terminalBio is building the foundational layer for molecular intelligence — open models that treat every atom, sequence, and structure as part of one unified compute substrate.
Read what drives every architectural decision we make — from atom tokenization to shared encoder weights.
Where most AI biology tools fragment the problem into separate protein models, RNA models, and small molecule models, terminalBio builds one model that speaks all of them. The same encoder. The same weights. The same representation space — because at the atomic level, biology is not a collection of domains. It is one problem.
At the atomic level, a protein carbon and an RNA carbon are the same thing. They should live in the same learned space — and in BindFM, they do.
We are building toward a future where any binding interaction — drug-target, aptamer-protein, CRISPR-DNA, antibody-antigen — can be predicted, designed, and optimized in silico. Not five separate tools running on five separate assumptions. One compute substrate. One biology.
Unified TB Biocompute is our answer to the fragmentation problem in computational biology. It is the operating system layer on top of which any molecular intelligence application can be built: drug discovery pipelines, aptamer therapeutics, synthetic gene circuits, rational vaccine design.
We build from first principles. No borrowed encoders from protein language models. No assumptions imported from evolutionary biology that fail at the cross-modality boundary. When a protein carbon passes through our EGNN encoder, it follows the exact same learned transformation as an RNA carbon. This is not a design constraint. It is the entire point.
Everything we build is open source under the MIT license. Science only advances by building on prior work — and that only works when the work is actually accessible. Our models live on GitHub and Hugging Face, runnable in a free Colab notebook in under forty minutes.
Every model is built from scratch. Inheriting evolutionary protein language model representations introduces the wrong inductive biases at every modality boundary crossing — a carbon atom is not a protein residue.
A protein carbon and an RNA carbon are both carbons. They should live in the same learned embedding space. In BindFM they do — 197-dimensional atom tokens processed through identical SE(3)-equivariant transformations.
All models are released under the MIT license. Science advances by building on prior work. That only works if the work is actually accessible — to every lab, everywhere, regardless of compute budget.
TB Biocompute is an OS-layer abstraction: one substrate on which drug discovery, aptamer therapeutics, synthetic circuits, and rational vaccine design all run — interoperably, by design.
SE(3)-equivariance is not an optional feature — it is a physical constraint. Our EGNN encoder respects the rotational and translational symmetry of 3D molecular systems at every layer.
Prediction is the first step. The end goal is generation — de novo design of binders, aptamers, and therapeutics that have never existed, optimized toward measurable binding affinity in silico before any wet lab step.
We are building toward a world where any molecular binding interaction can be predicted, designed, and optimized entirely in silico. Not five separate specialized tools — one compute layer.
Building terminalBio from first principles — every architectural decision driven by the conviction that molecular biology is one physics problem, not a taxonomy of domain-specific benchmarks. Kaggle Silver Medalist (89th of 1867 teams, Stanford RNA 3D Structure competition, March 2026).