How Protein Supplements May Improve Athletic Performance and Recovery After Exercise

A comprehensive search strategy was developed using Medical Subject Headings (MeSH) and free-text search terms to systematically search the SPORTDiscus, PubMed, Ovid, Web of Science, MEDLINE, CINAHL, and Scopus databases on September 26, 2024. The keywords and subject headings were finalised through discussion between the two authors (SZ & XZ). Detailed search strings for each database are provided in Supplementary File S1. A total of 6,129 studies were retrieved using the Covidence online tool for systematic review.

Following the PICOS principle, non-human studies and non-comparative studies were excluded. Eligible studies were randomised controlled trials (RCTs) that included protein supplements. The studies without a control group or protein group were not considered. Non-original studies (e.g. letters, reviews, or editorials) and studies lacking available data for extraction were also excluded.

Participants in RCTs must be athletes. Studies were only included if the experimental group received a protein supplement and the control group did not receive any form of protein supplementation (e.g. placebo, carbohydrate, or no supplement). Trials comparing different types, timing, or doses of protein supplements without a non-protein control group were excluded to ensure the presence of a proper comparison group. Outcomes in studies were required to be related to athletic performance, including endurance performance and muscle strength, or post-exercise recovery, encompassing glycogen resynthesis and fatigue.

The automatic tool in Covidence recommended by PRISMA was used to generate the flow diagram [39]. Two reviewers (SZ & XZ) independently screened titles and abstracts, followed by full texts against eligibility criteria using the Covidence online tool.

The risk of bias for all included studies was independently assessed using the guidelines and criteria outlined in the Cochrane Handbook for Systematic Reviews of Interventions. Two authors (SZ & XZ) assessed the included studies through the Cochrane risk of bias (RoB) criteria in RCTs within Covidence. Seven areas of bias were evaluated: (1) random sequence generation; (2) allocation concealment; (3) blinding of participants and personnel; (4) blinding of outcome assessment; (5) incomplete outcome data; (6) selective reporting; and (7) other bias. The risk of bias was classified as low, unclear, or high. After independent assessments, the authors reached a consensus through discussion. Final results were recorded in an Excel template and input into R software to create risk of bias summary plots using the robvis package [40]. Studies with more than two and fewer than four areas marked as unclear risk were classified as moderate risk overall.

To assess publication bias, we used the PublicationBias package in R [41,42], which provides a sensitivity analysis approach that does not rely on funnel plot symmetry. Instead, it estimates how strong publication bias would need to be to reduce the observed effect size or its confidence interval to a specified threshold.

Assuming that statistically significant and positive results were more likely to be published (favor_positive = TRUE), we created a significance-based funnel plot to visually compare affirmative (significant positive) and non-affirmative (non-significant or negative) studies. The worst-case estimate, based on non-affirmative studies only, provides a benchmark for evaluating robustness. We also calculated the s-value, defined as the minimum selection ratio (η) required to attenuate the effect to the null. An s-value of “not possible” indicates that no plausible publication bias could fully explain the observed effect.

To supplement this newer method, we also applied traditional techniques using the metafor package. We generated a standard funnel plot, conducted Egger's test, and applied the trim-and-fill method when Egger's test yielded p < 0.05, providing additional evidence from a Frequentist perspective.

Data were extracted independently by two authors (SZ & XZ) using Covidence, with conflicts resolved through discussion with the third author (ZN). For each study, characteristics including intervention group, control group, first author, publication year, study design, country, participants' ages, body mass index (BMI), athlete type, protein type, extra protein dose, overall protein dose, duration, washout or follow-up period, and outcome measure were extracted. The outcomes included time to exhaustion, lower body strength, upper body strength, one-repetition maximum (1 RM), cycling time trials, maximum voluntary contraction (MVC), counter-movement jump (CMJ), anaerobic peak and average power, vertical jump, cycling distance, hand grip strength, maximum speed, average speed, blood glucose, muscle glycogen, blood lactate, muscle soreness, physical and mental fatigue, and VO₂max.

Data were presented as mean ± standard deviation (M ± SD). When data were not presented as exact numbers, GetData Graph Digitiser [43] was used to extract data from graphs. We used an online tool called Meta-analysis Accelerator to convert data that was not initially in M ± SD format and to calculate the change values (mean and standard deviation) accounting for pre- and post-intervention baseline measurements [44]. Since none of the studies provided correlation coefficients, a correlation of 0.5 was assumed for all trials, following the recommendation of Follmann et al [45].

The Metafor package in R was used to calculate the standardised mean difference (SMD) according to the formula:

Bayesian mixed-effects models implemented in the R package brms [46] were used to analyse variation in effect sizes. We fitted the models assuming a normal distribution and included random effects for within-study ID and between-study ID to account for within-study and between-study heterogeneity in each result according to two formulas:

In addition to heterogeneity indicators, we tested two indices through the metainc package [47] including the dissimilarity index (DI) and the across-studies inconsistency (ASI) to assess the potential inconsistency in each model. These indices were considered more suitable for Bayesian meta-analysis compared to traditional heterogeneity indicators such as I² or Q-test [48]. A DI ≥ 50% and an ASI ≥ 25% were considered indicative of important inconsistency. These indices differ from previously existing measures by considering effect size (ES) in the context of decision thresholds (DTs). We set three decision thresholds as suggested for SMD, namely 0.2, 0.5, and 0.8 [48].

