Search results and study characteristics

We identified 256 potentially relevant articles, and of these, twelve were excluded (five were the wrong publication type, one was published in a potentially predatory journal, five were not published in English language and one was not available). Of the 244 articles assessed for eligibility, 153 were excluded based on the title or abstract, and fifty-two were excluded due to ineligibility issues after full-text screening. In total, thirty-nine articles were eligible for inclusion in this systematic review and meta-analysis (Fig. 1), comprising a total of 512 and 423 rodents in the intervention and comparator groups, respectively. Of the reviewed articles, twenty-eight used rats and eleven used mice. Eligible articles using rodents other than rats and mice were not identified. The study characteristics are presented in Table 1 and Supplementary Table 1. The included studies were published between the years 1986 and 2021. The vast majority of the studies used male rodents; only two of the thirty-nine reviewed articles reported the use of female rodents, both were mice studies^{(Reference Bjorndal, Burri and Wergedahl34,Reference Parolini, Vik and Busnelli35)} , and one article did not state the sex of the mice^{(Reference Moriya, Hosokawa and Miyashita36)}. None of the articles reported using both male and female rodents. All reviewed articles stated the age and/or bodyweight at the initiation of the intervention period. The age of the rodents at the start of the intervention ranged from 4 weeks to 15 weeks, i.e. periadolescent and young adults. The durations of the intervention periods were between 1 and 12 weeks, with 4 weeks being the most common duration of intervention. The included articles reported the number of animals in each experimental group, with one exception^{(Reference Chevrier, Mitchell and Rioux37)}, and the group sizes ranged from four to fourteen rodents.

Table 1. Study characteristics and outcomes

CYP7A1, cholesterol 7 α-hydroxylase; N/A, data not available; NS, not statistically significant; SOAT-2, sterol O-acyltransferase 2; TBA, total bile acids; TC, total cholesterol.

The blood samples for analyses of TC, HDL-cholesterol and LDL-cholesterol were collected at the time of euthanasia for all articles except one, where blood samples were collected 1 week prior to euthanasia^{(Reference Parolini, Vik and Busnelli35)}. Blood was sampled from fasting rats in the majority of the articles^{(Reference Vikoren, Drotningsvik and Bergseth31,Reference Vikoren, Drotningsvik and Mjos32,Reference Bjorndal, Burri and Wergedahl34–Reference Mijiti, Mori and Huang53)} , while twelve articles declared that animals were not fasted^{(Reference Hosomi, Maeda and Ikeda54–Reference Hosomi, Fukunaga and Arai65)}. Two articles presented results where half of the rats in each of the experimental groups were non-fasting while the rest of the animals were in a fasting state^{(Reference Hurley, Galibois and Jacques66,Reference Shukla, Bettzieche and Hirche67)} , and four articles did not include information about the prandial state of the animals^{(Reference Parolini, Vik and Busnelli35,Reference Wergedahl, Liaset and Gudbrandsen68–Reference Hosomi, Miyauchi and Yamamoto70)} . The details for prandial conditions at euthanasia are presented in online Supplementary Table 1.

Most of the included studies were designed with ad libitum access to feed^{(Reference Vikoren, Drotningsvik and Bergseth31,Reference Vikoren, Drotningsvik and Mjos32,Reference Bjorndal, Burri and Wergedahl34,Reference Moriya, Hosokawa and Miyashita36–Reference Affane, Louala and El Imane Harrat46,Reference Hosomi, Nishimoto and Kobayashi48,Reference Vik, Tillander and Skorve50,Reference Mijiti, Mori and Huang53,Reference Hosomi, Maeda and Ikeda54,Reference Maeda, Hosomi and Fukuda56,Reference Hosomi, Fukunaga and Arai59,Reference Kawabata, Mizushige and Uozumi62,Reference Jacques, Deshaies and Savoie63,Reference Hosomi, Fukunaga and Arai65,Reference Hurley, Galibois and Jacques66,Reference Wergedahl, Liaset and Gudbrandsen68)} . The remaining studies were designed with pair feeding^{(Reference Liaset, Madsen and Hao49,Reference Maeda, Hosomi and Yokoyama64)} , slightly restricted access to feed^{(Reference Shukla, Bettzieche and Hirche67)} or gave no information on whether feed was freely available or controlled^{(Reference Parolini, Vik and Busnelli35,Reference Wang, Gagnon and Nair47,Reference Louala, Hamza-Reguig and Benyahia-Mostefaoui51,Reference Wang, Nair and Gagnon52,Reference Hosomi, Fukunaga and Arai57,Reference Hosomi, Fukunaga and Arai58,Reference Maeda, Hosomi and Koizumi60,Reference Mizushige, Kawabata and Uozumi61,Reference Hosomi, Fukunaga and Arai69,Reference Hosomi, Miyauchi and Yamamoto70)} . One article contained one study with ad libitum access to feed and one study with pair feeding^{(Reference Liaset, Hao and Jorgensen55)}.

