Contents

Pages

Contact Griffin Labs

Previous innovation Fundamental AI 5 min read

AI for metabolic simulations.

Neural differential equations, published as an arXiv preprint in December 2025.

Highlights

  1. 90% of drugs fail in human trials, 30% of them because of unforeseen side effects.
  2. Existing metabolic models are limited in causal discovery and accuracy.
  3. Our neural differential equations simulate metabolic pathways with over 90% lower error and 1000× faster inference.

1 The problem

90% of drugs fail in human trials.

30% of them fail because of unforeseen side effects. This is an incredibly costly endeavor: big pharma reports that one failed drug costs $1 billion.

We believe that through more accurate computational models of disease, we can pick better drug targets by predicting their side effects early. This would increase the success rate of all the following stages in the drug development pipeline.

Figure 1

Probability of success across the drug development pipeline

a Today

Probability of success Target identification In vitro testing Animal trials Human trials

b With better target identification

Probability of success Target identification In vitro testing Animal trials Human trials
Figure 1. Schematic. Picking better drug targets at the first stage, by predicting their side effects early, raises the success rate of every following stage.

2 Existing methods

The limits of existing metabolic models.

1 Genome-scale metabolic models
  • Limited causal discovery, since they are based on knowledge bases of isolated interactions.
  • Limited by the steady-state assumption (a stable flux of 0), which cannot capture disease progression from a healthy state, since that involves flux changing across time steps.
2 Equations derived by hand
Very poor modeling accuracy, as Costello and Martin (2018) showed.

3 Our approach

Neural differential equations for metabolic simulations.

Toward the vision of accurate computational models for human biology, we developed neural differential equations (NDEs) for metabolic simulations.

On time-series multiomics data from engineered E. coli strains, our NDEs cut prediction error by more than 90% on both the limonene and isopentenol pathways, and ran inference 1000× faster than the existing machine-learning baseline.

Figure 2

Predicted metabolite dynamics on held-out strains

a Acetyl-CoA

Limonene pathway

Acetyl-CoA over time, limonene pathway: the NDE prediction follows the shape of the measured curve, while the baseline falls steadily below zero.

b HMG-CoA

Isopentenol pathway

HMG-CoA over time, isopentenol pathway: the NDE prediction closely tracks the measured rise and fall, while the baseline rises in a straight line.

c Isopentenol

Isopentenol pathway

Isopentenol over time: the NDE prediction closely tracks the measured production curve, while the baseline overshoots it.
Figure 2. NDE predictions (orange, dashed) against measured concentrations (blue) on held-out medium-producing strains, with the machine-learning baseline of Costello and Martin (2018) in green. Time in interpolated units. From Habaraduwa and Lixandru (2025), Figure 3.

Figure 3

A learned vector field

Learned vector field of Acetyl-CoA against HMG-CoA: black streamlines flow toward a central diagonal band of fast dynamics, shown in red, with slower regions in blue.
Figure 3. A (toy) learned vector field in two dimensions. Color shows the magnitude of the field: red marks faster dynamics, blue slower changes. Black streamlines show the direction in which the system evolves. From Habaraduwa and Lixandru (2025), Figure 2.

4 The team

The team

Andrei Lixandru on LinkedIn

Project lead

Coordinated the NDEs project and co-authored the preprint.

AI scientist trained in leading AI and computational biology labs, with a track record of innovation in machine learning research (FrODO, axPPO, NDEs).

Udesh Habaraduwa on LinkedIn

Machine learning researcher

Turned NDEs for metabolic simulations from an intuition into a rigorous scientific study, showing the promise of physics-informed neural networks for simulations in biomedicine.

PhD researcher and host of the One Deeper podcast.

Contact

Get in touch.

For questions about NDEs for metabolic simulations or the preprint.