Gracilaria tenuistipitata extract was prepared following the protocol described by Thépot et al. (25). The red-seaweed sample was collected in May 2022 from the Sonadia Sea Beach coast (21°29′N 91°54′E), Moheskhali, Bangladesh [water salinity ~32 ppt, temperature ~26°C] and thoroughly washed with tap water to remove unwanted contaminants (i.e., salt, sediments, sands, invertebrates, and epiphytes). The clean seaweed samples were then kept in a freeze drier at −80°C for 72 h. Once dried, the sample was grounded and sieved (200 µm mesh size) to produce a fine powder that was then vacuum-sealed and stored at −20°C for further processing. The Gracilaria extract (GE) was made with 70% methanol in a 1:10 (m:v) ratio under dark conditions about 12 h in each time. The crude extract was filtered (Whatman^® no. 2) and then evaporated slowly in a rotary evaporator (IKA^® RV3 Eco). Once the methanol had completely evaporated, the extract was then stored at −20°C for use. The various steps involved in making the GE are depicted in Figure 1 . A representative amount of GE (triplicated) was dried using a freeze dryer at −80°C. The dried GE was then assessed for its nutrient composition (%) including protein (12.47 ± 1.57), lipid (1.35 ± 0.14), and ash (27.02 ± 1.50) based on a standard protocol (AOAC, 2023). The bioactive compounds, including total phenolic content (TPC) and total flavonoid content (TFC), were determined using the procedure described by Sobuj et al. (26). The results showed a TPC of 63.46 ± 1.77 mg GAE/g (gallic acid equivalents per gram) and a TFC of 29.71 ± 1.09 mg QE/g (quercetin equivalent per gram) using methanol as the solvent. The GE was added to the feed mixture along with all the other ingredients during the water addition step to make dough. Afterward, the dough was pelletized measuring 1.0–3.0 mm using a laboratory pelletizer. The pellets were dried at 50°C for 12 h and stored in airtight polythene bags in a refrigerator at 4°C until used. The feed formulation and chemical composition of the diet are presented in Table 1 .

Monosex Nile tilapia, O. niloticus, fry were purchased from a local fish hatchery (BRAC Fisheries, Khulna, Bangladesh). Nile tilapia fry were placed in eight experimental tanks, each with a 100-L water holding capacity, continuous aeration, and a water temperature of 28°C. They were then acclimatized in tanks for a period of 2 weeks where fish were fed the commercial tilapia diet (Quality Feeds Limited, Bangladesh) twice daily. After 2 weeks of acclimatization, fish were fasted for 24 h, then 200 similar-sized (1.61 ± 0.15 g) fry were randomly assigned into the same eight tanks divided into two dietary groups (four tanks per diet with 25 fish per tank). One quadruplicate group served as a control termed CON, while another group was fed the CON diet added with 1% Gracilaria extract (dry weight basis of feed) named GE. The concentration of GE (~1% of feed) was determined based on previous findings (27). The experiment duration was 10 weeks in which fish were kept under ambient temperature conditions (28°C) for the first 6 weeks. In the following 2 weeks, the water temperature was gradually increased by 1°C every 2 days from the initial temperature to 35°C to test the fish ability to withstand temperature changes. Subsequently, it was maintained at 32°C for 2 weeks to acclimatize the fish to a temperature 4°C higher than their original ambient temperature of 28°C. An overview of the experimental time chart is illustrated in Figure 2 . Aerators were used in every tank to ensure enough oxygen (>5 mg/L). Fish were fed twice daily, at 08:00 and 16:00 h, until they appeared to be satisfied. The amount of feed delivered and the bulk weight of each replication tank were assessed every 2 weeks to monitor fish performance over time. To keep the water quality favorable for fish, regular siphoning was ensured to take out the unused feed and feces. Every morning, one-third of the tank water was replaced with temperature-controlled reservoir tank water to maintain the desired water temperature for fish.