Assessments of risk of bias and quality of evidence

The reviewed articles were assessed with the SYRCLE’s risk of bias tool as presented in Supplementary Table 2. The use of sequence generation (#1) for reducing selection bias was not reported in any of the reviewed articles, and all articles were therefore graded with ‘No’. Body weight and age were defined as the baseline characteristics (#2) that were compared between groups, and the majority of the articles used animals that were similar at baseline and were graded with ‘Yes’. Eight articles did include a baseline comparison of animals^{(Reference Jacques, Deshaies and Savoie39,Reference Demonty, Deshaies and Lamarche40,Reference Louala, Hamza-Reguig and Benyahia-Mostefaoui51,Reference Liaset, Hao and Jorgensen55,Reference Jacques, Deshaies and Savoie63,Reference Hurley, Galibois and Jacques66–Reference Wergedahl, Liaset and Gudbrandsen68)} , thus resulting in an unclear risk of bias. One article included experimental groups that differed in bodyweight at baseline, corresponding to 2–3 d of growth, possibly causing an increased risk of bias^{(Reference Drotningsvik, Vikoren and Mjos44)}. There was an unclear risk of bias in all studies for the allocation concealment (#3), thus all articles received ‘Unclear’ on this entry. Although none of the articles reported whether random housing (#4) was used, all articles received ‘Yes’, i.e. a low risk of bias, based on the signalling question ‘Is it unlikely that the outcome or the outcome measurement was influenced by not randomly housing the animals?’. Only one article^{(Reference Vikoren, Drotningsvik and Bergseth41)} reported the use of blinding in relation to performance bias (#5) and detection bias (#7). For the random outcome assessment (#6), we graded all articles as having an unclear risk of bias. When scoring the articles for incomplete outcome data (#8), most articles received ‘Yes’, i.e. a low risk of bias, as all the expected data were included or any missing data were accounted for. Ten articles received ‘No’, as they did not disclose why animals were missing from the reported outcomes^{(Reference Parolini, Vik and Busnelli35,Reference Chevrier, Mitchell and Rioux37–Reference Jacques, Deshaies and Savoie39,Reference Wang, Gagnon and Nair47,Reference Liaset, Madsen and Hao49,Reference Vik, Tillander and Skorve50,Reference Wang, Nair and Gagnon52,Reference Jacques, Deshaies and Savoie63,Reference Hurley, Galibois and Jacques66)} and three studies did not disclose the number of animals in each assessment^{(Reference Demonty, Deshaies and Lamarche40,Reference Liaset, Hao and Jorgensen55,Reference Hosomi, Fukunaga and Arai69)} and therefore received grading as ‘Unclear’. For the last two entries; selective outcome reporting (#9) and other sources of bias (#10), we graded all articles as having an unclear risk of bias. To summarise, the included articles were scored with one or two ‘No’ and between one and five ‘Yes’, with the rest of the scores being ‘Unclear’. Since few of the entries in the SYRCLE’s tool were addressed, we conclude that the risk of bias was unclear for the included articles.

