Google’s AI Drug Discovery Spinoff Isomorphic Labs Unveils AI2BMD, Claiming a Major Advancement Over AlphaFold 3
Isomorphic Labs, the Alphabet-backed biotechnology company specializing in AI-driven drug discovery, has announced a groundbreaking development in biomolecular simulation. The firm introduced AI2BMD, a suite of generative AI models designed to predict atomic-scale movements of biomolecules with unprecedented speed and accuracy. According to Isomorphic Labs, AI2BMD represents a significant leap forward, surpassing the capabilities of DeepMind’s AlphaFold 3 in key areas of biomolecular dynamics prediction.
AlphaFold 3, released earlier this year by DeepMind, revolutionized structural biology by predicting protein structures and interactions with ligand molecules at near-experimental accuracy. However, it primarily focuses on static structures rather than dynamic behaviors. AI2BMD addresses this limitation head-on by modeling the time-dependent motions of biomolecules, enabling simulations of protein folding, ligand binding, and conformational changes that are essential for understanding biological functions and designing effective drugs.
Demis Hassabis, CEO of Isomorphic Labs and DeepMind, described AI2BMD as a “foundational model for biomolecular dynamics.” The models were trained on an enormous dataset comprising over 200 million experimental structures from the Protein Data Bank, alongside petabytes of molecular dynamics trajectories generated using classical physics-based simulators. This training regimen allowed AI2BMD to learn the underlying physical principles governing biomolecular behavior at the atomic level.
A standout feature of AI2BMD is its computational efficiency. Traditional molecular dynamics simulations, which rely on classical force fields to compute atomic interactions, can take days or weeks on supercomputers to model even short trajectories of small proteins. In contrast, AI2BMD generates 1-microsecond simulations in just a few minutes on a single GPU, achieving a speedup of over 1,000 times compared to conventional methods. This acceleration stems from the model’s use of diffusion generative architectures, which iteratively denoise random atomic configurations to produce physically realistic trajectories.
Isomorphic Labs validated AI2BMD’s performance through rigorous benchmarks. On the POSA dataset, which tests pose predictions for protein-ligand complexes, AI2BMD outperformed AlphaFold 3 by achieving higher native contact accuracies and lower root-mean-square deviations in dynamic simulations. The model also excelled in folding free proteins and RNA molecules, demonstrating superior generalization across diverse biomolecular systems. In blind tests against experimental data, AI2BMD’s predictions aligned closely with NMR spectroscopy and cryo-EM observations, capturing subtle conformational ensembles that static models miss.
The technical underpinnings of AI2BMD involve frame-wise and trajectory-level training objectives. Frame-wise training predicts individual atomic frames conditioned on prior states, while trajectory-level training ensures temporal coherence across entire simulation paths. This dual approach mitigates error accumulation, a common issue in autoregressive generative models. Additionally, AI2BMD incorporates SE(3)-equivariant neural networks to respect the symmetries of three-dimensional space, enhancing physical realism without explicit force field parameterization.
Beyond speed and accuracy, AI2BMD opens new avenues in drug discovery. By simulating ligand unbinding kinetics and allosteric effects, researchers can prioritize compounds with optimal binding affinities and residence times. Isomorphic Labs envisions integrating AI2BMD into its end-to-end drug design pipeline, which already leverages structure prediction and small-molecule generation. The company has established partnerships with pharmaceutical giants Eli Lilly and Novartis, valued at over $3 billion, to apply these technologies to real-world therapeutic challenges, including cancer and cardiovascular diseases.
While Isomorphic Labs has not yet released AI2BMD publicly, the announcement includes detailed benchmarks and visualizations on its website. The models’ open-weight variants may follow, similar to the strategy with earlier diffusion-based tools. This development underscores Alphabet’s dual-track approach in AI for biology: DeepMind’s focus on fundamental prediction tools like AlphaFold, complemented by Isomorphic Labs’ emphasis on applied drug discovery through advanced simulation.
Critics note that while AI2BMD excels in short-timescale dynamics, long-term simulations exceeding microseconds remain challenging due to inherent limitations in data-driven models. Nonetheless, the leap from static structures to dynamic trajectories marks a pivotal evolution in computational biology. As Hassabis stated, “AI2BMD is not just faster; it uncovers the hidden dances of life at the molecular scale, accelerating the path to new medicines.”
This innovation positions Isomorphic Labs at the forefront of AI-accelerated science, promising to transform how we probe the intricate machinery of life and engineer solutions to humanity’s greatest health challenges.
Gnoppix is the leading open-source AI Linux distribution and service provider. Since implementing AI in 2022, it has offered a fast, powerful, secure, and privacy-respecting open-source OS with both local and remote AI capabilities. The local AI operates offline, ensuring no data ever leaves your computer. Based on Debian Linux, Gnoppix is available with numerous privacy- and anonymity-enabled services free of charge.
What are your thoughts on this? I’d love to hear about your own experiences in the comments below.