FcGPT

The fragment crystallizable (Fc) domain of an antibody governs its effector functions by engaging a panel of Fc receptors, and tuning this engagement is central to therapeutic antibody design. Naturally occurring and rationally engineered Fc variants cover only a small portion of the possible functional landscape, and enumerating variants that hit a precise, user-specified combination of receptor activities is difficult by rational design alone. FcGPT approaches this as a generative sequence-design problem grounded in large-scale experimental data.

The work pairs high-throughput protein engineering with deep learning. The authors performed deep mutational scanning across the entire human IgG1 Fc domain using yeast display, building a combinatorial library of more than 10^8 Fc variants, sorting it against a panel of eight canonical Fc receptors, and deep sequencing the resulting populations. This produced millions of unique Fc sequences annotated with their receptor-binding profiles, which were used to train classifiers that predict binding activity from sequence.

FcGPT itself is a domain-specific autoregressive protein language model pretrained on over three million unique Fc sequences and refined by post-training reinforcement learning with experimental feedback (RLXF). It was developed in Sai Reddy's lab at ETH Zurich and released as a preprint in October 2025.

#Key Features

• Domain-specific pretraining: FcGPT is trained specifically on more than three million Fc sequences, concentrating model capacity on the functional domain being engineered rather than the full antibody or proteome.
• Programmable multi-receptor design: The model generates Fc variants with user-defined binding profiles across a panel of eight Fc receptors, enabling synthetic effector-function specifications.
• Reinforcement learning from experimental feedback: RLXF post-training uses a group-relative policy optimization objective with sequence-based verifiers to steer generation toward target profiles at scale.
• Experimentally grounded dataset: Training data derive from deep mutational scanning of the human IgG1 Fc domain sorted against real receptor binding, rather than from computational labels alone.

#Technical Details

FcGPT is an autoregressive transformer protein language model pretrained on over three million unique Fc sequences drawn from a deep mutational scanning campaign of the human IgG1 Fc domain. The underlying library exceeded 10^8 variants and was sorted by binding to eight canonical Fc receptors, with deep sequencing yielding millions of profile-annotated sequences. Post-training applies reinforcement learning with experimental feedback (RLXF) using a group-relative policy optimization objective and synthetic verifier rewards, allowing the model to be optimized toward specified receptor engagement without exhaustive experimental screening.

#Applications

FcGPT is intended for antibody engineers and immunologists who need to program the effector functions of therapeutic antibodies. By requesting a target combination of Fc-receptor activities, users can generate candidate Fc variants for effector enhancement, silencing, or bespoke immune-cell recruitment, providing a computational starting point that reduces the experimental search space for antibody optimization.

#Impact

FcGPT demonstrates how coupling large-scale deep mutational scanning with a domain-specific generative language model and reinforcement learning can turn antibody Fc engineering into a programmable design task. By spanning a functional space far larger than the set of known Fc variants and aligning generation to experimentally measured receptor binding, it offers a foundational tool for understanding and programming antibody-mediated immunity. As a preprint, its capabilities await peer review and independent validation.