All included articles were assessed for the quality of evidence for the primary outcomes (online Supplementary Table 3). All articles were peer reviewed (#1), included information on the animal model and strain used (#2) and gave details on the statistical method (#7) and descriptive statistics with a measure of variability (#8) for the serum/plasma TC concentration. All articles except one^{(Reference Moriya, Hosokawa and Miyashita36)} stated the sex of the experimental animals (#3). Only five articles^{(Reference Vikoren, Drotningsvik and Bergseth31,Reference Vikoren, Drotningsvik and Mjos32,Reference Vikoren, Drotningsvik and Bergseth41,Reference Drotningsvik, Mjos and Pampanin42,Reference Drotningsvik, Vikoren and Mjos44)} provided information about husbandry conditions and actions to improve animal welfare of the experimental animals (#4). For item #5, i.e. description of the procedures, we checked for information on when the blood was sampled, any use of anaesthesia when blood was sampled for TC analysis and if information regarding the prandial condition of the animals were provided. All articles described when blood was collected for TC analysis, but three articles did not provide information on prandial status^{(Reference Wergedahl, Liaset and Gudbrandsen68–Reference Hosomi, Miyauchi and Yamamoto70)} and one article gave no information regarding any use of anaesthesia at the time of sampling^{(Reference Liaset, Hao and Jorgensen55)}. Eighteen articles described the analysis of serum/plasma TC, including the name and brand of assays and kits (#6), while the other articles referred to the laboratory that conducted the analysis or the instrument that was used. Compliance with animal welfare regulations (#9) was declared in all articles except the two oldest; Jacques et al. 1986^{(Reference Jacques, Deshaies and Savoie39)} and Jacques et al. 1988^{(Reference Jacques, Deshaies and Savoie63)}. Statement of potential conflict of interests (#10) was not declared in twenty-one articles. The overall quality was high for the included articles, with a mean score of 8 (range 6–10).

Details on the rodent models

The included papers used the following rat models: Wistar rats^{(Reference Jacques, Leblanc and Papineau38,Reference Narayan, Yamaguchi and Hosokawa43,Reference Affane, Louala and El Imane Harrat46,Reference Hosomi, Nishimoto and Kobayashi48,Reference Liaset, Madsen and Hao49,Reference Louala, Hamza-Reguig and Benyahia-Mostefaoui51,Reference Hosomi, Maeda and Ikeda54,Reference Liaset, Hao and Jorgensen55,Reference Hosomi, Fukunaga and Arai57–Reference Hosomi, Fukunaga and Arai59,Reference Hosomi, Fukunaga and Arai65,Reference Hosomi, Fukunaga and Arai69,Reference Hosomi, Miyauchi and Yamamoto70)} , Sprague Dawley rats^{(Reference Jacques, Deshaies and Savoie39,Reference Demonty, Deshaies and Lamarche40,Reference Song, Hooiveld and Li45,Reference Mizushige, Kawabata and Uozumi61–Reference Jacques, Deshaies and Savoie63,Reference Hurley, Galibois and Jacques66,Reference Shukla, Bettzieche and Hirche67)} or obese Zucker fa/fa rats^{(Reference Vikoren, Drotningsvik and Bergseth31,Reference Vikoren, Drotningsvik and Bergseth41,Reference Drotningsvik, Mjos and Pampanin42,Reference Drotningsvik, Vikoren and Mjos44)} or combined studies in Zucker fa/fa rats with Wistar rats^{(Reference Wergedahl, Liaset and Gudbrandsen68)} or Long-Evans rats^{(Reference Vikoren, Drotningsvik and Mjos32)} (Table 1). Several mouse models were utilised: C57BL/6J mice^{(Reference Moriya, Hosokawa and Miyashita36,Reference Wang, Gagnon and Nair47,Reference Vik, Tillander and Skorve50,Reference Wang, Nair and Gagnon52,Reference Mijiti, Mori and Huang53)} , apolipoprotein E null mice (apoE^-/-, established in strain C57BL/6)^{(Reference Parolini, Vik and Busnelli35)}, ApoB100 only, LDL receptor knockout male mice (LDLR^-/-/ApoB^100/100)^{(Reference Chevrier, Mitchell and Rioux37)}, transgenic mice expressing human tumour necrosis factor-α (hTNFα, established in strain C57BL/6)^{(Reference Bjorndal, Burri and Wergedahl34)}, leptin-deficient ob/ob mice^{(Reference Maeda, Hosomi and Yokoyama64)} and KK-A ^y obese mice^{(Reference Maeda, Hosomi and Fukuda56,Reference Maeda, Hosomi and Koizumi60)} .

