A Turing-Style Test for In-Silico Antibodies: How Sampling Mode Makes WGAN-GP Beat VAE in the Wet Lab
- 1Moderna, Inc.
Abstract
Recently, it has become feasible to generate antibodies in silico using AI-based approaches such as deep learning, natural language processing, and diffusion models. This opens the door to computational antibody design as a complement to laboratory-based methods (animal immunization, hybridoma technology, and molecular display) for biologic drug discovery. However, when proposing human antibody sequences, or libraries thereof, it remains essential to determine whether they can be expressed, purified, and biophysically characterized using standardized laboratory assays typically applied early in discovery campaigns. In this work, we devised a Turing imitation game inspired experiment to compare two deep learning methods, Generative Adversarial Networks (GANs) and Variational Autoencoders (VAEs), for generating experimentally viable de novo antibody sequences. Under the sampling protocols tested here, a Wasserstein GAN with gradient penalty produced sequences that were experimentally validated at a markedly higher rate (92/93; 99%) than those from a VAE (2/40; 5%). Importantly, the two sets of experimental candidates were not sampled identically: GAN sequences were drawn unconditionally from the trained generator, whereas VAE sequences were generated from latent seeds corresponding to marketed antibodies not seen during training. This methodological asymmetry, rather than model architecture alone, largely accounts for the contrasting outcomes, and when the VAE was instead seeded from its training distribution its in-silico developability profiles were comparable to the GAN's. Our findings highlight both the promise of computationally driven discovery of antibody-based biotherapeutics and the decisive role of sampling strategy in determining experimental success.
Related
- What makes a meta-analysis believable?
- What Makes a Child Go Bad? Genes, Choice, and Punishment in Original Sin
- When location makes oxidative stress lethal: mitoxyperilysis at the immunometabolic stress interface
- Flexible compression of sensory evidence makes human decisions more resilient to noisy neural computations
- From Turing Patterns to Spike Self-Replication and Nucleation in a Growing-Domain Tulip Breaking Virus Model
BibTeX
@misc{kummer2026turingstyle,
title = {{A Turing-Style Test for In-Silico Antibodies: How Sampling Mode Makes WGAN-GP Beat VAE in the Wet Lab}},
author = {A. Kummer and F. Mahmoudinobar and W. Liu and J. W. Davis and E. J. Ma and S. Kumar},
howpublished = {bioRxiv},
year = {2026},
month = sep,
doi = {10.64898/2026.09.11.750936},
url = {https://www.biorxiv.org/content/10.64898/2026.09.11.750936v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - A Turing-Style Test for In-Silico Antibodies: How Sampling Mode Makes WGAN-GP Beat VAE in the Wet Lab
AU - A. Kummer
AU - F. Mahmoudinobar
AU - W. Liu
AU - J. W. Davis
AU - E. J. Ma
AU - S. Kumar
PY - 2026
DA - 2026/09/11
PB - bioRxiv
DO - 10.64898/2026.09.11.750936
UR - https://www.biorxiv.org/content/10.64898/2026.09.11.750936v1
AB - Recently, it has become feasible to generate antibodies in silico using AI-based approaches such as deep learning, natural language processing, and diffusion models. This opens the door to computational antibody design as a complement to laboratory-based methods (animal immunization, hybridoma technology, and molecular display) for biologic drug discovery. However, when proposing human antibody sequences, or libraries thereof, it remains essential to determine whether they can be expressed, purified, and biophysically characterized using standardized laboratory assays typically applied early in discovery campaigns. In this work, we devised a Turing imitation game inspired experiment to compare two deep learning methods, Generative Adversarial Networks (GANs) and Variational Autoencoders (VAEs), for generating experimentally viable de novo antibody sequences. Under the sampling protocols tested here, a Wasserstein GAN with gradient penalty produced sequences that were experimentally validated at a markedly higher rate (92/93; 99%) than those from a VAE (2/40; 5%). Importantly, the two sets of experimental candidates were not sampled identically: GAN sequences were drawn unconditionally from the trained generator, whereas VAE sequences were generated from latent seeds corresponding to marketed antibodies not seen during training. This methodological asymmetry, rather than model architecture alone, largely accounts for the contrasting outcomes, and when the VAE was instead seeded from its training distribution its in-silico developability profiles were comparable to the GAN's. Our findings highlight both the promise of computationally driven discovery of antibody-based biotherapeutics and the decisive role of sampling strategy in determining experimental success.
ER -