Following the feeding trial, fish were counted and bulk weighed in each tank. Several performance parameters such as final body weight (FBW), specific growth rate (SGR), feed intake (FI), feed conversion ratio (FCR), hepatosomatic index (HSI), viscerasomatic index (VSI), and survival rate (SR) were calculated using the formula mentioned in an earlier study (28).

Eight randomly selected fish (two fish per each replicate) from each dietary group were sampled for liver and intraperitoneal fat histology. The collected samples were cleaned with normal saline solution to remove blood and other unwanted substances and were preserved immediately in 10% buffered formalin and stored at 4°C until further processing. The muscle and fat samples were preserved in 70% alcohol at 4°C until histological analysis. Assessment of lipid accumulation in the liver was performed using Oil Red O staining, whereas muscle and intraperitoneal fat samples were stained with hematoxylin and eosin (H&E). The histology was performed based on standard protocol. Briefly, fixed samples were dehydrated by passing through a graded series of alcohol. After dehydration, the samples were molted with wax and kept in cool storage. The samples were then sectioned (5 μm) using a microtome. The sectioned tissues were stained with H&E and observed under a light microscope (Olympus, Germany) at 400× magnification with photos captured with an onboard camera (BX40F4, Olympus, Tokyo, Japan) connected to the microscope. Muscle microstructure such as fiber diameter and density were assessed by nonparametric statistical procedures according to Rowlerson et al. (29). The adipocyte number and diameter were determined from 30 intact cells from each dietary group according to Chaklader et al. (30). All images were captured using a light digital microscope connected to a camera and ImageJ software (version: 1.53). The protocol used to calculate lipid droplet accumulation in the liver was described in Supplementary Figure S1 ).

At the end of the feeding trial, 4 fish from each replicate tank (16 fish/dietary group) were chosen at random, and blood samples were taken by puncturing the caudal vein and kept in non-heparinized tubes, left to settle for 6 h, and then centrifuged at 4°C for 10 min to extract serum that was kept at −20°C. The lysozyme activity of serum was analyzed following Siddik et al. (31). The levels of aspartate aminotransferase (AST), alanine transaminase (ALT), cholesterol, triglycerides, and glycogen were analyzed by an automated blood analyzer (SLIM; SEAC Inc., Florence, Italy) following Blanc et al. (32).

Fish that were being considered for blood collection were dissected to obtain liver tissue for measuring the antioxidant response. The activities of antioxidant enzymes including malondialdehyde (MDA), superoxide dismutase (SOD), and catalase (CAT) were assessed using commercial kits from ZellBio GmbH, Lonsee, Germany.

Total RNA from frozen hind gut samples of GE and control-fed fish was isolated using RNeasy Mini Plus Kit (Qiagen, Hilden, Germany) following the manufacturer’s instructions. The quality and quantity of the RNA were evaluated through agarose gel electrophoresis and spectrophotometry (NanoDrop^® ND-2000), respectively. cDNA synthesis was performed using a PrimeScript™ RT reagent kit (Takara, Japan). The real-time quantitative PCR (RT-qPCR) primers used in this study for Nile tilapia are listed in Table 2 . The melting curve of the amplicon from the primers was obtained using uMELT software. Gene expression levels were measured by RT-qPCR with PowerUp™ SYBR Green Master Mix (Thermo Scientific, USA) on the 7500 Real-Time PCR System (Applied Biosystems, USA). The following conditions were used for real-time PCR: initial denaturation for 2 min at 95°C, followed by 40 cycles of amplification stating 30 s denaturation at 95°C, annealing for 1 min at 60°C, and extension at 72°C for 30 s. The melting stage in PCR began with the 95°C heating step for 15 s and cycled to 70°C cooling for 1 minute, with the continuous increase of 0.015°C per second. The analysis of RT-qPCR data for relative gene expression levels was normalized to the β-actin content in each sample and quantified using the 2^−ΔΔCt method as outlined by Livak and Schmittgen (33).

