All the dietary ingredients were collected from the local market. Dried chela fish (Chela cachius) and kachki fish (Corica soborna) were ground into powder and mixed at a 1:1 ratio to make the fishmeal. The proximate composition of these ingredients Four test diets were formulated to contain 40% crude protein and 14% crude lipid. Diets 1, 2, and 3 contained 27% protein from fishmeal, protein concentrate, or meat and bone meal, respectively, while diet 4 contained 27% protein from a mixture of the three protein sources on an equal basis (9% from each). Mustard oil cake was included at a fixed level of 10% in all the diets. Rice bran was included primarily as a carbohydrate source. The diets were formulated to be as iso-proteinous and iso-lipidic as possible by adjusting proportions of soybean meal and soybean oil. All the ingredients were finely ground and sieved through 0.5-mm mesh. All the dry ingredients were weighed according to the formulation and mixed for 30 min using a mixer. Then soybean oil was added and mixed for another 15 min. Finally, water (about 35%–40% of the dry weight of ingredients) was added to the mixture, which was further mixed for 15 min and then made into pellets (1.0–2.2 mm) with a kitchen-type pellet machine (MUM 5580, Type: UM 60 ST2-M, Bosch, Leipzig, Germany). The pellets were sun-dried at ambient temperature (35?C–40?C) for 5 to 6 h and packed in airtight polythene bags and stored at −20?C until fed. Proximate analysis of ingredients, diets, and whole body composition were carried out in triplicate. Crude protein was determined by the Kjeldahl method with a Tecator Kjeltec System (Kjeltec System 1002, Tecator, Sweden). Crude lipid, ash, crude fiber and moisture contents were analyzed using standard methods. Amino acid concentrations of diets were analyzed using high performance liquid chromatography. The feeding trial was conducted for ten weeks in 12 earthen experimental ponds situated at the Faculty of Fisheries, Bangladesh Agricultural University (BAU), Mymensingh, Bangladesh, between August and October 2007. The ponds were square (30 m2) with a water depth maintained at a maximum of 1.2 m using a fine-meshed PVC overflow pipe. Prior to use, the ponds were dried and lime was applied at 250 kg/ha. One week after liming, the ponds were filled with water. The pond water was exchanged at regular intervals from a deep tube-well supply to maintain good water quality. A. testudineus fry (0.53 ± 0.02 g) were obtained from a local fish vendor and randomly allocated to the 12 ponds. The four diets were assigned to four treatments each with three replicates. The fish were stocked at 120/pond (40,000/ha). At the beginning the fish were hand fed daily the respective diets at 10% of their biomass. This was gradually reduced to 5% in the fourth week for the duration of the experiment. Daily ration size was divided into three equal feedings at 08.30, 13.00, and 17.30 h. Sampling was carried out fortnightly using a cast net, and fish were bulk weighed to adjust the ration size. At the beginning of the experiment five fish from the stock pond and, at the end of the trial, five fish from each replicate pond were sampled for analysis of whole body composition. Water temperature, pH, and dissolved oxygen (DO) were measured weekly at 10.00 h. After ten weeks, the fish were harvested by draining the ponds, and all the fish from each pond were washed, counted, bulked weighed, and survival (%) was calculated. Weight gain (%), specific growth rate (SGR, % day), apparent feed intake (AFI, g/fish), apparent feed conversion ratio (AFCR), apparent protein efficiency ratio (APER), and survival (%) of fish were estimated. A simple economic analysis was performed to estimate the gross return generated from different dietary treatments after the ten-week feeding trial. The economic analysis was based on the local market price of the ingredients and fish in Mymensingh in 2007. All data were analyzed using one-way analysis of variance (Super- ANOVA, ver 1.11, Abacus Concepts, Berkeley, Calif.). Percentage survival data were arcsin-square-root transformed before statistical analysis. Levels of significance between individual treatments (P < 0.05) were evaluated by Duncan’s multiple range test (Package super-ANOVA, Abacus Concepts, Berkeley, Calif.).