dAMN: a genome scale neural-mechanistic hybrid model to predict bacterial growth dynamics

This study introduces dAMN, a hybrid neural-mechanistic model that integrates neural networks with genome-scale dynamic flux balance analysis to accurately predict bacterial growth dynamics and substrate consumption across diverse nutrient environments, including realistic lag phase adaptation and acetate overflow phenomena.

The Gist

Imagine you are trying to predict how fast a bacterial city will grow and how much food it will eat.

For a long time, scientists have had two main ways to do this, and both had big problems:

1. The "Strict Accountant" (Mechanistic Models): This approach uses a massive rulebook of chemistry (like a genome-scale map) to calculate exactly what the bacteria should do. It's very accurate regarding the rules of physics and chemistry, but it's rigid. It assumes the bacteria instantly switch gears the moment food changes, which isn't how real life works. It also struggles to predict the "lag phase"—that awkward pause where bacteria are just waking up and adjusting before they start growing.
2. The "Gambler" (Pure AI/Neural Networks): This approach uses Artificial Intelligence to look at past data and guess the future. It's great at spotting patterns, but it doesn't really understand the rules of biology. It might guess that bacteria grow on poison or eat more food than exists in the universe because it doesn't have a "reality check."

Enter dAMN: The "Hybrid Chef"

The paper introduces a new tool called dAMN (dynamic Artificial Metabolic Networks). Think of dAMN as the perfect marriage between the Strict Accountant and the Gambler. It's a "hybrid" model that uses the best of both worlds.

Here is how it works, using a simple analogy:

The Restaurant Analogy

Imagine a bacterial colony is a busy restaurant kitchen.

• The Menu (The Genome): The kitchen has a massive menu (the genome-scale model) listing every possible dish (chemical reaction) they can cook.
• The Ingredients (The Nutrients): The customers bring in specific ingredients (sugars, amino acids).
• The Chef (The Neural Network): This is the AI part. The Chef looks at the ingredients on the counter and says, "Okay, I see we have glucose and succinate. I know from experience that when we get this mix, the kitchen takes 20 minutes to warm up (the lag phase), and then we start cooking at a specific speed."
• The Rules (The Mechanistic Constraints): This is the Accountant part. Even though the Chef is making guesses, the Accountant is standing right there with a clipboard. If the Chef tries to cook a dish that requires 100 eggs but the kitchen only has 10, the Accountant slaps the table and says, "No! That violates the laws of conservation!" The Chef must adjust the recipe to fit the rules.

What Makes dAMN Special?

1. It Understands the "Wake-Up" Time (Lag Phase)
Real bacteria don't start growing the second they touch food. They need time to build their machinery. Old models often skipped this or got it wrong. dAMN learns this "waking up" time directly from the data. It's like the model knows, "Oh, this specific mix of ingredients is weird; the bacteria will need a coffee break before they start working."

2. It Predicts the Future Without Being Told
In the study, the scientists trained dAMN on bacteria eating simple food. Then, they asked the model to predict what would happen if the bacteria were given a completely new mix of foods it had never seen before.

• The Result: dAMN didn't just guess; it figured out the logic. It predicted that the bacteria would eat the "easy" food first (glucose), get full, and then switch to the "harder" food (acetate) later. This is called a diauxic shift. Even though the model wasn't explicitly taught this rule, the combination of AI intuition and chemical rules forced it to discover this behavior naturally.

3. It's a "Self-Correcting" System
The model has a "loss function," which is just a fancy way of saying "a grading system."

• If the model predicts the bacteria grow too fast, the Accountant (rules) says "Too fast!"
• If the model predicts the bacteria eat more sugar than is available, the Accountant says "Impossible!"
• The model gets a "bad grade" and tries again until it finds a prediction that is both smart (fits the data) and legal (fits the laws of chemistry).

Why Should You Care?

This is a big deal for science and industry.

• For Medicine: We could predict how bacteria causing infections will grow in different parts of the body (which have different nutrients) and figure out how to stop them.
• For Industry: If you want to use bacteria to make biofuels or medicines, you need to know exactly how they will behave in a giant tank. dAMN lets scientists simulate thousands of different "recipes" for the bacteria's food to find the perfect one, without having to run expensive and slow experiments for every single possibility.

In a nutshell: dAMN is a super-smart computer program that learns from real bacteria but is forced to obey the laws of physics. It can look at a bowl of food and accurately predict how a bacterial city will wake up, eat, grow, and switch diets, even if it's never seen that specific bowl of food before.

Technical Summary

1. Problem Statement

Predicting bacterial growth dynamics across diverse nutrient environments remains a central challenge in systems biology. Existing approaches suffer from specific limitations:

• Kinetic Models: While mechanistically accurate, they require extensive parameterization and are not scalable to genome-scale networks.
• Standard Flux Balance Analysis (FBA): Offers scalability and stoichiometric consistency but assumes steady-state, preventing the simulation of dynamic batch processes (e.g., substrate depletion over time).
• Dynamic FBA (dFBA): Extends FBA to temporal processes but often produces biologically unrealistic instantaneous flux shifts, struggles with numerical stiffness, and typically fails to capture the lag phase (the adaptation period before growth begins).
• Existing Hybrid/ML Models: Recent attempts to combine machine learning with metabolic models (e.g., PINNs, Neural-ODEs, AMNs) often fail to simultaneously predict full growth curves, handle unseen media compositions, or explicitly model lag phases without requiring rich omics data or fixed substrate sets.

