Transcriptional targeting of cytotoxic genes is an important way to control toxicity associated with gene transfer therapies, but supposedly, tissue-specific promoters are often either very weak and/or leaky. In addition, the phenotypic leakiness of such tissue-specific promoters is dependent upon the toxicity of the gene being used. Therefore, we devised a transcriptional feedback loop to restrict gene expression of very potent genes to melanoma cells. We screened different elements of the human tyrosinase promoter to find one which gave no detectable expression in non-melanoma cells but was active in melanoma cell lines. This weak, but highly tissue specific, element (Tyr-300) was then used as the basis for a transcriptional amplification feedback loop in which a consensus heat shock element (HSE) was cloned upstream of Tyr-300. The cytotoxic gene was cloned downstream of the HSE-Tyr-300 element along with a mutated form of the heat shock factor-1 (HSF-1) transcription factor, which no longer requires cellular stress to activate its trimerisation, nuclear localisation and transcriptional activation properties. Low levels of expression from Tyr-300 initiated expression of both the cytotoxic and the HSF-1 genes in melanoma cells. Gradual build up of HSF-1 amplified expression through binding to the HSE to give levels of cytotoxicity similar to that provided by a CMV promoter. However, no leakiness was observed in multiple non-melanoma cell lines tested. In addition to amplifying low levels of weak tissue-specific expression, the use of HSF-1 also leads to activation of endogenous stress-related genes such as hsp70. Induction of these genes, in the presence of cell killing by the cytotoxic gene, is a highly immunostimulatory event which enhances the antitumour vaccination effects of direct tumour cell destruction. Having demonstrated the compatibility of the component elements in plasmid form, we incorporated the feedback loop into a hybrid LTR-modified retroviral vector and confirmed that the system can be effective in the form of a viral vector. The format of the feedback loop described here could be exploited for any tissue type in which a highly tissue-specific element can be identified but which is itself too weak to be effective therapeutically.