A series of 2,4-dinitrobenzamide mustards were prepared from 5-chloro-2,4-dinitrobenzoic acid or the corresponding 5-dimesylate mustard as potential prodrugs for gene-directed enzyme prodrug therapy (GDEPT) with the E. coli nfsB nitroreductase (NTR). The compounds, including 32 new examples, were evaluated in four pairs of NTR+ve/-ve cell lines for selective cytotoxicity (IC50 and IC50 ratios), in multicellular layer (MCL) cultures for bystander effects, and for in vivo activity against tumors grown from stably NTR transfected EMT6 and WiDr cells in nude mice. Multivariate regression analysis of the IC50 results was undertaken using a partial least-squares projection to latent structures model. In NTR-ve lines, cytotoxicity correlated positively with logP, negatively with hydrogen bond acceptors (HA) and donors (HD) in the amide side chain, and positively with the reactivity of the less-reactive leaving group of the mustard function, likely reflecting toxicity due to DNA monoadducts. Potency and selectivity for NTR+ve lines was increased by logP and HD, decreased by HA, and was positively correlated with the leaving group efficiency of the more-reactive group, likely reflecting DNA crosslinking. NTR selectivity was greatest for asymmetric chloro/mesylate and bromo/mesylate mustards. Bystander effects in the MCL assay also correlated positively with logP and negatively with leaving group reactivity, presumably reflecting the transcellular diffusion/reaction properties of the activated metabolites. A total of 18 of 22 mustards showed equal or greater bystander efficiencies in MCLs than the aziridinylbenzamide CB 1954, which is currently in clinical trial for NTR-GDEPT. The dibromo and bromomesylate mustards were surprisingly well tolerated in mice. High MTD/IC50 (NTR+ve) ratios translated into curative activity of several compounds against NTR+ve tumors. A bromomesylate mustard showed superior activity against WiDr tumors grown from 1:9 mixtures of NTR+ve and NTR-ve cells, indicating a strong bystander effect in vivo.