Weakly informative priors (mean = 2, sd = 0.5) were used for the random effects. While not universally optimal, weakly informative priors are widely regarded as a best-practice default in Bayesian meta-analysis, especially when prior knowledge is limited [49]. Athletic performance and post-exercise recovery were analysed separately, with each data point categorised by protein supplementation strategy (PRO vs. PRO-CHO), protein source, additional protein dose, total protein intake, energy matching, fasted state, study design (parallel or crossover), acute or chronic ingestion and protein timing. Secondly, to explore the sources of heterogeneity, we further categorised the outcome measures by establishing muscle model, endurance model, glycogen model, and fatigue model. We then applied interaction models using the brms package to analyse the data. This approach helped reduce data complexity and mitigate heterogeneity.

Fifteen models were fitted for each of the athletic performance and post-exercise recovery datasets. Ten Bayesian models (Fasted model, Energy model, Design model, Acute-chronic model, Endurance model, Muscle model, Glycogen model, Fatigue model and Interaction model) used in this study were established post hoc. Model specifications were refined based on reviewer feedback and adjustments made during the analytical process to address initial errors and improve model fit. Details of the fitted models are provided below:

Four chains of 10,000 iterations were applied for each model. To assess Markov chain convergence in this study, we used the Rhat statistic as the sole diagnostic metric. The Rhat value compares between-chain and within-chain variances, with values near 1 indicating adequate convergence of the chains. The degree of support for a difference from zero was assessed using Bayes Factors (BFs). The function bayesfactor_parameters from the bayestestR package was used to implement the reciprocal of the Savage-Dickey density ratio, which was used to estimate BFs [51]. Bayes Factors were interpreted on a conventional scale, with values exceeding 3 indicating moderate evidence and those above 10 reflecting strong support for the presence of an effect [52]. For complex hierarchical models, computing the 95% highest-density interval (HDI), which represents the most plausible outcomes in the posterior distribution, is seamless, as opposed to approximating p-values and confidence intervals, which involve additional assumptions [53].

The quality of the protein evidence was assessed using the GRADE approach, which evaluates the risk of bias, inconsistency, indirectness, and imprecision of effect estimates. The GRADE approach classifies the quality of evidence as high, moderate, low, or very low. Furthermore, in the evaluation of inconsistency, in addition to the heterogeneity indicator (I²), we also used the previously calculated inconsistency indicator (DI & ASI) to help judge the quality of the results. If the inconsistency indicator showed moderate or above inconsistency, the level of the results was downgraded by at least one level.

In this study, a moderation analysis was conducted using the brms package in R, incorporating four trait moderator variables: age, weight, percentage of women among participants, and sample size. Additionally, three dose moderators concerning duration of protein supplementation, additional protein supplement dose, and total daily protein intake were incorporated into a moderation analysis to investigate the optimal daily protein consumption and intervention duration for athletes. The analysis utilised a Bayesian framework, allowing for the estimation of effect sizes while accounting for measurement error and hierarchical data structure.

Figure 1 illustrates the selection process and information sources. Covidence automatically removed 1,079 duplicate records, and two duplicates were removed manually. A total of 6,129 studies were screened, with 1,270 marked as ineligible by an automatic tool and 3,662 manually excluded as irrelevant. After full-text screening of 116 studies, 54 were excluded, resulting in 62 studies being included in the meta-analysis. An additional 13 RCTs were identified through citation searching, yielding a final total of 75 RCTs.

The detailed characteristics of each included study are provided in Supplementary File S2. A total of 75 studies involving 1206 athletes (220 females and 916 males) were included, with 70 athletes lacking gender information [54–58]. Twenty-nine studies were randomised controlled trials, and 46 were randomised crossover trials. Geographically, 37 studies (49%) were conducted in Europe, 21 (28%) in North America, seven (9%) in Oceania, six (8%) in Asia, three (4%) in Africa, and one (1%) in South America.

Forty-four studies used protein-carbohydrate supplements as the intervention, while 27 used pure protein supplements. Two studies implemented high-protein diets [59,60] and two utilised enriched protein yoghurts [54] or protein plus probiotic supplements [61] as their interventions. Eleven studies did not report daily energy intake for either the intervention or control groups [10,27,61–69].

Among the 75 included studies, 8 studies (10.67%) did not specify the protein source and another 8 studies (10.67%) used supplements combining multiple protein types. Thirty-five studies (46.67%) used whey protein, six (8.00%) used soy protein, one (1.33%) used branched-chain amino acid (BCAA) supplements, nine (12.00%) used casein protein, and ten (13.33%) used milk protein. Additionally, one study each (1.33%) used hydrolysed collagen, egg white protein, wheat protein, or beef protein.

The participants' sports disciplines included cycling, triathlon, wrestling, boxing, orienteering, running, field hockey, weightlifting, football, soccer, basketball, rugby, volleyball, track and field, sailing, and judo. Nine studies did not report the athletic discipline of the participants [54,59,60,70–75]. Regarding additional protein ingestion from supplements, six studies reported more than 2 g/kg/day [76–81], four studies reported between 1 and 2 g/kg/day [70,82–84], and the remainder reported less than 1 g/kg/day.

