Electrocatalysts for the nitrogen reduction reaction (NRR) have attracted ever-growing attention due to their applications in renewable energy alternative processes to fossil fuels. However, the activation of the inert N–N bond requires multiple complex charge injections, which complicates the design of the catalysts. Here, by combining atomic-scale screening and machine learning (ML) methods, we explore the rational design of NRR single-atom catalysts (SACs) supported by molybdenum disulfide (MoS₂). Our work reveals that the activity of NRR SACs is highly dependent on the number of unpaired d electrons of TMs, with “positive” samples with high activity favoring higher values while “negative” cases distribute at lower values, both varying with the doping conditions of the host. We find that the substitution of sulfur with boron can activate the intrinsic NRR activity of certain TMs such as Ti and V, which are otherwise inactive over pristine MoS₂. Importantly, among the various descriptors used in ML, the charged state of adsorbed TMs plays a key role in donating an electron to the π* anti-bonding orbital of N₂via the back-donation mechanism. Our work shows a feasible strategy for the rational design of NRR SACs and retrieval of the decisive feature of active catalysts.