Rationally designing drugs that last longer in the face of biological evolution is a critical objective of drug discovery. However, this goal is thwarted by the diversity and stochasticity of evolutionary trajectories that drive uncertainty in the clinic. Although biophysical models can qualitatively predict whether a mutation causes resistance, they cannot quantitatively predict the relative abundance of resistance mutations in patient populations. We present stochastic, first-principle models that are parameterized on a large in vitro dataset and that accurately predict the epidemiological abundance of resistance mutations across multiple leukemia clinical trials. The ability to forecast resistance variants requires an understanding of their underlying mutation biases. Beyond leukemia, a meta-analysis across prostate cancer, breast cancer, and gastrointestinal stromal tumors suggests that resistance evolution in the adjuvant setting is influenced by mutational bias. Our analysis establishes a principle for rational drug design: when evolution favors the most probable mutant, so should drug design.

合理设计在生物进化过程中持续时间更长的药物是药物发现的关键目标。但是，这一目标因进化轨迹的多样性和随机性而受到阻碍，这些变化轨迹在临床上带来了不确定性。尽管生物物理模型可以定性地预测突变是否引起抗药性，但它们不能定量地预测患者群体中抗药性突变的相对丰度。我们提出了在大型体外参数化的随机，第一性原理模型数据集，可以准确预测多个白血病临床试验中耐药突变的流行病学数量。预测抗性变异的能力需要了解其潜在的突变偏向。除白血病外，对前列腺癌，乳腺癌和胃肠道间质瘤的荟萃分析表明，佐剂环境中的耐药性进化受到突变偏倚的影响。我们的分析建立了合理药物设计的原则：当进化倾向于最可能的突变体时，药物设计也应如此。