FlashABB

FlashABB is a pretrained antibody structure prediction model developed by the Oxford Protein Informatics Group (OPIG), Charlotte M. Deane's group at the University of Oxford, and released as a bioRxiv preprint in June 2026. It predicts full paired heavy/light antibody 3D structures directly from sequence, and does so faster than protein language models (PLMs) can produce their embeddings — the headline claim captured by the preprint's title, "Modelling antibody structures at the speed of language."

The central problem FlashABB addresses is throughput. Antibody discovery and engineering increasingly operate at repertoire scale, where immune sequencing campaigns yield millions of paired sequences. Conventional structure predictors — and even the sequence-only PLMs often used as fast proxies — become a computational bottleneck at this volume, forcing practitioners to subsample or rely on sequence features alone. FlashABB is designed so that explicit 3D structure prediction is no longer the slow step, making structure-aware analysis tractable across entire repertoires.

The model's key architectural contribution is Flashpoint Attention, a fast, linear-memory analogue of the Invariant Point Attention (IPA) module that underpins AlphaFold2-style structure modules. By reformulating the geometric attention used for structure refinement to scale more favourably in memory and time, FlashABB collapses the cost gap between generating a structure and generating an embedding. It is distributed as a fixed pretrained checkpoint, loadable in a single line, lowering the barrier for routine use.

#Key Features

• Flashpoint Attention: A fast, linear-memory reformulation of Invariant Point Attention that delivers the geometric reasoning of IPA-based structure modules without its quadratic memory and runtime overhead, enabling structure prediction at PLM-level speed.
• Paired heavy/light prediction: Predicts complete paired variable-domain (Fv) structures from heavy|light sequence input, outputting full-atom and backbone coordinates that can be written directly to PDB files.
• Repertoire-scale throughput: Demonstrated on millions of sequences, making explicit structure prediction practical for whole-repertoire analyses rather than only curated subsets.
• Fixed pretrained checkpoint: Ships as a ready-to-use model loaded via from flash_abb import pretrained, with no training or fine-tuning required to obtain structures.
• Downstream developability and embeddings: Companion components include FlashTAP for Therapeutic Antibody Profiler-style developability scores (PSH, PPC, PNC, SFvCSP) and FlashABB-SSS for per-residue structure-aware embeddings that combine sequence and predicted 3D context.

#Technical Details

FlashABB is a transformer-based structure predictor whose structure module replaces standard Invariant Point Attention with Flashpoint Attention, an analogue engineered for linear memory scaling so that the per-sequence cost of producing coordinates is comparable to — and in the authors' reported regime faster than — generating embeddings from a protein language model. The model takes paired antibody sequences in heavy|light format and emits 3D coordinates and backbone coordinates. It is provided as a fixed pretrained checkpoint (pip install flash-abb), accessed through pretrained(), with sibling checkpoints pretrained_tap() for developability scoring and pretrained_sss() for structure-aware sequence embeddings. The accompanying preprint demonstrates the model at repertoire scale — applying it across datasets of millions of antibody sequences to drive downstream stability and developability prediction.

#Applications

FlashABB targets antibody discovery and therapeutic engineering workflows where structural context is valuable but throughput has historically forced compromises. By making per-sequence structure prediction as cheap as embedding generation, it enables structure-aware triage, stability and developability assessment, and liability screening across entire immune repertoires rather than hand-picked candidates. The FlashTAP and FlashABB-SSS components let teams move directly from raw paired sequences to developability scores and structure-aware representations, supporting candidate ranking, antibody humanization, and large-scale repertoire mining for protein engineers and computational immunologists.

#Impact

FlashABB extends OPIG's established line of antibody-specific tools (such as AbLang and ABodyBuilder) by attacking the speed dimension directly: its Flashpoint Attention demonstrates that the geometric attention central to modern structure prediction can be made memory- and runtime-efficient enough to operate at language-model speeds. This reframes structure prediction from an occasional, curated step into a routine, repertoire-scale primitive, opening the door to structure-conditioned analyses over the full output of immune sequencing campaigns. As a recent preprint, its benchmark standing and adoption remain to be established, and its scope is currently limited to paired heavy/light antibodies rather than nanobodies or general proteins.