Ångström AI recently launched!
"Accelerating molecular simulation using generative AI."
TLDR: They build fast and experimentally accurate generative AI-based simulations of molecular interactions for pharma and biotech companies. These simulations can tell us whether a drug is going to bind to a protein or how quickly a drug will act once it is consumed by a patient. Replace wet lab experiments with gen AI molecular simulations.
Founded by Javier Antoran, Laurence Midgley, José Miguel Hernández-Lobato & Gabor Csanyi
The Problem
During drug discovery and development, pharmaceutical companies need to understand how drug molecules interact with other molecules in the human body to determine the drugs’ efficacy and safety.
Conventional methods to assess molecular interactions are unsatisfactory:
- Wet lab experiments are accurate but slow and expensive.
- Machine learning prediction methods, like AlphaFold, are fast but inaccurate. They are limited by the quality and quantity of training data, which must be generated by lab experiments.
- Molecular Dynamics simulates interactions by rolling out the equations of physics, offering a balance between wet lab and machine learning predictions in accuracy and speed. These factors depend on the model of physics used. More accurate models are more expensive to run. This makes the method compute bottlenecked rather than data bottlenecked.
Here is a video of a simulation of 64 water molecules ⤵️
https://www.youtube.com/watch?v=18AuJEVnwWA
Solution: Fast and Accurate Molecular Simulations
Ångström AI runs molecular simulations, keeping them in the compute-constrained regime, but they combine 1) quantum mechanically accurate models of physics 2) generative AI that allows them to run these models quickly.
- They use the MACE (multi-atomic cluster expansion) physics model, which accurately reproduces quantum-mechanical interactions. It was developed by Gabor, their co-founder. Ångström AI recently showed MACE simulations are the first-ever to provide accuracy comparable to lab experiments when estimating the hydration-free energies (a quantity relevant to drug bioavailability). Below is a plot from their recent publication. However, MACE is computationally expensive; each of the results from the below plot required 1 week of compute on 8 A100 GPUs.
- They use diffusion models to accelerate MACE simulations, making them computationally affordable. Their models generate states consistent with physics, but the transitions between states are non-physical and significantly faster. Here is a video of their AI simulating the same water box as above.
https://www.youtube.com/watch?v=aK1MNLsQg2Y
Demos!
Here are a couple of examples to show off the type of stuff you can do with their generative models.
Speeding up Supercool Water
Supercool water—liquid water below freezing, here at -40°C—is notoriously difficult to simulate with traditional methods because cold molecules move slower. The plot below shows the correlation between water molecule orientations across simulation steps. Their AI introduces about 10,000 times more information per step compared to traditional simulations.
Hydrating Methane
Modeling interactions between molecules and water allows them to calculate how quickly the molecules will dissolve and their bioavailability as drugs. Here are videos of a traditional simulation of a methane molecule interacting with water and their AI simulation. The startup's AI prioritizes the movement of the methane molecule and its surrounding water, which are the parts that matter for solubility and bioavailability calculations.
These are the first ever genAI accelerated, physically accurate molecular dynamics simulations incorporating the interaction of many molecules. Ångström AI is scaling up - so stay tuned!