The risk-of-bias summary is illustrated in Figure 2, and the specific risk-of-bias graph in each study is provided in Supplementary File S3. The Cochrane Risk of Bias tool (RoB) was utilised to assess the included studies, with results visualised using the R package robvis. Some studies were rated as high risk because they used an inappropriate randomisation method, such as block randomisation (4%) [55,75,85], and some authors explicitly acknowledged that the lack of blinding could introduce bias (4%) [86–88]. The majority of studies were rated as having some concerns in the domain of blinding of outcome assessors due to insufficient reporting (84%). Some studies were rated as having some concerns in allocation concealment and blinding of participants and personnel due to a lack of information. Overall, nearly 75% of studies were assessed as low risk of bias, fewer than 15% as moderate risk, and fewer than 10% as high risk.

The meta-analysis was divided into fifteen sections, each presenting four outcome measures: muscle strength, endurance performance, glycogen resynthesis, and fatigue recovery. Detailed results in each model are presented in Supplementary File S4. The Markov chain convergence in all fifteen models showed good convergence based on the Rhat parameter. The Rhat value in all results was close to 1.00. Therefore, we did not present the results of Markov chain convergence in the text.

In the null model, the overall effect size was calculated for four outcomes. All forest plots and density plots of four outcomes in the null model are presented in Supplementary File S5. Sixty-four studies involving 1,048 athletes were included in the analysis of endurance performance. The Bayesian meta-analysis showed a statistically significant effect [μ(SMD): 0.21, 95% CI: 0.07 to 0.35; HDI: 0.07 to 0.34; BF: 2.76], with high between-study heterogeneity and low within-study heterogeneity [τ_within: 0.10, within I²: 3.85%, τ_between: 0.43, between I²: 96.15%]. Thirty studies, including 548 athletes, were included in the analysis of muscle strength in the null model, with no statistically significant effect [μ(SMD): 0.31, 95% CI: –0.01 to 0.64; HDI: –0.01 to 0.63; BF: 0.54], with high between-study heterogeneity and low within-study heterogeneity [τ_within: 0.06, within I²: 0.52%, τ_between: 0.83, between I²: 99.48%].

Thirty-two studies with 425 athletes were included in the analysis of glycogen resynthesis in the null model. No statistically significant effect was observed in the meta-analysis [μ(SMD): 0.17, 95% CI: –0.01 to 0.35; HDI: –0.01 to 0.36; BF: 0.26], with low between-study heterogeneity and moderate within-study heterogeneity [τ_within: 0.31, within I²: 68.54%, τ_between: 0.21, between I²: 31.46%]. Forty-three studies containing 663 athletes were included in the analysis of post-exercise fatigue recovery in the null model. No statistically significant effect was observed in the meta-analysis [μ(SMD): 0.16, 95% CI: –0.01 to 0.33; HDI: –0.01 to 0.32; BF: 0.23], with high between-study heterogeneity and low within-study heterogeneity [τ_within: 0.16, within I²: 13.79%, τ_between: 0.40, between I²: 86.21%].

Five comparisons including protein supplements (PRO) vs. placebo (PLA), protein supplements (PRO) vs. carbohydrate supplements (CHO), protein plus probiotic vs. placebo (PLA), protein-carbohydrate supplements (CHOPRO) vs. carbohydrate supplements (CHO), and protein-carbohydrate supplements (CHOPRO) vs. placebo (PLA) were conducted.

The forest plot is shown in Figure 3. In athletic performance outcomes, including muscle strength and endurance performance, fifteen studies involving 334 athletes were included in the PRO vs. PLA comparison. Statistically significant effects were observed for both endurance performance [μ(SMD): 0.37, 95% CI: 0.02 to 0.71; HDI: 0.07 to 0.73; BF: 1.6] and for muscle strength [μ(SMD): 0.72, 95% CI: 0.18 to 1.27; HDI: 0.18 to 1.26; BF: 4.37]. Eight studies involving 98 athletes were included in the CHOPRO vs. PLA comparison. No statistically significant effect was observed for muscle strength, whereas a statistically significant effect was observed for endurance performance [μ(SMD): 0.57, 95% CI: 0.2 to 0.93; HDI: 0.19 to 0.93; BF: 7.69]. Moderate between-study heterogeneity and low within-study heterogeneity were observed in the supplements model for endurance [τ_within: 0.08, within I²: 2.22%, τ_between: 0.45, between I²: 70.09%] and muscle strength [τ_within: 0.06, within I²: 0.52%, τ_between: 0.79, between I²: 90.75%].

In post-exercise recovery outcomes, including glycogen resynthesis and post-exercise fatigue recovery, eight studies involving 146 athletes were included in the PRO vs. CHO comparison. No statistically significant effect was found for fatigue, whereas a statistically significant effect was observed for glycogen resynthesis [μ(SMD): 0.83, 95% CI: 0.21 to 1.46; HDI: 0.21 to 1.46; BF: 4.84]. Low between-study and within-study heterogeneity were observed in the supplements model for glycogen resynthesis [τ_within: 0.21, within I²: 13.15%, τ_between: 0.18, between I²: 9.92%] and fatigue [τ_within: 0.15, within I²: 7%, τ_between: 0.37, between I²: 42.86%].