The Wistar rat, the Sprague Dawley rat, the Long-Evans rat and the C57BL/6J mouse are regarded as being normocholesterolemic when fed a regular diet. The obese Zucker fa/fa rat, which is one of the most widely used rat models for studies of metabolic complications and for possible treatments of obesity in humans, is hyperphagic and spontaneously develops metabolic abnormalities including elevated concentrations of serum LDL-, HDL- and VLDL-cholesterol^{(Reference Bray71)} and has impaired bile secretory function^{(Reference Pizarro, Balasubramaniyan and Solis72)}. The apolipoprotein E null mouse^{(Reference Zhang, Reddick and Piedrahita73)} and the LDLR-deficient mouse^{(Reference Powell-Braxton, Veniant and Latvala74)} are hypercholesterolemic. The KK mouse carrying the yellow obese gene A ^y (KK-A ^y ) is obese and develops type 2 diabetes^{(Reference Iwatsuka, Shino and Suzuoki75)}. The leptin-deficient ob/ob mouse is obese, has elevated cholesterol in circulation and impaired transport of cholesterol from circulation to faeces^{(Reference Duong, Uno and Nankivell76)}. The transgenic hTNFα mouse experiences metabolic alterations induced by inflammation^{(Reference Hayward, Jones and Saparov77)} but does not have elevated serum cholesterol concentration^{(Reference Glosli, Gudbrandsen and Mullen78)}.

Details on the designs of the diets

All reviewed articles included a control group that was fed a diet with casein as the sole protein source for direct comparison of the intervention fish protein source. The dietary protein content was similar in the control (casein) diet and the fish intervention diet for all included studies. The dietary total protein content was between 15 and 25 wt% for all described experiments. All studies used semi-purified diets, i.e. diets combining a fish protein source or casein with pure ingredients or chemicals with a varying degree of refinement.

The reviewed articles show great heterogeneity regarding both the fish species investigated, and which parts of the fish that were used and how these were processed. Nine varieties of fish were explored, and the proteins from fish replaced between 15 and 100 % of the casein in the experimental diets, as detailed in Table 1. Articles used muscles from cod that were defatted^{(Reference Jacques, Leblanc and Papineau38–Reference Demonty, Deshaies and Lamarche40,Reference Jacques, Deshaies and Savoie63,Reference Hurley, Galibois and Jacques66)} or lyophilised^{(Reference Vikoren, Drotningsvik and Bergseth41)}, muscles from pollock that were defatted^{(Reference Hosomi, Nishimoto and Kobayashi48,Reference Hosomi, Maeda and Ikeda54,Reference Hosomi, Fukunaga and Arai57,Reference Hosomi, Fukunaga and Arai58,Reference Maeda, Hosomi and Koizumi60–Reference Kawabata, Mizushige and Uozumi62,Reference Maeda, Hosomi and Yokoyama64,Reference Shukla, Bettzieche and Hirche67,Reference Hosomi, Fukunaga and Arai69)} or defatted and hydrolysed^{(Reference Hosomi, Fukunaga and Arai59)} and frames from saithe^{(Reference Liaset, Madsen and Hao49)} or water-soluble (press liquid) meal from headed and gutted blue whiting^{(Reference Drotningsvik, Vikoren and Mjos44)}. Also, hydrolysed muscles^{(Reference Chevrier, Mitchell and Rioux37,Reference Wergedahl, Liaset and Gudbrandsen68)} , lyophilised muscles^{(Reference Vikoren, Drotningsvik and Bergseth31,Reference Vikoren, Drotningsvik and Mjos32)} , hydrolysed by-products^{(Reference Parolini, Vik and Busnelli35,Reference Drotningsvik, Mjos and Pampanin42,Reference Vik, Tillander and Skorve50,Reference Liaset, Hao and Jorgensen55)} and protamine (from milt)^{(Reference Mijiti, Mori and Huang53,Reference Hosomi, Fukunaga and Arai65,Reference Hosomi, Miyauchi and Yamamoto70)} from salmon were investigated. From herring, hydrolysed by-products^{(Reference Drotningsvik, Mjos and Pampanin42)}, spray-dried roe and milt^{(Reference Bjorndal, Burri and Wergedahl34)}, defatted roe protein^{(Reference Moriya, Hosokawa and Miyashita36)}, milt protein hydrolysate^{(Reference Wang, Gagnon and Nair47)} and dried or hydrolysed milt^{(Reference Wang, Nair and Gagnon52)} were tested. Studies were also conducted using defatted muscles^{(Reference Louala, Hamza-Reguig and Benyahia-Mostefaoui51)}, defatted and hydrolysed muscles^{(Reference Affane, Louala and El Imane Harrat46)}, defatted and hydrolysed by-products^{(Reference Affane, Louala and El Imane Harrat46)} from sardine, defatted muscles from tuna^{(Reference Narayan, Yamaguchi and Hosokawa43,Reference Hosomi, Maeda and Ikeda54,Reference Maeda, Hosomi and Fukuda56)} and defatted muscles from carp^{(Reference Song, Hooiveld and Li45)}.

