Manifold Bio's mBER Scale
- Jeff Huber described Manifold Bio's work connecting in-silico, in-vitro and in-vivo loops for AI-driven discovery. - He noted mBER, an open-source antibody design tool, was validated at million-scale with Nvidia collaboration. - The example shows computational antibody design scaling into experimental validation workflows in drug discovery (x.com).
Antibody drug discovery starts with a simple problem: find a protein that sticks to a disease target, then prove it still works outside a computer. Manifold Bio says its open-source mBER system now does that at million-design scale. (github.com, manifold.bio) mBER, short for Manifold Binder Engineering and Refinement, is an antibody binder design framework that uses AlphaFold-Multimer, structure templates, and sequence conditioning to generate antibody-format binders. Manifold released the code on GitHub in October 2025 and describes it as open source under an MIT license. (github.com, manifold.bio) In its preprint, Manifold said it designed two libraries with more than 1 million single-domain antibodies, or VHH binders, across 436 targets and experimentally screened those libraries against 145 targets. The company reported a dataset of more than 100 million protein-protein interactions and said 45% of screened targets produced specific, significant design success. (biorxiv.org, manifold.bio) That scale matters because protein design has often outrun lab validation: software can propose huge numbers of candidates, but wet-lab testing is slower and more expensive. Manifold’s pitch is that it can connect in-silico design, in-vitro screening, and in-vivo testing in one loop instead of treating them as separate steps. (manifold.bio, manifold.bio) The company’s website says its broader platform combines high-throughput in vivo screening with artificial-intelligence-guided design to build tissue-targeted biologic drugs. In December 2025, Manifold also announced a collaboration with Roche focused on antibody-based blood-brain barrier shuttles for neurological diseases. (manifold.bio, manifold.bio) Manifold and Nvidia said in March 2026 that they had entered a joint study on artificial-intelligence-driven protein binder design, using Nvidia’s computational infrastructure alongside Manifold’s experimental platform. Manifold said the mBER work was the first reported use of that experimental approach and framed the Nvidia study as a next step in scaling the same design-and-test cycle. (manifold.bio) Jeff Huber, an executive chair at Manifold, described the effort in a post on X as a system that links computational design with lab and animal testing loops. His example centered on mBER as evidence that antibody design can move from software output into large experimental workflows instead of stopping at a small proof-of-concept set. (x.com, manifold.bio) The main caveat is that the core mBER results are still posted as a bioRxiv preprint, not a peer-reviewed journal paper. But the public code, the reported million-scale screens, and the Nvidia-backed follow-on study give a concrete picture of what Manifold is trying to build: an antibody design loop that runs at the scale of modern machine learning. (biorxiv.org, github.com, manifold.bio)