The protein source model divided the data into ten protein types for the analysis of athletic performance outcomes. The forest plot is shown in Figure 4. Only one protein source (whey protein) showed statistically significant effects for both endurance performance [μ(SMD): 0.28, 95% CI: 0.07 to 0.49; HDI: 0.07 to 0.49; BF: 1.52] and muscle strength [μ(SMD): 0.53, 95% CI: 0.01 to 1.05; HDI: 0.004 to 1.05; BF: 1.04]. No statistically significant effects were observed in other types of protein sources. Moderate between-study heterogeneity and low within-study heterogeneity were observed in the supplements model for endurance [τ_within: 0.09, within I²: 2.39%, τ_between: 0.48, between I²: 68.06%] and muscle strength [τ_within: 0.06, within I²: 0.36%, τ_between: 0.94, between I²: 89.51%].

For post-exercise recovery outcomes, the data were divided into ten protein types. In all types of protein sources, no statistically significant effect was observed. Low between-study and within-study heterogeneity were observed in the protein source model for glycogen resynthesis [τ_within: 0.27, within I²: 15.8%, τ_between: 0.22, between I²: 10.49%] and fatigue [τ_within: 0.16, within I²: 7.41%, τ_between: 0.41, between I²: 46.22%].

In the protein dose model, data were divided into three types (0–1 g/kg, 1–2 g/kg, and 2–3 g/kg) based on the extra protein dose in daily protein supplement ingestion. The forest plot can be seen in Supplementary File S6. In athletic performance, only an extra protein dose ingested with 0–1 g/kg a day from protein supplements showed a statistically significant effect in endurance performance [μ(SMD): 0.25, 95% CI: 0.10 to 0.41; HDI: 0.10 to 0.40; BF: 4.99], with moderate between-study and low within-study heterogeneity [τ_within: 0.09, within I²: 2.78%, τ_between: 0.44, between I²: 66.37%]. Fifty-five studies with 901 athletes were included in this group.

In post-exercise recovery, only the extra protein dose ingested with 0–1 g/kg from protein supplements showed a small statistically significant effect in fatigue recovery [μ(SMD): 0.18, 95% CI: 0.02 to 0.36; HDI: 0.01 to 0.36; BF: 0.44], with low between-study and within-study heterogeneity [τ_within: 0.14, within I²: 7.73%, τ_between: 0.29, between I²: 33.35%]. Although the 95% HDI excluded zero, indicating a small but credible positive effect, the Bayes Factor (BF = 0.44) offered only weak evidence for the alternative hypothesis, suggesting the effect should be interpreted with caution. Thirty-five studies and 551 athletes were included in this group.

In the overall dose model, the data were divided into three types (0–1 g/kg, 1–2 g/kg, and 2–3 g/kg) based on the overall protein dose in daily protein intake from both protein supplements and three daily meals. The forest plot is illustrated in Supplementary File S6. In terms of athletic performance and post-exercise recovery, no statistically significant effects were observed for endurance, muscle strength, glycogen resynthesis, or fatigue recovery across any of the three protein dose categories.

In the protein timing model, data were divided into two parts (day and night). The forest plot is shown in Supplementary File S6. Only the group that consumed protein supplements during the day showed a statistically significant effect in endurance performance [μ(SMD): 0.25, 95% CI: 0.09 to 0.41; HDI: 0.09 to 0.40; BF: 3.97], with low between-study and within-study heterogeneity [τ_within: 0.08, within I²: 6.04%, τ_between: 0.14, between I²: 18.49%]. Twenty studies with 313 athletes were included in this group. No statistically significant effect was observed for muscle strength, glycogen resynthesis, and fatigue.

In the energy model, data were divided into two groups: energy-matched and energy-unmatched. The forest plot is presented in Supplementary File S6. Statistically significant effects on endurance [μ(SMD): 0.47, 95% CI: 0.24 to 0.70; HDI: 0.24 to 0.70; BF: 147.66] and muscle strength [μ(SMD): 0.52, 95% CI: 0.01 to 1.04; HDI: 0.01 to 1.04; BF: 0.99] were observed only in the energy-unmatched group. Low within-study heterogeneity and moderate between-study heterogeneity were observed in endurance performance [τ_within: 0.08, within I²: 2.35%, τ_between: 0.42, between I²: 64.66%], and low within-study heterogeneity and high between-study heterogeneity were detected in muscle strength [τ_within: 0.06, within I²: 0.47%, τ_between: 0.84, between I²: 91.73%].

In the results of post-exercise recovery, no statistically significant effect was found for either glycogen resynthesis or fatigue recovery.

In the fasted model, data were divided into two groups: fasted and fed. The corresponding forest plot is presented in Supplementary File S6. Regarding athletic performance, a statistically significant effect was observed only for endurance in the fed group [μ(SMD): 0.15, 95% CI: 0.03 to 0.26; HDI: 0.03 to 0.26; BF: 0.60], with low within-study and between-study heterogeneity [τ_within: 0.06, within I²: 4.18%, τ_between: 0.15, between I²: 26.13%]. Although the 95% HDI excluded zero, indicating a small positive effect of feeding on endurance performance, the Bayes Factor (BF = 0.6) provided anecdotal evidence in favour of the null hypothesis, suggesting that the evidence for a true effect remains weak.