The proteins from fish were included in a variety of diet designs, including diets with regular contents of fat and simple carbohydrates (≤ 7 wt% fat and ≤ 10 wt% of simple carbohydrates)^{(Reference Vikoren, Drotningsvik and Bergseth31,Reference Vikoren, Drotningsvik and Mjos32,Reference Moriya, Hosokawa and Miyashita36,Reference Vikoren, Drotningsvik and Bergseth41–Reference Song, Hooiveld and Li45,Reference Hosomi, Nishimoto and Kobayashi48,Reference Louala, Hamza-Reguig and Benyahia-Mostefaoui51,Reference Hosomi, Fukunaga and Arai57–Reference Hosomi, Fukunaga and Arai59,Reference Hosomi, Fukunaga and Arai65,Reference Hosomi, Fukunaga and Arai69,Reference Hosomi, Miyauchi and Yamamoto70)} , high-fat diets (≥ 30 wt% fat) with regular content of simple carbohydrates (≤ 10 wt%)^{(Reference Wang, Gagnon and Nair47,Reference Wang, Nair and Gagnon52,Reference Mijiti, Mori and Huang53,Reference Hosomi, Fukunaga and Arai69)} , high-fat diets with moderately high content of simple carbohydrates (11–54 wt%)^{(Reference Chevrier, Mitchell and Rioux37)}, diets with moderately high-fat content (8–29 wt%) and regular content of simple carbohydrates (≤ 10 wt%)^{(Reference Bjorndal, Burri and Wergedahl34,Reference Jacques, Deshaies and Savoie39,Reference Demonty, Deshaies and Lamarche40,Reference Affane, Louala and El Imane Harrat46,Reference Liaset, Madsen and Hao49,Reference Vik, Tillander and Skorve50,Reference Hosomi, Maeda and Ikeda54,Reference Maeda, Hosomi and Fukuda56,Reference Maeda, Hosomi and Koizumi60,Reference Jacques, Deshaies and Savoie63,Reference Maeda, Hosomi and Yokoyama64)} , diets with moderately high-fat content and moderately high content of simple carbohydrates^{(Reference Parolini, Vik and Busnelli35,Reference Jacques, Leblanc and Papineau38,Reference Liaset, Hao and Jorgensen55,Reference Mizushige, Kawabata and Uozumi61,Reference Kawabata, Mizushige and Uozumi62,Reference Shukla, Bettzieche and Hirche67,Reference Wergedahl, Liaset and Gudbrandsen68)} and diets with moderately high-fat content (10–30 %) and high content of simple carbohydrates (≥ 55 wt%)^{(Reference Hurley, Galibois and Jacques66)}. In some of the articles, cholesterol^{(Reference Jacques, Leblanc and Papineau38,Reference Demonty, Deshaies and Lamarche40,Reference Hurley, Galibois and Jacques66,Reference Shukla, Bettzieche and Hirche67)} or a combination of cholesterol and cholate^{(Reference Jacques, Deshaies and Savoie39,Reference Louala, Hamza-Reguig and Benyahia-Mostefaoui51,Reference Hosomi, Maeda and Ikeda54,Reference Hosomi, Fukunaga and Arai58,Reference Hosomi, Fukunaga and Arai59,Reference Jacques, Deshaies and Savoie63,Reference Hosomi, Fukunaga and Arai69)} was added to the diets to promote elevation in the circulating cholesterol concentration. Cholate reduces the cholesterol catabolism in the liver by inhibiting the production of bile acids through down-regulation of hepatic CYP7A1 activity. None of the studies added cholesterol in order to balance the cholesterol content between the diets.

