpolated ground-truth oracle ($0.82$).
- The advantage of pNODE increased with longer prediction horizons, maintaining high AUC ( $\approx 0.81-0.82$ ) at 58 days where the oracle's performance declined.

3. Visualization of Antibiotic-Dependent Flow Fields

By projecting the genus-level pNODE dynamics onto a TaxUMAP embedding, the authors visualized the learned vector field.

Different antibiotic regimens (perturbation states) were shown to reshape the flow field, steering microbial communities toward distinct ecological attractors.
Specific antibiotics (e.g., penicillins) were observed to drive communities toward Enterococcus-dominated regions, providing a mechanistic visualization of how treatment choices influence ecological trajectories.

Significance and Claims

The paper claims that pNODE represents a significant advancement in modeling clinical microbiome dynamics by:

Bridging the Gap Between Data and Dynamics: It demonstrates that flexible, non-parametric neural ODEs can outperform classical ecological models (gLV) even when the underlying system is theoretically gLV-like, particularly under the sparse and noisy conditions of real-world clinical data.
Predictive Utility Without Outcome Training: The model successfully predicts downstream clinical events (BSIs) based solely on learning microbial dynamics and antibiotic perturbations. This suggests the model captures the upstream ecological mechanisms (intestinal domination) that precede infection, rather than merely correlating with the outcome.
Generative Capability for "What-If" Scenarios: The framework enables the simulation of prospective microbial trajectories under different antibiotic regimens. This capability lays the groundwork for microbiome-aware treatment strategies, where clinicians could potentially evaluate the ecological consequences of different antibiotic choices before administration.
Interpretability via Flow Fields: The ability to visualize antibiotic-dependent flow fields offers a new way to understand how specific drug combinations drive the microbiome toward clinically relevant states (e.g., pathogen domination), generating testable hypotheses about ecological resilience and susceptibility.

The authors conclude that while pNODE is a promising tool for predictive modeling, it is currently limited to single-center data and binary perturbation encoding. Future work is needed to validate these dynamics across diverse institutions and to incorporate richer features such as dosage, absolute abundance, and host variables to support prospective clinical decision-making.