Blood red blood cell (RBC) level and its cellular and nuclear abnormalities in fish were measured in Nile tilapia at three different temperatures across the feeding trial in D42 (pre-temperature stress at 28°C), D56 (post-temperature stress at 35°C), and D70 (post-temperature adaptation at 32°C). Two fish from each replicate tank were selected for blood sampling and euthanized after recording their weights for biomass calculations. RBC counts were determined using a hemocytometer. The blood smear slides were prepared immediately after collecting the blood. The blood was spread along the edge of the slide and tilted at a 45° angle to create a thin, even smear. After air-drying for 5 min, the smear was fixed in methanol for approximately 2 min. Once dry and fixed, it was stained with a 5% Giemsa solution. The slides were then air-dried overnight and mounted with dibutylphthalate polystyrene xylene (DPX). The samples were examined under an optical microscope (G-206, Italy) using a 100× objective lens to assess RBC cellular and nuclear abnormalities. RBC cellular abnormalities were categorized as elongated (significantly longer than wider), fusion (joining of more than two cells to create a larger mass), and twin (two cells connected by their surfaces). RBC nuclear abnormalities consisted of micronucleus (circular chromatin bodies resembling the central nucleus), nuclear bridge (a strip of nuclear material connecting two nuclei within separate erythrocytes or within a single erythrocyte), and nuclear degeneration (nuclear condensation and extrusion leading to the creation of a pyrenocyte structure, which is subsequently engulfed and broken down by macrophages).

Eight fish from each dietary treatment were randomly selected to obtain hind gut samples, following appropriate biosafety measures. The gut samples, which included mucosa and digesta, were homogenized using a TissueLyser II (Qiagen, Hilden, Germany). Genomic DNA extraction was then carried out using the DNeasy PowerSoil Pro Kit (Qiagen, Hilden, Germany) following the instructions provided by the manufacturer. The DNA was quantified using a NanoDrop 2000c (Thermo Fisher Scientific, USA), and its quality was assessed by running it on a 1% agarose gel. To prepare the final master mix, 50 μL of Hot Start 2× Master mix (New England BioLabs Inc., USA), as per the manufacturer’s instructions, was combined with 2 μL of template DNA, 1 μL each of V3V4 primers (10 µM) with Illumina overhang adapter, and 21 μL of nuclease-free treated water. The mixture was then subjected to 35 cycles of amplification using a BioRad S1000 Thermal Cycler (BioRad Laboratories Inc., USA). Positive amplicons were purified using AMPure beads and indexed according to the Illumina 16S Sequencing Library Preparation protocol (Part # 15044223 Rev. B). Finally, the equimolar amplicons were pooled and sequenced on an Illumina MiSeq platform (Illumina Inc., San Diego, California, USA) using the paired-end, v3 kit with 600 cycles.

Paired-end amplicon sequence data (gz format) were imported in qiime2 (v2021.11) for further processing. The denoising of reads was performed using DADA2 followed by trimming of demultiplexed reads with parameters such as -p-trim-left-f l0; -p-trunc-len-f 260; -p-trim-left-r 10; -p-trunc-len-r 220. Chimeric sequences with >0.05% error rates were removed and non-chimeric reads (82%) indicating biological features were tabulated for feature frequency amplicon sequence variants (ASVs). The lowest non-zero frequency of 10 was set to filter the feature ASV table. Phylogenetic classification of ASVs into different taxa levels was performed using the qiime2 “classify consensus- blast” plug-in against SILVA 138 release (34). The feature table collapsed with taxonomy and subsequently removed mitochondrial and chloroplast sequences (<1% of reads). There were variations in Clostridium classification, and as a result, these variations were classified as “Clostridium sensu stricto 1-9” throughout the dataset. To make a homogeneous dataset for Clostridium, we renamed all “Clostridium sensu stricto 1-9” as “Clostridium”, and other non-classified reads into the “Unclassified” bacterial group. The final set of data was normalized (also called “rarefaction”) at an even depth of 15,236 for bacterial community analysis. Metagenome prediction of functional features from 16S rRNA data was performed with the PICRUSt2 pipeline (35).

