The bioactivity of compounds plays an important role in drug development and discovery. Existing machine learning approaches have poor generalizability in bioactivity prediction due to the small number of compounds in each assay and incompatible measurements among assays. In this paper, we propose ActFound, a bioactivity foundation model trained on 1.6 million experimentally measured bioactivities and 35,644 assays from ChEMBL. The key idea of ActFound is to use pairwise learning to learn the relative bioactivity differences between two compounds within the same assay to circumvent the incompatibility among assays. ActFound further exploits meta-learning to jointly optimize the model from all assays. On six real-world bioactivity datasets, ActFound demonstrates accurate in-domain prediction and strong generalization across assay types and molecular scaffolds. We also demonstrate that ActFound can be used as an accurate alternative to the leading physics-based computational tool FEP+(OPLS4) by achieving comparable performance when using only a few data points for fine-tuning. Our promising results indicate that ActFound could be an effective bioactivity foundation model for compound bioactivity prediction, paving the way for machine-learning-based drug development and discovery.

化合物的生物活性在药物开发和发现中起着重要作用。由于每次测定中的化合物数量较少以及测定之间的测量不兼容，现有的机器学习方法在生物活性预测方面的通用性较差。在本文中，我们提出了 ActFound，这是一种生物活性基础模型，经过 160 万个实验测量的生物活性和来自 ChEMBL 的 35,644 次检测进行训练。 ActFound 的关键思想是使用配对学习来了解同一测定中两种化合物之间的相对生物活性差异，以避免测定之间的不兼容性。 ActFound 进一步利用元学习来联合优化所有分析的模型。在六个真实世界的生物活性数据集上，ActFound 展示了准确的域内预测以及跨检测类型和分子支架的强大泛化能力。我们还证明，ActFound 可以作为领先的基于物理的计算工具 FEP+(OPLS4) 的准确替代品，只需使用几个数据点进行微调即可实现相当的性能。我们令人鼓舞的结果表明，ActFound 可能成为化合物生物活性预测的有效生物活性基础模型，为基于机器学习的药物开发和发现铺平道路。