Cholesterol metabolism

All reviewed articles describe the effects of fish or fish protein intake, i.e. from muscles or by-products, on serum/plasma TC concentration measured at the endpoint. In addition, many articles describe the effects on HDL-cholesterol, LDL-cholesterol and hepatic TC concentration, the daily faecal excretion of TC and bile acids and a selection of enzymes relevant to cholesterol metabolism, as detailed in Table 1. Only in one article was the circulating TC concentration measured at baseline after rats were fed a cholesterologenic pre-diet to induce hypercholesterolaemia, but statistical testing to compare TC concentrations at baseline and endpoint was not conducted^{(Reference Louala, Hamza-Reguig and Benyahia-Mostefaoui51)}. Below we present the findings reported in the reviewed articles, organised by family of the fish sources.

Meta-analyses

The main meta-analysis of differences in endpoint serum/plasma TC concentration in the intervention and comparator groups includes data from all thirty-nine reviewed articles, representing findings in 512 rodents in the intervention groups and 423 rodents in the comparator groups (Fig. 2). The meta-analysis revealed that diets containing proteins from fish resulted in lower circulating TC concentration in rodents compared with their respective control group, with a mean difference −0·24 mmol/l and 95 % confidence interval −0·34, −0·15 mmol/l, with an overall test for effect Z = 5·10 and P < 0·00001. When testing for statistical heterogeneity, χ ² was 222·05 (P < 0·00001) and I² was 71 %, thus reflecting that both the direction of the between-group differences and the magnitude of effect were highly heterogeneous in our meta-analysis.

Fig. 2. Meta-analysis using a random effects model presenting the effects of intake of proteins from fish muscles or fish residuals on circulating total cholesterol concentration (mmol/l) as a forest plot. The studies are described as [type of fish and dose], [fish fraction], [description of diet], [w/o cholesterol and cholate], [rodent strain], [first author and year of publication]. Type of fish and dose: the dose of fish protein is shown as percent of total dietary content; E1, spine hydrolysed with Acid Protease A; E2, spine hydrolysed with Umamizyme; E4, backbone and heads hydrolysed with Alcalase fish fraction; B, by-products; M, muscle; MiH, hydrolysed milt; MiP, dried milt powder; P, protamine; R, roe. Description of diet: HFD, high-fat diet; HG, headed and gutted; MFD, medium-fat diet; RGD, regular diet; s, high sucrose content; c, high corn starch content; b, added beef tallow; m, added menhaden oil; f, added fish oil; t, added tuna oil; 1, animals were fasted at euthanisation; 2, animals were nonfasted at euthanisation with or without enrichment with cholesterol and cholate: NACC, diet is not added cholesterol or cholate; AC, diet is added cholesterol; ACC, diet is added cholesterol and cholate. Rodent strain: apoE, apolipoprotein E null mice; C57, C57BL/6J mice; hTNFa, transgenic hTNFα mice; KKAy, KK-A^y mice; LDLR, ApoB100 only LDL-receptor knockout mice; LE, Long-Evans rats; obob, leptin-deficient ob/ob mice; sd, Sprague Dawley rats; W, Wistar rats; Z, Zucker fa/fa rats.

To further explore the heterogeneity in the main meta-analysis, a series of subgroup analyses were conducted (Fig. 3 and online Supplementary Table 4). First, diets containing proteins from fish muscles or fish by-products were tested individually, and these meta-analyses gave a mean difference (95 % CI) of –0·27 (–0·37, 0·17) mmol/l for serum/plasma TC concentration for fish muscle proteins and -0·16 (-0·38, 0·06) mmol/l for fish by-product proteins, with P < 0·00001 (Z = 5·16) and P 0·16 (Z = 1·40), respectively. The heterogeneity was still high for both analyses, with Chi² = 149·51 (P < 0·00001) and I² = 72 % for fish muscle proteins and Chi² = 72·47 (P < 0·00001) and I² = 71 % for fish by-product proteins. When the subgroups were compared, no significant difference was detected between the effect of fish muscles and fish by-products on circulating TC concentration (P 0·37).

