A blinded, prospective benchmark of in silico antibody discovery anchored to experimental affinity and developability
20 affiliations
- 1Specifica, an IQVIA Business, Santa Fe, NM, USA. [email protected]
- 2Carterra, Salt Lake City, UT, USA
- 3Sapidyne Instruments, Inc. GmbH, Boise, ID, USA
- 4Mosaic Biosciences, Boulder, CO, USA
- 5GENEWIZ, Azenta Life Sciences, South Plainfield, NJ, USA
- 6Takeda, Inc., Cambridge, MA, USA
- 7Aureka Biotechnologies, Laguna Hills, CA, USA
- 8University of Cambridge, Cambridge, UK
- 9Shenzhen University of Advanced Technology (SUAT), Shenzhen, China
- 10The George Washington University, Washington, D.C., USA
- 11Xencor, Inc., Pasadena, CA, USA
- 12Biomedical Informatics and Data Science, Yale School of Medicine, New Haven, CT, USA
- 13Institute of Computational Life Sciences, Zurich University of Applied Sciences (ZHAW), Zurich, Switzerland
- 14Incyte, Inc., Wilmington, DE, USA
- 15Merck KGaA, Darmstadt, Germany
- 16Bristol-Myers Squibb, Princeton, NJ, USA
- 17Seismic Therapeutic, Watertown, MA, USA
- 18Locksmith Bio, Dover, DE, USA
- 19InstaDeep, London, UK
- 20Scripps Research Institute, San Diego, CA, USA
Abstract
Experimentally validated prospective, blinded benchmarks are needed to separate durable advances from hype in computational antibody design. Here AIntibody, a challenge inspired by the Critical Assessment of Structure Prediction, tests 511 artificial intelligence (AI)-designed or predicted antibodies from 29 organizations on three tasks: in silico affinity maturation from phase 1 sequencing outputs, affinity ranking within heavy-chain complementarity-determining region 3 (HCDR3) clusters of a selection output and CDR design of proteins not included in a selection output. Validated with diverse experimental assays, several groups produced developable antibodies with affinities <100 pM. However, these successes were exceptions that did not transfer across tasks. Affinity-matured antibodies were modeled effectively. Except for one model, predicting high-affinity clones from clustered HCDR3 datasets was worse than random clone picking. Out-of-library design was highly variable for most method submissions, with many failing to outperform standard selections. The AIntibody challenge shows that AI can optimize antibodies in defined, biologically grounded regimes, in addition to highlighting critical gaps including affinity prediction and library-inspired antibody design and cross-task generalization.
Related
- An interpretable open platform for sequence-based antibody developability prediction
- kitAb: an automated framework for assay-specific antibody developability modeling and large-scale landscape analysis
- A structural antibody benchmark for leakage-aware deep-learning evaluation
- 211 A discovery-to-therapeutics platform for functional antibody targeting of complex membrane-embedded tumor antigens: validation with claudin family members
- From Benchmark to Bench: Can Agents Survive Real-World Drug Discovery?
