TerraBind

TerraBind is a protein-ligand foundation model for structure and binding affinity prediction developed by Terray Therapeutics, the AI-driven drug discovery company headquartered in Monrovia (Los Angeles), California. Introduced in February 2026 alongside an arXiv preprint accepted to ICML 2026, it is the model behind Terray's "EMMI Predict" product. TerraBind targets a core bottleneck in computational drug discovery: the all-atom diffusion modules used by recent co-folding models such as AlphaFold 3, Chai-1, and Boltz-2 are accurate but computationally expensive, which makes them slow for the large-scale virtual screening campaigns that early-stage discovery depends on.

The central premise of TerraBind is that full atomic resolution is not necessary to predict binding affinity well. Instead of generating all-atom coordinates with a diffusion process, the model learns over a coarse-grained representation — protein residue C-beta atoms and ligand heavy atoms only — and maps that representation directly to a binding affinity estimate. By removing the diffusion-based coordinate generation step entirely, TerraBind sidesteps the dominant inference cost of the co-folding paradigm while retaining the geometric reasoning needed for accurate affinity ranking.

The result is a model that Terray reports as roughly 26x faster than Boltz-2 on a single GPU while delivering about 20% higher affinity-prediction accuracy, positioning TerraBind as a practical screening engine for industrial drug discovery rather than a general-purpose structure predictor.

#Key Features

• Diffusion-Free Affinity Prediction: Replaces the all-atom diffusion coordinate-generation module of co-folding models with a direct mapping from coarse-grained structure to binding affinity, eliminating the dominant inference cost.
• Coarse-Grained Structural Representation: Operates on protein C-beta atoms and ligand heavy atoms only, demonstrating that full atomic resolution is unnecessary for strong affinity ranking.
• Two Complementary Variants: TerraBind-Seq combines ESM-2 protein embeddings with COATI-3 molecular representations for a sequence-only setting, while TerraBind-Struct learns coarse structural representations in a Pairformer trunk and maps them directly to affinity.
• Uncertainty Quantification: Incorporates epistemic neural networks (Epinets) to produce calibrated uncertainty estimates, enabling probabilistic acquisition strategies such as Expected Maximum (EMAX) for compound selection.
• Validated Pose Accuracy: Matches diffusion-based baselines on pose-quality benchmarks including PoseBusters and FoldBench despite never running a diffusion sampler.

#Technical Details

TerraBind is a multimodal transformer system. The structure-based variant, TerraBind-Struct, encodes protein C-beta atoms and ligand heavy atoms and processes their pairwise relationships in a Pairformer trunk (the pair-representation architecture popularized by AlphaFold-style models), then regresses binding affinity with a likelihood-based prediction head that yields uncertainty estimates rather than point values. The sequence-based variant, TerraBind-Seq, pairs ESM-2 protein language-model embeddings with molecular representations from COATI-3, Terray's chemical foundation model pretrained on more than one billion small molecules. A streamlined pocket-focused variant restricts inputs to the binding-site region.

On reported benchmarks, TerraBind runs approximately 26x faster than Boltz-2 on a single NVIDIA A6000 GPU. It improves Pearson correlation for binding affinity by roughly 16-20% over Boltz-2 across independent test sets, including a 16% gain on the CASP16 affinity challenge (where it is competitive with the IsoDDE method) and a ~20% average gain across 18 internal drug-discovery targets backed by more than 25,000 assay readouts, outperforming Boltz-2 on 15 of those 18 targets. The Epinet-based EMAX acquisition strategy is reported to deliver roughly a 6x improvement in molecular selection efficiency over greedy selection.

#Applications

TerraBind is designed for industrial structure-based drug discovery, where its speed makes it suited to large-scale virtual screening and hit-to-lead affinity ranking that would be impractical with slower diffusion-based co-folding models. Medicinal chemists can rank candidate compounds against a target rapidly, while the model's calibrated uncertainty estimates support active-learning and probabilistic acquisition workflows that prioritize the compounds most likely to be high-affinity binders. Within Terray's EMMI Predict platform, it is paired with the company's high-throughput chemistry and screening data to drive iterative design-make-test cycles.

#Impact

TerraBind challenges the prevailing assumption that accurate affinity prediction requires all-atom diffusion, showing that a coarse-grained representation combined with strong protein (ESM-2) and chemical (COATI-3) foundation models can match diffusion baselines on pose quality while exceeding them on affinity at a fraction of the compute. Its roughly 26x speedup over Boltz-2 directly addresses the throughput limits that constrain screening campaigns, and acceptance to ICML 2026 signals interest from the broader machine-learning community. Key caveats remain: TerraBind is a commercial product from Terray Therapeutics with model weights not openly released, much of the headline evaluation relies on proprietary internal targets that are difficult to reproduce independently, and the model is specialized for affinity and pose ranking rather than the broad multi-entity structure prediction offered by generalist co-folding models.