Fig. 3. Subgroup analyses for meta-analysis using a random effects model presenting the effects of intake of proteins from fish muscles or fish residuals on circulating total cholesterol concentration (mmol/l) as a forest plot. We refer to online Supplementary Table 4 for further details on heterogeneity, effect and statistical significance.

Next, subgroup analyses by fish family were conducted, including both fish muscle proteins and fish by-product proteins combined to investigate the effect on circulating TC concentration. The subgroup of rodents fed proteins from the Gadidae family (cod, pollock, saithe and blue whiting) were the largest contributors to the main meta-analysis, with a total of 514 rodents included in the analysis. The difference between the intervention and comparator groups was significant (P < 0·00001), with a mean difference of −0·28 (95 % CI; −0·38, −0·19) mmol/l and I² of 60 %. For the subgroup fed proteins from the Salmonidae family (205 rodents included), the test for overall effect was borderline significant with P 0·05 and a mean difference of −0·24 (–0·48, 0·00) mmol/l, whereas the subgroups fed dietary proteins from the Clupeidae family (sardine and herring, a total of 150 rodents included) did not affect circulating TC concentration (P 0·44 and mean difference −0·14 (–0·50, 0·22) mmol/l). Subgroup analyses were not conducted for the groups fed proteins from the Scombriadae (tuna) and Cyprinidae (carp) families due to their small sample sizes. The comparison of the effect of Gadidae, Salmonidae and Clupeidae families on circulating TC concentration showed no significant differences between the subgroups (P 0·73).

A subgroup analysis was conducted to investigate whether a total replacement of casein or a replacement of 50 % or less of casein with proteins from fish had different impact on the circulating TC concentration. Both total and partial replacement with proteins from fish resulted in significantly lower serum/plasma TC concentration (mean difference −0·32 (–0·49, −0·15) with P 0·0002, and −0·24 (-0·35, −0·13) with P < 0·0001, respectively), with no significant difference between the subgroups (P 0·29).

We expected that the circulating TC concentration would increase over time in especially two subcategories; the genetically obese rodents which spontaneously develop hypercholesterolaemia after birth (the leptin-resistant Zucker fa/fa rats and the leptin-deficient ob/ob mice) and rodents fed diets enriched with cholesterol alone or in combination with cholate. A subgroup analysis of studies using these rodents v. the rest of the included studies revealed that both subgroups had significantly lower mean differences in serum/plasma TC concentration when intervention and comparator groups were compared; for the subgroup comprising fa/fa rats, ob/ob mice and rodents fed diets enriched with cholesterol/cholate, the mean difference was −0·53 (–0·72, −0·35) with P < 0·00001, and for the subgroup consisting of the remaining groups the mean difference was −0·13 (-0·23, −0·02) with P 0·02. The difference between the subgroups was significant (P 0·0002), with the strongest effect seen in the former subgroup.

The sensitivity analysis, conducted through a leave-one-out analysis of studies with the highest positive and negative effect sizes, and the studies with the highest risk of bias or lowest quality of evidence were excluded sequentially from the meta-analysis, did not alter the outcome measure (data not presented).

The asymmetry of the funnel plot (Fig. 4) was visually inspected to assess the possibility that relevant studies are missing from our meta-analysis. We consider the asymmetry to be low, with four data points from the article by Hurley et al. 1995^{(Reference Hurley, Galibois and Jacques66)} standing out towards the lower left corner of the plot. Although the use of a funnel plot is not recommended when the statistical heterogeneity is large (> 50 %)^{(Reference Ioannidis and Trikalinos79)}, as is the case for the present meta-analysis with I² of 71 %, the low asymmetry in the funnel plot suggests that the risk for publication bias is low in the current meta-analysis.

Fig. 4. Funnel plot showing the effect estimate with 95 % CI for the effect of intake of diets containing proteins from fish on circulating total cholesterol concentration.