The ASV table, taxonomy, and metadata files were imported into R software (v4.22) for diversity and composition analysis (36). The number of shared and unique taxa were calculated using the MicEco package (37). Species richness (observed), Chao1, Shannon, and Simpson diversity were considered for alpha-diversity measurements. Beta-diversity was performed as a UniFrac distance metric (Unweighted and Weighted) and PERMANOVA was conducted for the visualization of feeding effects on beta-dispersion with 1,000 permutations using vegan and phyloseq (38) R packages. The distance between samples for a group was calculated as the “Bray–Curtis “ distance. We considered 1% taxa abundance per sample as a threshold for composition analysis at the phylum and genus level (39). Further analysis of differential abundance between groups was performed as Linear Discriminant Analysis (LDA) using the MicrobiomeMarker R package (40). An LDA cutoff value of 2.0 and a p-value of <0.05 were considered as statistically significant for compositional difference analysis.

Data were assessed for normality using the Shapiro–Wilk test and then analyzed with two-tailed Student’s t-tests and two-way analysis of variance (ANOVA). Two-tailed Student’s t-tests were applied to compare the effects of GE diet to control. The total RBC count, along with its cellular and nuclear abnormalities, was analyzed using two-way ANOVA. Additionally, alpha-diversity, beta-diversity, and compositional log-fold changes between the two dietary groups in microbiome composition were analyzed using two-tailed t-tests. Gut microbiome analysis was carried out using the R software. A p-value of <0.05, <0.01, and <0.001were considered statistically significant between the two dietary groups of GE and CON.

Nile tilapia fed the GE diet showed a significant improvement in growth performance and feed utilization compared to the CON ( Figures 3A–E , p < 0.05). However, the somatic indices such as hepatosomatic index and viscerasomatic index were not influenced by the GE supplementation ( Figures 3F, G , p > 0.05). Additionally, there was no significant difference in the survival rate between the two dietary groups ( Figure 3H ). The higher growth performance was supported by the enhanced dorsal muscle fiber diameter in fish fed the GE diet ( Figures 3I–K ).

Hepatic lipid droplets significantly reduced in the GE-fed fish compared to the control, while liver weight exhibited no significant difference between the dietary groups ( Figures 4A–C ). Likewise, the hepatic enzymes AST and ALT were lower in GE-fed fish compared to control-fed fish, whereas cholesterol, triglyceride, and glycogen levels were not affected by seaweed supplementation ( Figures 4D–H ). Adipocyte numbers and fatty tissue diameter in fish fed the control diet were significantly decreased in comparison to fish fed the control diet ( Figures 4I–K ). The frequency of adipocyte diameter increased in GE-fed fish up to 50 μm, while adipocytes in fish fed the CON diet were greater in diameter (>50 to 200 μm) ( Figure 4L ).

The level of MDA significantly reduced in fish fed the GE-supplemented diet while CAT and SOD levels exhibited no difference between the two dietary groups ( Figures 5A–C , p < 0.05). The lysozyme activity was enhanced in the GE group compared to the CON group ( Figure 5D ). Dietary GE supplementation significantly upregulated the expression of tight junction proteins (occludin, claudin1, and ZO-1) and the antioxidant gene, nrf2, while a downregulation of hsp70 was observed in the GE group compared to the CON group. The pro-inflammatory cytokines tnf-α and il-1β were found unaffected by seaweed supplementation ( Figure 5E , p > 0.05).

RBC count and cellular and nuclear abnormalities were found significantly different by both the dietary groups and temperatures ( Figure 6 , p < 0.05). Fish fed the GE-supplemented diet significantly enhanced the RBC level when compared to the control. Likewise, the total frequencies of RBC cellular and nuclear abnormalities were found significantly lowered in the control group when compared to the control. However, fish at 35°C post-temperature stress produced more cellular and nuclear abnormalities compared to pre-temperature stress and post- temperature adaptation at 28°C and 32°C, respectively.