For post-exercise recovery outcomes, a statistically significant effect was observed only for fatigue recovery in the fasted group [μ(SMD): 0.30, 95% CI: 0.06 to 0.54; HDI: 0.06 to 0.54; BF: 1.24], with low within-study and between-study heterogeneity [τ_within: 0.14, within I²: 9.36%, τ_between: 0.2, between I²: 13.08%].

In the design model, data were divided into two groups: parallel and crossover. The corresponding forest plot is presented in Supplementary File S6. Regarding athletic performance, a statistically significant effect was observed for endurance [μ(SMD): 0.39, 95% CI: 0.14 to 0.65; HDI: 0.14 to 0.65; BF: 5.78] and muscle strength [μ(SMD): 0.53, 95% CI: 0.01 to 0.97; HDI: 0.10 to 0.96; BF: 2.01] only in the parallel group. Low within-study heterogeneity and moderate between-study heterogeneity were observed in endurance performance [τ_within: 0.09, within I²: 2.78%, τ_between: 0.44, between I²: 66.37%], and low within-study heterogeneity and high between-study heterogeneity were found in muscle strength [τ_within: 0.06, within I²: 0.50%, τ_between: 0.81, between I²: 91.16%].

For post-exercise recovery outcomes, no statistically significant effect was observed in either the parallel or crossover group. The corresponding forest plot is presented in Supplementary File S6.

In the acute-chronic model (acute vs chronic), a statistically significant effect on muscle strength was observed in the chronic protein group [μ(SMD): 0.5, 95% CI: 0.05 to 0.96; HDI: 0.05 to 0.95; BF: 1.36], with low within-study heterogeneity and high between-study heterogeneity [τ_within: 0.06, within I²: 0.48%, τ_between: 0.83, between I²: 91.55%]. For endurance performance, a statistical significance was found in both the chronic protein group [μ(SMD): 0.26, 95% CI: 0.03 to 0.49; HDI: 0.03 to 0.49; BF: 0.66] and the acute protein group [μ(SMD): 0.18, 95% CI: 0.001 to 0.36; HDI: 0.004 to 0.37; BF: 0.33], with low within-study heterogeneity and moderate between-study heterogeneity [τ_within: 0.08, within I²: 2.14%, τ_between: 0.45, between I²: 67.75%]. While both posterior distributions excluded zero, indicating credible positive effects, the Bayes Factors (chronic: BF = 0.66; acute: BF = 0.33) suggested only anecdotal to weak support for the alternative hypothesis, with stronger uncertainty observed in the acute group. The corresponding forest plot is presented in Supplementary File S6.

In the endurance model, a statistically significant effect was observed in aerobic [μ(SMD): 0.23, 95% CI: 0.04 to 0.42; HDI: 0.03 to 0.42; BF: 0.74] and anaerobic performance [μ(SMD): 0.47, 95% CI: 0.05 to 0.88; HDI: 0.05 to 0.88; BF: 1.21] when assessed through cycling-based tests, with low within-study heterogeneity and moderate between-study heterogeneity [τ_within: 0.09, within I²: 2.69%, τ_between: 0.45, between I²: 67.37%]. The corresponding forest plot is presented in Supplementary File S6.

In the muscle model, a statistically significant effect was observed only in lower body strength [μ(SMD): 0.52, 95% CI: 0.06 to 0.97; HDI: 0.05 to 0.97; BF: 1.47], with low within-study heterogeneity and high between-study heterogeneity [τ_within: 0.07, within I²: 0.67%, τ_between: 0.81, between I²: 89.75%]. The corresponding forest plot is presented in Supplementary File S6.

In the glycogen model, no statistically significant effect was observed in either glucose or muscle glycogen indicators. The corresponding forest plot is presented in Supplementary File S6.

In the fatigue model, a statistically significant effect was found only in fatigue index [μ(SMD): 1.06, 95% CI: 0.25 to 1.90; HDI: 0.24 to 1.89; BF: 5.63], with low within-study and between-study heterogeneity [τ_within: 0.14, within I²: 6.55%, τ_between: 0.36, between I²: 43.32%]. The corresponding forest plot is presented in Supplementary File S6.

In the interaction model one (Endurance × Energy), a statistically significant effect was only observed in the energy-unmatched condition for anaerobic endurance performance [μ(SMD): 1.38, 95% CI: 0.19 to 2.46], with low within-study and between-study heterogeneity [τ_within: 0.1, within I²: 5.18%, τ_between: 0.27, between I²: 37.79%].

In the interaction model two (Muscle Strength × Protein Timing), no statistically significant effects were observed in any timing condition across all muscle strength indicators.

In the interaction model three (Muscle strength × Supplements), a statistically significant effect was only observed in jump performance (PRO vs PLA) [μ(SMD): 2.21, 95% CI: 0.97 to 3.41], with low within-study and moderate between-study heterogeneity [τ_within: 0.07, within I²: 2.97%, τ_between: 0.3, between I²: 54.58%].