Dietary intake

If a complete or partial replacement of casein with proteins from fish affected the dietary intake, this may in turn have influenced biochemical parameters and physiological outcomes. Most of the articles reported either the energy intake or the amount of feed consumed (online Supplementary Table 1). The majority of studies, i.e. thirty-three of the thirty-nine articles included, reported no differences in dietary intake between the intervention and control groups^{(Reference Vikoren, Drotningsvik and Bergseth31,Reference Vikoren, Drotningsvik and Mjos32,Reference Bjorndal, Burri and Wergedahl34,Reference Moriya, Hosokawa and Miyashita36–Reference Drotningsvik, Vikoren and Mjos44,Reference Wang, Gagnon and Nair47–Reference Liaset, Madsen and Hao49,Reference Louala, Hamza-Reguig and Benyahia-Mostefaoui51–Reference Hosomi, Maeda and Ikeda54,Reference Maeda, Hosomi and Fukuda56–Reference Shukla, Bettzieche and Hirche67,Reference Hosomi, Fukunaga and Arai69,Reference Hosomi, Miyauchi and Yamamoto70)} . The dietary intake was higher in Sprague Dawley rats fed a diet with 100 % of protein from carp muscles^{(Reference Song, Hooiveld and Li45)} and was lower in Wistar rats consuming hydrolysed salmon by-products^{(Reference Liaset, Hao and Jorgensen55)} or hydrolysed muscles or by-products from sardine^{(Reference Affane, Louala and El Imane Harrat46)} as the sole protein sources, when compared with rats fed casein diets. Two articles did not declare any information on the dietary intake^{(Reference Parolini, Vik and Busnelli35,Reference Wergedahl, Liaset and Gudbrandsen68)} , and in one article the feed intake was measured but statistical testing was not performed^{(Reference Vik, Tillander and Skorve50)}.

Adiposity and bodyweight gain

A complete or partial replacement of casein with proteins from fish in the diet may affect adiposity in rodents, which in turn may affect the cholesterol metabolism. Adiposity was quantified in twenty-five articles (online Supplementary Table 1), and of these, twenty articles reported that the absolute or relative weights of adipose tissues were similar in rodents fed proteins from fish when compared with the control group^{(Reference Vikoren, Drotningsvik and Bergseth31,Reference Vikoren, Drotningsvik and Mjos32,Reference Moriya, Hosokawa and Miyashita36,Reference Chevrier, Mitchell and Rioux37,Reference Vikoren, Drotningsvik and Bergseth41,Reference Drotningsvik, Mjos and Pampanin42,Reference Song, Hooiveld and Li45,Reference Hosomi, Nishimoto and Kobayashi48,Reference Liaset, Madsen and Hao49,Reference Wang, Nair and Gagnon52,Reference Hosomi, Maeda and Ikeda54,Reference Maeda, Hosomi and Fukuda56–Reference Kawabata, Mizushige and Uozumi62,Reference Hosomi, Fukunaga and Arai65,Reference Hosomi, Fukunaga and Arai69)} . A lower weight of one or multiple white adipose tissues was reported in Wistar rats consuming a diet with either hydrolysed salmon by-products as the sole protein source^{(Reference Liaset, Hao and Jorgensen55)}, a diet with protamine as 25 % of protein^{(Reference Hosomi, Miyauchi and Yamamoto70)}, or hydrolysed sardine by-products as 100 % of dietary protein^{(Reference Affane, Louala and El Imane Harrat46)}, and in C57BL/6J mice fed a diet with 20 % of protein from protamine^{(Reference Mijiti, Mori and Huang53)}. One article described a higher relative weight of epididymal WAT in leptin-deficient ob/ob mice fed 100 % of protein from pollock muscles, but was similar to the control group with regard to relative weights of mesenteric, perirenal + retroperitoneal and inguinal WAT^{(Reference Maeda, Hosomi and Yokoyama64)}.

Thirty-three articles provided information regarding bodyweight gain (often referred to as growth) during the intervention period. The reported bodyweight gain is presented in Supplementary Table 1, but is not further described here since differences in bodyweight do not necessarily reflect dissimilarities in adiposity or muscle mass between the experimental groups.