SynFlowNet: Towards Molecule Design with Guaranteed Synthesis Pathways

最近生成模型的突破使得许多人提出了用于药物发现的分子生成模型。虽然这些模型在捕捉类似药物的基元方面表现良好，但它们通常会产生合成上无法实现的分子。这是因为它们被训练以一种逼近训练分布的方式来组合原子或碎片，但它们并不明确地意识到制造实验室中的分子的合成限制。为了解决这个问题，我们介绍了SynFlowNet，这是一个使用经过化学验证的反应和反应物来顺序构建新分子的GFlowNet模型。我们使用合成可达性分数和独立的反合成工具来评估我们的方法。SynFlowNet始终会采样出合成可行的分子，同时仍能找到多样性和高效的候选物。此外，我们将使用SynFlowNet设计的分子与实验验证的活性物质进行比较，并发现它们具有类似的重量、SA分数和预测的蛋白质结合亲和力等感兴趣的性质。

SynFlowNet: A GFlowNet Model for Synthetically Feasible Molecular Generation

The paper proposes SynFlowNet, a GFlowNet model that uses chemically validated reactions and reactants to sequentially build new molecules, addressing the issue of generative models producing synthetically inaccessible molecules.

SynFlowNet consistently samples synthetically feasible molecules while still being able to find diverse and high-utility candidates. The approach is evaluated using synthetic accessibility scores and an independent retrosynthesis tool. The designed molecules show comparable properties of interest to experimentally validated actives. The paper also discusses the limitations and future directions of the proposed approach.

Recent breakthroughs in generative modelling for drug discovery include 'MoleculeVAE: A Generative Model for Molecular Graphs' and 'Graph Convolutional Policy Network for Goal-Directed Molecular Graph Generation'.