2. Methodology: The dAMN Architecture

The authors propose dAMN (dynamic Artificial Metabolic Networks), a hybrid framework that integrates data-driven neural networks with mechanistic constraints from genome-scale dFBA.

Core Architecture

The model operates on discrete time steps ($t$) and consists of two main components:

1. Neural Network Components:
   • Lag Phase Predictor: A feedforward neural network ($NN_{lag}$) takes the initial medium composition ($C_{init}$) as input and predicts the lag time ($t_{lag}$) and the stiffness ($k_{lag}$) of the adaptation phase.
   • Flux Predictor: A second feedforward neural network ($NN_{flux}$) predicts reaction fluxes ($v(t)$) based on the current state (time and concentrations).
2. Mechanistic Constraints (dFBA):
   • The predicted fluxes are constrained by the stoichiometric matrix ($S$) derived from a genome-scale metabolic model (e.g., E. coli iML1515, P. putida iJN1463).
   • The system enforces mass balance: $S \cdot v(t) = 0$ for internal metabolites.
   • A transport matrix ($R$) maps reaction fluxes to changes in extracellular metabolite concentrations.

Mathematical Formulation

The model simulates growth using the following logic:

• Lag Phase Function: An exponential function $f(t)$ modulates growth, starting at 0 and reaching 1 after $t_{lag}$. This prevents instantaneous growth at $t=0$.
• Concentration Update: The concentration of metabolites at time $t$ is updated based on the previous concentration, the lag function, and the fluxes:
  $$C(t) = C(t-1) + f(t) \cdot R \cdot v(t) \cdot \Delta t$$
• Training Objective: The model is trained using a multi-objective loss function that minimizes the error between predicted and measured concentrations while enforcing physical constraints:
  • Data Fit: Minimize difference between predicted and experimental concentrations.
  • Biomass Monotonicity: Ensure biomass does not decrease over time.
  • Stoichiometry: Enforce $S \cdot v(t) = 0$.
  • Irreversibility: Ensure all fluxes are non-negative.
  • Weighting: Loss weights decay exponentially over time to prioritize early-stage fitting.

Datasets

The model was trained and tested on three datasets:

1. M28 (E. coli): 280 different media conditions with various combinations of sugars, amino acids, and nucleobases.
2. Putida (P. putida): 81 different media conditions.
3. Millard (E. coli): Time-series data for glucose, acetate, and biomass to test substrate switching (diauxie).

Two evaluation strategies were used:

• Forecast Set: Training on the first 2/3 of time points, predicting the final 1/3.
• Media Set: Training on 2/3 of media conditions, predicting growth on completely unseen media combinations.

3. Key Contributions

• Hybrid Neural-Mechanistic Framework: Successfully integrates neural networks for dynamic parameter inference (fluxes, lag times) with the rigorous stoichiometric constraints of genome-scale models.
• Explicit Lag Phase Modeling: Unlike standard dFBA, dAMN learns the lag phase duration and shape directly from nutrient composition, enabling realistic adaptation dynamics.
• Generalization to Unseen Media: The model can predict full growth curves for novel combinations of nutrients without retraining, a capability lacking in many previous hybrid models.
• Emergent Substrate Dynamics: The model infers realistic substrate consumption and overflow (e.g., acetate production) solely from biomass constraints, without explicit supervision on intermediate metabolite dynamics.
• Minimal Data Requirement: Requires only initial medium composition and biomass measurements, avoiding the need for extensive kinetic parameters or omics data.

4. Results

• Predictive Accuracy:
  • Forecasting: Achieved a median $R^2 \approx 0.98$ for E. coli and $0.98$ for P. putida when predicting future time points.
  • Generalization: Achieved a median $R^2 \approx 0.96$ (E. coli) and $0.94$ (P. putida) when predicting growth on unseen media.
• Biological Phenomena Reproduction:
  • Successfully reproduced acetate overflow and the glucose-acetate consumption shift (diauxie) in E. coli, even when trained only on biomass and glucose data.
  • Accurately predicted substrate depletion curves that mirrored biomass accumulation.
• Comparison with PINNs:
  • When compared to a Physics-Informed Neural Network (PINN) based on a specific ODE model, dAMN outperformed the PINN unless the PINN was provided with perfect kinetic parameters. dAMN achieved high accuracy with fewer assumptions.
• Robustness: Performance remained high ($R^2 > 0.85$) even when training data was reduced by half, though it struggled slightly with undetectable lag phases or sharp stationary-phase drops.

5. Significance

dAMN represents a significant advancement in computational systems biology by bridging the gap between purely data-driven models and mechanistic constraint-based models.

• Practical Application: It enables the rapid prediction of bacterial growth in complex, combinatorial nutrient environments, which is crucial for bioprocess engineering, synthetic biology, and metabolic engineering.
• Scientific Insight: The ability to learn lag phases and diauxic shifts from minimal data suggests that the stoichiometric constraints of the genome-scale model, when coupled with flexible neural networks, are sufficient to capture complex regulatory behaviors without explicitly modeling gene regulation.
• Scalability: By leveraging existing genome-scale models (GEMs), the approach is scalable to different organisms and does not require the manual curation of kinetic parameters for every reaction.

The source code, models, and data are publicly available, facilitating further adoption and extension of the framework.