Compared to the previous models, all interaction models showed a substantial reduction in heterogeneity. The heterogeneity in the original endurance performance model primarily stemmed from differences in energy matching between groups, while the heterogeneity in the muscle strength model was mainly attributed to variations in protein timing and types of protein supplements.

Four trait moderators were added to the meta-regression. The regression plot is presented in Figure 5 for athletic performance and in Figure 6 for post-exercise recovery, while the detailed results are provided in Supplementary File S7.

In terms of athletic performance, age was the only moderator that showed a significant negative effect, indicating that the effectiveness of protein on athletic performance decreases with increasing age [coefficient estimate: –0.02, 95% CI: –0.04 to –0.001; R² = 2%]. Other moderators were not found to have statistical significance in athletic performance. However, from the graph, the influence of weight and female proportion showed a negative trend.

In post-exercise recovery, no moderators were found to have a significant effect.

Based on the trait moderation analysis, the models including the four trait moderators all showed low R² values, indicating that they cannot fully explain the heterogeneity. Therefore, it is unlikely that the heterogeneity originates from these four moderators (age, weight, sample size, and gender).

Three dose moderators, duration of protein supplementation (in days), extra protein supplement dose and total daily protein intake, were added to the moderation analysis. Regarding the duration of protein supplement interventions, the results suggest that a period of 40 to 65 days is optimal for enhancing athletic performance [40 days: coefficient estimate: 0.45, 95% CI: 0.07 to 0.88; R²: 4%; 65 days: coefficient estimate: 0.60, 95% CI: 0.23 to 0.99; R²: 4%], while a duration of 40 to 80 days appears suitable for promoting post-exercise recovery [40 days: coefficient estimate: 0.33, 95% CI: 0.08 to 0.62; R²: 10%; 80 days: coefficient estimate: 0.57, 95% CI: 0.10 to 1.05; R²: 10%]. The regression plots are presented in Figure 7.

Regarding the extra protein dose from supplements, the results show that a daily intake of 1 g/kg of additional protein from supplements [coefficient estimates: 0.27, 95% CI: 0.05 to 0.50; R²: 4%] is more effective for improving athletic performance than 1.5 g/kg [coefficient estimates: 0.20, 95% CI: –0.16 to 0.56; R²: 4%] and 2 g/kg [coefficient estimates: 0.13, 95% CI: –0.33 to 0.57; R²: 4%]. For promoting post-exercise recovery, an extra dose of 0.5 g/kg [coefficient estimates: 0.16, 95% CI: 0.06 to 0.26; R²: 7%] shows better outcomes compared to 1 g/kg [coefficient estimates: 0.09, 95% CI: –0.1 to 0.26; R²: 7%] and 2 g/kg [coefficient estimates: 0.17, 95% CI: –0.11 to 0.49; R²: 7%]. The regression plots are presented in Figure 8.

Regarding total daily protein intake, the results indicate that consuming 2 g/kg per day yields better improvements in athletic performance [coefficient estimate: 0.33, 95% CI: 0.05 to 0.63; R²: 63%] compared to 1 g/kg [coefficient estimate: 0.05, 95% CI: –0.52 to 0.56; R²: 63%] and 1.5 g/kg [coefficient estimate: 0.21, 95% CI: –0.05 to 0.50; R²: 63%]. For promoting post-exercise recovery, an intake of 2 g/kg also shows more favourable effects [coefficient estimate: 0.30, 95% CI: 0.001 to 0.63; R²: 21%] than 1 g/kg [coefficient estimate: –0.02, 95% CI: –0.50 to 0.44; R²: 21%] and 1.5 g/kg [coefficient estimate: 0.21, 95% CI: –0.06 to 0.49; R²: 21%]. The regression plots are presented in Figure 9.

The R² density plots are presented in Supplementary File S8. The results indicate that the overall protein dose model is the optimal model for both athletic performance and post-exercise recovery (R² for performance: 63%; R² for recovery: 21%).

The quality grade for each outcome was determined based on sample size, meta-analysis results, and quality assessment, including risk of bias, result inconsistency, indirectness, and imprecision of effect estimates. The results showed that the quality of evidence for muscle strength, glycogen resynthesis, and fatigue was rated as very low due to high heterogeneity or inconsistency, non-significant findings, and small sample sizes. The quality of evidence for endurance performance was rated as low because of high heterogeneity and inconsistency. The GRADE summary is presented in Figure 10.

Additionally, we conducted GRADE assessments for all models. In the interaction models, we found that heterogeneity was generally reduced, leading to upgraded quality ratings (high quality for endurance and moderate quality for muscle strength). The previously observed high heterogeneity in the null model for endurance performance was mainly attributed to differences among endurance indicators and variations in macronutrient distribution. For muscle strength, the high heterogeneity in the null model stemmed from differences in strength measurement indicators, timing of protein intake, and types of supplementation used. The additional GRADE summary is provided in Supplementary File S9.

