terminalBio's unified molecular intelligence engine — BindFM, XUZU, RNJ — is being directed at the hardest open problems across five research domains. One compute substrate. Many frontiers.
Cancer is, at its most fundamental level, a molecular recognition failure — misfolded proteins escape immune surveillance, oncoproteins hijack signaling cascades, and cancer cells evolve resistance by altering binding interfaces faster than conventional drugs can follow.
terminalBio's approach applies BindFM's cross-modal binding prediction to the specific challenge of oncology: predicting how mutant oncoproteins differ in their binding behavior from wild-type, mapping the full interaction surface of cancer driver mutations, and identifying de novo small-molecule or aptamer inhibitors using XUZU's generative engine.
Resistance emerges at the binding interface. If you can model every binding interaction as one continuous function, you can predict resistance mutations before they appear in the clinic.
The RNJ model — which achieved Kaggle Silver (89th of 1867 globally) for RNA 3D structure prediction — directly applies here: oncogenic RNA folds, miRNA-mRNA interactions in cancer gene regulation, and the structural basis of aptamer-based cancer targeting are all tractable within the same framework that solved competitive RNA folding.
Current focus areas include: KRAS G12C binding pocket modeling, aptamer design against HER2 and PD-L1 extracellular domains, and RNA-based regulatory network mapping in triple-negative breast cancer.
Drug discovery is a binding problem. Every step in the pipeline — hit identification, lead optimization, selectivity profiling, ADMET prediction — reduces to predicting or engineering molecular recognition events across the protein-ligand interface.
BindFM was designed with this pipeline in mind. Its protein-small molecule binding head is trained on binding affinity datasets across multiple assay types (Kd, Ki, IC50, kon/koff), giving it sensitivity to the kinetic dimensions of drug binding that static docking programs miss.
A drug doesn't just bind — it binds at a rate, for a duration, with a selectivity profile. The next generation of binding models must predict all three simultaneously.
Beyond prediction, XUZU's generative OT-CFM (optimal transport conditional flow matching) decoder can be directed toward small-molecule generation conditioned on a target binding pocket — enabling fragment-growing and scaffold-hopping without enumerating chemical space explicitly.
Cross-modality is critical here: aptamers increasingly compete with small molecules in therapeutic applications. XUZU enables the same design loop — target structure in, candidate binder out — across both chemical and nucleic acid space, from one interface.
The proteome and the genome are not independent entities. They are the same molecular system observed at different levels of abstraction. Proteins bind DNA at transcription factor binding sites, RNA polymerase reads DNA templates, ribosomes translate mRNA — every cellular process is a series of nucleic acid-protein binding events.
BindFM's protein-nucleic acid binding head was designed precisely for this cross-domain landscape. The same weights that predict aptamer-protein affinity also predict DNA-binding domain interactions, transcription factor occupancy, and CRISPR-Cas9 PAM recognition events.
The proteome and genome are the same physics problem at different scales. A unified binding model doesn't need to be told which level it's operating at — it reads the atoms.
In genomics applications, RNJ's RNA 3D structure prediction capability maps directly to riboswitch function prediction, guide RNA secondary structure validation, and splicing regulatory element identification — all of which depend critically on RNA fold accuracy.
For proteomics, BindFM's protein-protein interaction head enables interactome-scale binding prediction — predicting which protein pairs interact, at what affinity, and how disease mutations alter the interaction network. This is the foundation for systems-level drug target discovery.
terminalBio is, at its core, a computational biology research organization. The models — BindFM, XUZU, RNJ — are as much research contributions as they are tools. Each architectural decision is a hypothesis about the geometry of molecular space.
The central hypothesis: SE(3)-equivariant graph neural networks operating directly on atomic coordinates are the correct inductive bias for any molecular prediction task. Not protein language models, not evolutionary sequence statistics — the 3D atomic graph.
The architecture is the hypothesis. EGNN shared weights across protein and RNA carbons are not an engineering shortcut — they are a claim about what binding is, at the physical level.
Active architectural research includes: scaling laws for equivariant molecular models (how does prediction accuracy scale with parameter count and training set size?), the design of FEBI (Fundamental Encoding Block Intelligence) and RJ (Relational Junction) — XUZU's proprietary transformer and graph attention blocks — and the GER (Generate-Evaluate-Refine) self-improvement loop that uses REINFORCE to push generated sequences toward lower predicted Kd.
TB Biocompute is the abstraction layer above the model weights — the API, evaluation harness, and model composition framework that allows models trained on different binding types to be composed into multi-step prediction pipelines.
Aging biology is, at the molecular level, a story about binding events going wrong slowly. Proteins misfold and aggregate. DNA repair machinery loses fidelity. Epigenetic marks drift from their regulatory positions. Senescent cells sustain SASP (Senescence-Associated Secretory Phenotype) signaling by altering their cytokine binding landscape.
Every one of these processes is a change in molecular binding behavior — and therefore addressable with a sufficiently general binding prediction framework.
Aging is entropy accumulating at binding interfaces. If you can model every binding interaction, you can model aging — and intervene at the molecular level before the phenotype becomes irreversible.
BindFM applies to aging research in several ways: predicting how amyloid beta and tau protein aggregation changes PPI networks in neurodegeneration, modeling how sirtuins and NAD+-binding enzymes change their kinetics as NAD+ levels decline with age, and predicting which senolytics disrupt senescence-sustaining PPI interactions.
XUZU's aptamer design capability is particularly relevant here: aptamers against aggregation-prone proteins (alpha-synuclein, amyloid beta, tau) represent a growing therapeutic modality for neurodegenerative aging pathologies. The zero-shot design capability — no SELEX required — dramatically lowers the experimental barrier to testing new candidates.
RNJ contributes through telomerase RNA structure modeling and the prediction of age-associated changes in RNA secondary structure stability — a component of the RNA degradation that accompanies cellular aging.