BibTeX
@article{erasmus2026blinded,
title = {{A blinded, prospective benchmark of in silico antibody discovery anchored to experimental affinity and developability}},
author = {M. Frank Erasmus and Daniel Bedinger and Elizabeth Hopkins and Ginger Ferguson and Justine Strickler and Christilyn P. Graff and Samantha R. Summers and Stacy L. Capehart and Joshua D. Slocum and Crystal Richardson and Sumit Kumar and ZhiFei Sun and Yujie Shang and Jixian Zhang and Ming Gu and Lixia Yi and Alon Wellner and Shuangjia Zheng and Wei Lu and Pietro Sormanni and Matthew Greenig and Haiping Zhang and Brendan T. Mann and Mahdi Baghbanzadeh and Ali Rahnavard and Gregory L. Moore and Huaiyu Sun and Ying Ding and Alex Nisthal and Jitendra Kanodia and Matthew J. Bernett and Aurélien Pélissier and Yanjun Shao and María Rodríguez Martínez and Karthik Ramesh and Horacio G. Nastri and Andreas Evers and Anhar Abdelatif and Andrew J. Bordner and Mykola Bordyuh and Lim Heo and Brian A. Kidd and Huseyin Serhat Tetikol and Shuai Wei and Jung-Eun Shin and Ryan Peckner and Leigh Manley and Ajitesh Lunge and Yashas Devasurmutt and Bora Guloglu and Liviu Copoiu and Miles McGibbon and Monica L. Fernández-Quintero and Nitesh Mishra and Sean Callaghan and Olivia M. Swanson and Daniel L. Bader and James A. Ferguson and Sai Sundar Rajan Raghavan and Benjamin Némoz},
journal = {Nature Biotechnology},
year = {2026},
month = aug,
doi = {10.1038/s41587-026-03238-6},
url = {https://doi.org/10.1038/s41587-026-03238-6},
note = {CC BY},
}
RIS
TY - JOUR
TI - A blinded, prospective benchmark of in silico antibody discovery anchored to experimental affinity and developability
AU - M. Frank Erasmus
AU - Daniel Bedinger
AU - Elizabeth Hopkins
AU - Ginger Ferguson
AU - Justine Strickler
AU - Christilyn P. Graff
AU - Samantha R. Summers
AU - Stacy L. Capehart
AU - Joshua D. Slocum
AU - Crystal Richardson
AU - Sumit Kumar
AU - ZhiFei Sun
AU - Yujie Shang
AU - Jixian Zhang
AU - Ming Gu
AU - Lixia Yi
AU - Alon Wellner
AU - Shuangjia Zheng
AU - Wei Lu
AU - Pietro Sormanni
AU - Matthew Greenig
AU - Haiping Zhang
AU - Brendan T. Mann
AU - Mahdi Baghbanzadeh
AU - Ali Rahnavard
AU - Gregory L. Moore
AU - Huaiyu Sun
AU - Ying Ding
AU - Alex Nisthal
AU - Jitendra Kanodia
AU - Matthew J. Bernett
AU - Aurélien Pélissier
AU - Yanjun Shao
AU - María Rodríguez Martínez
AU - Karthik Ramesh
AU - Horacio G. Nastri
AU - Andreas Evers
AU - Anhar Abdelatif
AU - Andrew J. Bordner
AU - Mykola Bordyuh
AU - Lim Heo
AU - Brian A. Kidd
AU - Huseyin Serhat Tetikol
AU - Shuai Wei
AU - Jung-Eun Shin
AU - Ryan Peckner
AU - Leigh Manley
AU - Ajitesh Lunge
AU - Yashas Devasurmutt
AU - Bora Guloglu
AU - Liviu Copoiu
AU - Miles McGibbon
AU - Monica L. Fernández-Quintero
AU - Nitesh Mishra
AU - Sean Callaghan
AU - Olivia M. Swanson
AU - Daniel L. Bader
AU - James A. Ferguson
AU - Sai Sundar Rajan Raghavan
AU - Benjamin Némoz
PY - 2026
DA - 2026/08/19
JO - Nature Biotechnology
DO - 10.1038/s41587-026-03238-6
UR - https://doi.org/10.1038/s41587-026-03238-6
AB - Experimentally validated prospective, blinded benchmarks are needed to separate durable advances from hype in computational antibody design. Here AIntibody, a challenge inspired by the Critical Assessment of Structure Prediction, tests 511 artificial intelligence (AI)-designed or predicted antibodies from 29 organizations on three tasks: in silico affinity maturation from phase 1 sequencing outputs, affinity ranking within heavy-chain complementarity-determining region 3 (HCDR3) clusters of a selection output and CDR design of proteins not included in a selection output. Validated with diverse experimental assays, several groups produced developable antibodies with affinities <100 pM. However, these successes were exceptions that did not transfer across tasks. Affinity-matured antibodies were modeled effectively. Except for one model, predicting high-affinity clones from clustered HCDR3 datasets was worse than random clone picking. Out-of-library design was highly variable for most method submissions, with many failing to outperform standard selections. The AIntibody challenge shows that AI can optimize antibodies in defined, biologically grounded regimes, in addition to highlighting critical gaps including affinity prediction and library-inspired antibody design and cross-task generalization.
SN - 0733-222X
SN - 1087-0156
SN - 1546-1696
SN - 2331-3684
ER -