Co-optimization of therapeutic antibody affinity and specificity using machine learning models that generalize to novel mutational space

Therapeutic antibody development requires selection and engineering of molecules with high affinity and other drug-like biophysical properties. However, optimizing antibody properties such as affinity is often detrimental to other properties such as stability and specificity, which can compromise safety and efficacy. Due to inherent tradeoffs between drug-like biophysical properties, co-optimization of multiple antibody properties remains a difficult and time-consuming process that impedes drug development. Here we have evaluated the use of machine learning to greatly simplify the identification of antibodies with co-optimal levels of affinity and specificity for a clinical-stage antibody (emibetuzumab) that displays both high levels of on-target (affinity) and off-target (non-specific) binding. We mutated sites in the antibody complementarity-determining regions that were predicted to mediate non-specific binding, sorted the antibody libraries for high and low levels of affinity and non-specific binding, and deep sequenced the enriched libraries. Interestingly, we found that machine learning models developed using binary datasets and supervised dimensionality reduction enabled predictions of continuous metrics that were strongly correlated with antibody affinity and non-specific binding. These models illustrated strong tradeoffs between antibody affinity and specificity, as increases in affinity along the co-optimal (Pareto) frontier required progressive reductions in specificity. Notably, models trained with deep learning features enabled extrapolation to predict novel antibody mutations that co-optimized affinity and specificity beyond what was possible for the original antibody library. These findings demonstrate the power of machine learning models to greatly expand the exploration of novel antibody sequence space and accelerate the development of highly potent, drug-like antibodies.

治疗性抗体的研发需要筛选并工程化改造具备高亲和力及其他类药生物物理属性的抗体分子。然而，优化亲和力这类抗体属性时，往往会对稳定性、特异性等其他属性造成损害，进而危及药物的安全性与有效性。由于类药生物物理属性之间存在固有权衡，多抗体属性的协同优化仍是一项困难且耗时的工作，阻碍了药物研发进程。本研究评估了机器学习的应用潜力，旨在大幅简化针对临床阶段抗体艾米妥珠单抗（emibetuzumab）的、具备协同最优亲和力与特异性水平的抗体的筛选流程；该抗体同时呈现出高水平的靶标结合（即亲和力）与非靶标结合（即非特异性结合）活性。我们对抗体互补决定区（complementarity-determining regions）中被预测介导非特异性结合的位点引入突变，基于亲和力与非特异性结合水平的高低对抗体文库进行分选，并对富集后的文库进行深度测序。有趣的是，本研究发现，使用二元数据集与监督降维方法构建的机器学习模型，能够预测与抗体亲和力及非特异性结合高度相关的连续量化指标。这些模型揭示了抗体亲和力与特异性之间存在显著权衡：若要沿着协同最优帕累托（Pareto）前沿提升亲和力，则需逐步降低特异性。值得注意的是，基于深度学习特征训练的模型能够实现外推预测，筛选出可在亲和力与特异性上实现协同优化的新型抗体突变体，其性能超越了原始抗体文库所能达到的水平。本研究结果证实，机器学习模型可极大拓展对新型抗体序列空间的探索，并加速高效、类药治疗性抗体的研发进程。