SpeciefAI

Therapeutic antibodies and nanobodies are increasingly delivered as mRNA, letting the host organism produce the protein in vivo. This shifts a key design constraint from protein space into nucleotide space: the encoding mRNA must both express efficiently in the target species and translate into a protein sequence that resembles that species' natural antibody repertoire, so as to minimize immunogenicity. Conventional humanization and species-adaptation tools operate on protein sequence and do not address the mRNA-level requirements of this delivery modality.

SpeciefAI, developed by Dominik Grabarczyk, Mikolaj Kocikowski, Shay B. Cohen, Javier Antonio Alfaro and colleagues at the University of Edinburgh (with collaborators at the University of Gdańsk and University of Calgary), is a transformer-based generative model that works directly in mRNA space. Posted to bioRxiv in March 2026, it generates novel framework regions (FRs) tailored to a given set of complementarity-determining regions (CDRs), harmonizing antibody and nanobody sequences for a chosen species while preserving the CDRs that define binding.

By learning the correspondence between FRs and CDRs at the mRNA level across six species, SpeciefAI couples species-specific protein nativeness with codon-level expressibility in a single generative framework.

#Key Features

• mRNA-space generation: Operates directly on mRNA sequences, simultaneously optimizing for native-like protein framework regions and efficient codon usage for the target species.
• CDR-conditioned framework design: Generates new framework regions conditioned on input CDRs, adapting an antibody or nanobody to a species while retaining its binding loops.
• Multi-species coverage: Learns FR-CDR correspondences across six species, supporting cross-species harmonization beyond human-only humanization.
• High species-specificity: Achieves strong nativeness scores (e.g., T20 of 0.95 for human and canine), indicating generated frameworks closely match natural species repertoires.
• Codon fidelity: Produces sequences with a mean absolute difference in codon adaptation index (CAI) of 0.013 relative to natural sequences, supporting expressibility in the host.

#Technical Details

SpeciefAI is a transformer-based language model trained on antibody and nanobody sequences represented as mRNA, learning the joint distribution over framework regions and codon usage conditioned on input CDRs across six species. Generated sequences closely track the distribution of natural sequences, with reported nativeness reflected in T20 scores up to 0.95 for human and canine outputs and a mean absolute CAI difference of 0.013 from natural references. The codebase, implemented primarily in Python with supporting tokenizer vocabularies and sample datasets, supports pretraining, finetuning, and sampling workflows. Code is available on GitHub and associated data and model artifacts (including canine data) are deposited on Zenodo under a CC-BY-4.0 license.

#Applications

SpeciefAI targets antibody engineers and therapeutic-discovery teams developing mRNA-encoded antibodies and nanobodies, particularly where cross-species adaptation matters — including veterinary therapeutics (e.g., canine) alongside human applications. By generating species-tailored frameworks around fixed CDRs, it supports humanization and broader species harmonization, library design, and preparation of candidates for in vivo mRNA expression with reduced immunogenicity risk.

#Impact

By formulating antibody species-adaptation as an mRNA-level generative task, SpeciefAI bridges protein nativeness and nucleotide expressibility that protein-only humanization tools handle separately, and extends adaptation beyond human to multiple species. Its open code and deposited data make it a practical starting point for mRNA antibody and nanobody design across species. As a preprint, its reported metrics await peer review and independent validation, and the deposited release reflects an early-stage public artifact.