计算机科学
人工智能
自然语言处理
抗体
蛋白质-蛋白质相互作用
模式识别(心理学)
计算生物学
语言模型
混合模型
班级(哲学)
机器学习
作者
Samad Amini,Yimin Huang,Mark C. Julian,Christina Palmer,Simone Sciabola,Y. C. Wang
出处
期刊:mAbs
[Landes Bioscience]
日期:2026-03-20
卷期号:18 (1): 2647489-2647489
被引量:7
标识
DOI:10.1080/19420862.2026.2647489
摘要
Protein language models (PLMs) provide a powerful framework for learning sequence - property relationships in antibodies. However, their performance and reliability in real-world industrial antibody discovery pipelines remain underexplored. Here, we systematically evaluate several state-of-the-art PLMs using internal datasets comprising antibody sequences and developability assay measurements from 33 historical therapeutic programs. The assays span three critical developability dimensions: polyspecificity reagent (PSR), hydrophobic interaction chromatography (HIC), and affinity-capture self-interaction nanoparticle spectroscopy (AC-SINS). Across all assays, domain-adaptive fine-tuning of PLMs on internal antibody sequence data consistently improves predictive performance relative to pretrained representations alone. In addition, we assess sequence likelihoods derived from pretrained PLMs as unsupervised indicators of developability risk and analyze their strengths and limitations across assay types. Together, these results demonstrate that PLMs can provide robust and complementary signals for antibody developability assessment, supporting their practical use in early-stage candidate optimization and selection.
科研通智能强力驱动
Strongly Powered by AbleSci AI