To adapt their behaviour in changing environments, cells sense concentrations by binding external ligands to their receptors. However, incorrect ligands may bind nonspecifically to receptors, and when their concentration is large, this binding activity may interfere with the sensing of the ligand of interest. Here, I derive analytically the physical limit to the accuracy of concentration sensing amid a large number of interfering ligands. A scaling transition is found when the mean bound time of correct ligands is twice that of incorrect ligands. I discuss how the physical bound can be approached by a cascade of receptor states generalizing kinetic proof-reading schemes.Because of their small sizes, biological systems typically operate with only a few copies of the molecules they sense and communicate with. In their pioneering work, Berg and Purcell derived the fundamental bound that the noise arising from these small numbers sets on the accuracy of concentration sensing [1]. Experimental progress in the characterization of single-cell variability [2] and sensing precision [3] has fueled a renewed interest in small-number noise and its implications for information processing [4][5][6]. General or refined bounds on sensing accuracy have been recently derived for single receptors [7][8][9], and extended to spatial [10][11][12][13][14] or temporal [15] gradient sensing, while the metabolic cost and trade-offs of sensing accuracy have been explored [16][17][18][19][20][21][22][23][24][25]. Much of this past work has assumed perfect specificity between the biological receptors and their cognate ligands. In realistic biological contexts, large numbers of spurious ligands may bind receptors nonspecifically, interfering with the ligand of interest [26]. This is the case in the problem of antigen recognition by T-cell receptors, where cells must react to a small number of specific foreign peptides among a large number of nonspecific selfpeptides [27]. Biochemical network architectures based on kinetic proofreading [28,29] have been shown to provide a solution to the discrimination problem, and have been studied in depth theoretically [26,[30][31][32]. However, no fundamental bound has been derived against which to compare the performance of these solutions, save for Ref.[33] where concepts of statistical decision theory were used to derive the minimal decision time to detect cognate ligands. In this paper I derive the fundamental limit on concentration sensing accuracy and ligand detection error in the presence of a large number of spurious ligands. The maximum likelihood estimate achieving the bound can be implemented biologically by simple networks based on push-pull reactions.Consider a mixture of two ligands, only one of which the biological system wishes to sense. The ligand of interest (hereafter referred to as the correct ligand) is present in concentration c, while the interfering or spurious ligand (called the incorrect ligand) is present in concentration c . The biological unit can sense ligands through N ident...