The funnel plots generated by the publicationbias and metafor packages are illustrated in Supplementary File S10. First, a multilevel Egger's test indicated the presence of publication bias in the athletic performance data, including both endurance and muscle strength outcomes (P < 0.0001), whereas no such bias was detected in the post-exercise recovery data (P = 0.27). Second, significance funnel plots generated using the PublicationBias package revealed that most studies were non-significant across all outcome types. However, the grey diamond (representing non-significant studies only) was closely aligned with the black diamond (representing all studies), suggesting that the overall findings were not substantially affected by publication bias. Third, the computed s-values indicated that, for both endurance and muscle strength, no plausible level of publication bias would be sufficient to attenuate the observed effects to null (s-value = not possible), supporting the robustness of the results.

Additionally, a trim-and-fill analysis using the metafor package further confirmed the robustness of the athletic performance data, with no imputed studies and no clear asymmetry observed in the funnel plot.

The results indicate that protein supplements offer significant benefits for both athletic performance and post-exercise recovery. Protein-carbohydrate supplements were found to improve endurance when compared to placebo supplements, though they did not improve muscle strength. In contrast, pure protein supplements enhanced both muscle strength and endurance relative to the placebo group, with no significant differences observed compared with carbohydrate supplements. Both protein-carbohydrate and pure protein supplements promoted more efficient glycogen resynthesis than placebo, and pure protein supplements additionally reduced fatigue. However, all significant findings were derived from studies in which the energy ratios between the experimental and control groups were unequal.

The effect of protein supplements on athletic performance is weak and only marginally significant, and even when statistical significance is detected, it originates from studies with unbalanced energy intake between the intervention and control groups [10,11,27,37,61–63,65–67,69,89–94]. In these energy-unmatched studies, protein supplementation was found to improve anaerobic and aerobic endurance in cycling tests, as well as lower-body strength in athletes. Whey protein remains the most effective type for improving athletic performance; however, the absence of significant effects for other protein types may result from the limited number of studies examining them.

Firstly, regarding endurance performance, many studies support the conclusion that in isocaloric trials, endurance performance does not improve with pure protein or protein-carbohydrate supplementation compared with control groups [10,69,95,96]. Jäger et al. concluded that adding protein to a carbohydrate beverage before or during endurance exercise does not generally improve performance, especially in isocaloric conditions [8]. Another review further noted that such performance improvements are likely attributable to the additional energy provided by protein co-ingestion, rather than protein itself [97]. Moreover, protein supplementation may offer greater benefits for individuals with low aerobic capacity, but not for those capable of substantial VO₂max improvements through endurance or high-intensity interval training [94]. Additionally, when daily nitrogen balance is already positive, the benefits of protein supplementation on subsequent endurance performance appear to be negligible [98]. Therefore, current evidence suggests that protein supplementation is not essential for enhancing endurance performance, particularly in well-trained athletes or when adequate carbohydrate intake is ensured.

Second, the effect of protein supplementation on muscle strength is similarly limited and appears to be even less effective than its impact on endurance performance. Long-term protein interventions tend to show better results than acute (single-dose) supplementation. Positive effects are primarily observed in improvements in lower-body strength, with many studies providing supportive evidence for this finding [54,70,71,99,100]. Jäger et al. summarised that protein supplementation exerts a small to modest impact on strength development in both men and women. While findings from individual studies are mixed, pooled analyses support a modest benefit, particularly when combined with sufficient training duration and total protein intake [8]. Recent evidence indicates that protein supplementation during resistance training results in small but significant gains in lean body mass and lower-body strength, especially in younger or trained individuals, while having minimal effects on fat mass, handgrip strength, and muscle fibre hypertrophy [32,101]. Additionally, a systematic review found that acute protein supplementation enhances myofibrillar protein synthesis following concurrent exercise, while longer-term supplementation shows inconsistent but sometimes positive effects on muscle mass and strength, with no benefits for aerobic capacity [102]. Therefore, protein supplementation should be viewed as a complementary strategy that supports muscular adaptations, particularly when combined with structured resistance training, but its benefits are limited and context-dependent.

Third, whey protein is superior to other types of protein sources and provides the majority of the benefits. However, this meta-analysis included a large number of studies that investigated whey protein (35 included studies), suggesting that no significant effects were found on other protein types, which may be due to insufficient studies. Although no significant effect was found, soy protein (6 included studies) appears to be the most promising alternative to whey protein for athletes. In the future, more robust evidence will be needed to prove this. Whey protein appears to play a role in enhancing lymphatic and immune system responses, making it an ideal protein source for athletes [8,103]. The latest review [104] shows that soy protein appears to be an effective alternative to whey protein in promoting optimal muscle mass and strength gains, but the data are limited, and its amino acid content is lower than that of whey protein. Therefore, its true effectiveness requires further evaluation in future studies.

In the moderation analysis, results indicated that older athletes experienced less benefit from protein supplementation. Evidence suggests that an additional intake of approximately 1 g/kg/day from supplements, resulting in a total daily protein intake of around 2 g/kg/day, is most effective for enhancing athletic performance. Furthermore, intervention durations between 40 and 65 days appear to yield the most favourable outcomes. These findings are consistent with previous research demonstrating an age-related decline in the efficacy of protein supplementation, as well as the importance of optimising dosage and duration. Specifically, Jäger et al. concluded that protein supplementation of 15 to 25 g per day over a period of 4 to 21 weeks positively influences performance [8]. It has also been reported that increasing age reduces the effectiveness of protein supplementation during resistance training, whereas greater training experience enhances it [105].

