1. Non-image forming, irradiance-dependent responses mediated by the human eye include synchronisation of the circadian axis and suppression of pineal melatonin production. The retinal photopigment(s) transducing these light responses in humans have not been characterised. 2. Using the ability of light to suppress nocturnal melatonin production, we aimed to investigate its spectral sensitivity and produce an action spectrum. Melatonin suppression was quantified in 22 volunteers in 215 light exposure trials using monochromatic light (30 min pulse administered at circadian time (CT) 16-18) of different wavelengths (lambda(max) 424, 456, 472, 496, 520 and 548 nm) and irradiances (0.7-65.0 microW cm(-2)). 3. At each wavelength, suppression of plasma melatonin increased with increasing irradiance. Irradiance-response curves (IRCs) were fitted and the generated half-maximal responses (IR(50)) were corrected for lens filtering and used to construct an action spectrum. 4. The resulting action spectrum showed unique short-wavelength sensitivity very different from the classical scotopic and photopic visual systems. The lack of fit (r(2) < 0.1) of our action spectrum with the published rod and cone absorption spectra precluded these photoreceptors from having a major role. Cryptochromes 1 and 2 also had a poor fit to the data. Fitting a series of Dartnall nomograms generated for rhodopsin-based photopigments over the lambda(max) range 420-480 nm showed that rhodopsin templates between lambda(max) 457 and 462 nm fitted the data well (r(2) > or =0.73). Of these, the best fit was to the rhodopsin template with lambda(max) 459 nm (r(2) = 0.74). 5. Our data strongly support a primary role for a novel short-wavelength photopigment in light-induced melatonin suppression and provide the first direct evidence of a non-rod, non-cone photoreceptive system in humans.