A total of 322, 761-bp quality reads were obtained after trimming, with an average of 20,172.6 ± 941.7 and ranging from 16,860 to 30,244 bp that were classified into 1,503 ASVs, six phyla, and 455 genera. The average good’s coverage index value of 0.998 and plateaued rarefaction curve were observed, indicating adequate depth and saturation level of all study sequences ( Figure 7A , Supplementary Figure S2 ). Only 63 ASVs were shared between CON and GE diet groups while the latter generated 548 additional unique ASVs compared to 218 in CON ( Figure 7B ). Consistent with ASVs, species diversity (observed ASVs and Chao index) and Shannon and Simpson diversities reflecting richness and evenness of top abundant taxa were significantly higher in the GE group, compared to CON ( Figure 7C ). Distinct bacterial communities in terms of bacterial presence–absence and relative abundance were observed between CON and GE diet groups in the beta ordination PCoA plot. Significant R and p-values in PERMANOVA indicate the influence of dietary GE in the gut bacterial diversity in terms of shifting in relative abundance of top abundant taxa and augmenting rare bacterial communities compared to the CON group ( Figures 7D, E ). The Bray–Curtis distance showed more balance variation in abundance for bacterial communities in the GE diet compared to CON feed ( Figure 7F ).

At the phylum level, Proteobacteria represented approximately half of the classified reads (45%) in both groups —37.8% in CON and 53.8% in GE, while Fusobacteriota (28%) and Actinobacteria (28.7%) composed 56.7% of reads in the CON diet. Firmicutes (13.8%) had a higher abundance than Fusobacteria (14.8%) and Actinobacteria (12.7%) in the fish gut-fed GE diet. Firmicutes abundance was only 2.6% in the CON diet, indicating the influential impact of GE diet on this phylum. Alongside Firmicutes and Bacteroidota, abundance also increased from <0.5% in CON to 3.8% in the GE diet ( Figure 8A ). At the genus level, Aeromicrobium, Escherichia-Shigella, and Cetobacterium comprised 84.6% of the total reads in fish fed the CON diet while Pseudomonas (32.3%) was the most abundant bacterial group in fish fed the GE diet, followed by Escherichia-Shigella (14.5%), Cetobacterium (13.3%), Aeromicrobium (4.9%), Cutibacterium (3.6%), and Sphingobacterium (2.2%) ( Figure 8B ). Dietary GE provision significantly promoted the abundance of Bacteroidota, Campilobactorota, Desulfobacterota, and Firmicutes at the phylum level compared to Nitrospirota in the CON diet ( Figure 9A ). At the order and genus level, Sphingobacterium, Prevotella, Lactobacilliles, Lactobacillus, Roseburia, Bifidobacterium, and Ruminococcus showed significantly higher abundance in the GE diet compared to Rhodopirellula and Rhizobiales in CON ( Figure 9B , Supplementary Figure S3 ). The GE diet also increased abundance for some opportunistic pathogens including Streptococcus, Staphylococcus, and Corynebacterium ( Figure 9C ). A shift in bacterial abundance in the gut also modulates metabolic pathways in the dietary GE group, specifically amino acid biosynthesis and metabolism compared to glucose–sucrose metabolism in the CON group ( Supplementary Figure S4 ). The microbial interaction based on ASV correlation showed the dominance of Proteobacteria and Firmicutes in the network. Some of the Proteobacteria and Firmicutes communities were found to be self-interactive ( Figure 10A ). Despite their high abundance, Actinobacteria and Fusobacteria were less interactive. Microbial co-occurrence network analysis revealed that Firmicutes led interactions in the community, despite the Proteobacteria-rich environment and their dominance in strong interactions. Firmicutes were associated with over 70% of medium and 50% of weak interactions, mostly with Bacteroidetes and Actinobacteria. Proteobacteria led the strong interactions and showed no involvement with Firmicutes ( Figure 10B ).