Regarding optimal intake levels, the recommended daily protein intake for maximising strength gains is above 1.6 g/kg/day [101,105], while for endurance athletes, approximately 1.85 g/kg/day appears to be most effective [106]. Additionally, for athletes aiming to maintain an anabolic environment, particularly during periods of energy restriction, higher protein intakes may be necessary. A daily intake of 1.4 to 2.0 g/kg/day is generally considered the minimum requirement, with greater amounts potentially needed to preserve fat-free mass [8]. These insights highlight the necessity of individualised protein strategies that consider factors such as age, training experience, exercise type, and nutritional status to optimise athletic performance.

In the moderation analysis regarding gender, although the results were not statistically significant, the regression plot suggested a negative trend. Whether protein intake is influenced by gender remains unclear. Most current studies have been conducted primarily on male participants [107,108], highlighting the need for more research focusing on female subjects to explore potential differences in the effects of protein supplementation between males and females. Moore et al. concluded that it is too early to recommend sex-specific carbohydrate or protein guidelines for female athletes if their energy needs are met. More research is needed using sport-specific protocols that control for factors such as prior exercise, nutritional status, contraceptive use, and menstrual cycle phase [109]. Therefore, future studies should prioritise gender-specific investigations to determine whether physiological and hormonal differences influence protein metabolism and supplementation outcomes. Such research is essential for developing tailored nutritional strategies that optimise performance and recovery in both male and female athletes.

Due to length constraints, examining the relationship between protein supplementation and all aspects of post-exercise recovery was challenging. Therefore, this meta-analysis focused on two key components: glycogen resynthesis (i.e. muscle glycogen and blood glucose) and fatigue. Unfortunately, only a small number of results showed statistical significance, and the observed significance was primarily attributable to studies in which the experimental and control groups had unequal energy intake, rather than the independent effect of protein supplementation.

Several studies have provided supporting evidence. Jäger et al. summarised that when combined with sub-optimal intake of carbohydrates (<1.2 g/kg/day), protein-carbohydrate supplements probably can heighten muscle glycogen recovery and may help mitigate changes in muscle damage markers [8]. Additionally, an RCT involving eight cyclists found protein-carbohydrate supplements could increase muscle glycogen by 128%, but these benefits may be due to the energy intake from extra carbohydrate ingestion, not the protein itself [110].

In terms of post-exercise fatigue recovery, only a statistically significant effect was observed in the fatigue index measured by the Wingate test. Some studies provided supporting evidence. Mhamed et al. concluded that the fatigue index reflects the degree of anaerobic fatigue by measuring the decline in power output during high-intensity exercise. Its improvement is likely due to increased muscle protein synthesis, reduced muscle inflammation, and timely protein intake post-exercise, which together enhance muscle recovery and endurance capacity [111]. A systematic review by Pasiakos et al. investigated the relationship between protein and muscle soreness [112]. In this review, some included studies showed a decrease in muscle soreness in groups consuming protein after initial exercise bout [10,113,114], whereas others did not [91,115,116]. Therefore, when protein supplements are provided, acute changes in post-exercise protein synthesis and anabolic intracellular signalling have not resulted in measurable reductions in muscle damage and enhanced recovery of muscle function [112]. The protein-rich supplementation regime seems to attenuate exercise-induced muscular and inflammatory stress responses [94]. However, apart from the fatigue index, no significant effects were observed in blood lactate levels or subjective fatigue assessments, indicating that the impact of protein supplementation on fatigue recovery requires further investigation. Additionally, some relevant biomarkers, such as creatine kinase (CK), were not included in this analysis, thus no definitive conclusion can be drawn regarding the effect of protein supplementation on fatigue recovery.

Several limitations should be acknowledged. First, although interaction models reduced heterogeneity, the limited sample size (~1,000 athletes) and persistent high heterogeneity in the main models led to multiple downgrades, thereby limiting the overall quality and credibility. More studies that include a large sample size are needed to provide robust evidence. Second, a potential risk of bias in the athletic performance outcomes reduces the credibility of the results. Third, this meta-analysis cannot provide comprehensive conclusions across all aspects of protein supplementation, particularly regarding athletic performance and post-exercise recovery. For instance, further research is required to clarify the efficacy of protein timing and sources. The limited number of studies has contributed to the inconsistency of the results. Moreover, key recovery-related biomarkers, such as CK which reflects muscle damage, were not extracted in this analysis; this omission may partly explain the negative findings. Finally, we found that the significant improvement in endurance performance associated with protein supplementation was mainly observed in studies where participants were not fasted before the intervention. This suggests that the observed effects may result from additional dietary intake rather than protein itself. Future studies should adopt more rigorous experimental designs to eliminate confounding factors and clarify the true effect of protein supplementation.

All data and supplementary files used in this study, including graphs, codes in R, and results, have been uploaded to the OSF database for sharing. (https://osf.io/y25va/?view_only=b8b9aa4fef4443dd88647974b6b3d7b8↗).