^{1}

^{2}

^{3}

^{4}

The authors have declared that no competing interests exist.

Conceived and designed the experiments: RJK. Performed the experiments: RJK CEC UK. Analyzed the data: RJK. Wrote the paper: RJK CEC CO UK.

Currently ketogenic diets (KDs) are hyped as an anti-tumor intervention aimed at exploiting the metabolic abnormalities of cancer cells. However, while data in humans is sparse, translation of murine tumor models to the clinic is further hampered by small sample sizes, heterogeneous settings and mixed results concerning tumor growth retardation. The aim was therefore to synthesize the evidence for a growth inhibiting effect of KDs when used as a monotherapy in mice.

We conducted a Bayesian random effects meta-analysis on all studies assessing the survival (defined as the time to reach a pre-defined endpoint such as tumor volume) of mice on an unrestricted KD compared to a high carbohydrate standard diet (SD). For 12 studies meeting the inclusion criteria either a mean survival time ratio (MR) or hazard ratio (HR) between the KD and SD groups could be obtained. The posterior estimates for the MR and HR averaged over four priors on the between-study heterogeneity τ^{2} were MR = 0.85 (95% highest posterior density interval (HPDI) = [0.73, 0.97]) and HR = 0.55 (95% HPDI = [0.26, 0.87]), indicating a significant overall benefit of the KD in terms of prolonged mean survival times and reduced hazard rate. All studies that used a brain tumor model also chose a late starting point for the KD (at least one day after tumor initiation) which accounted for 26% of the heterogeneity. In this subgroup the KD was less effective (MR = 0.89, 95% HPDI = [0.76, 1.04]).

There was an overall tumor growth delaying effect of unrestricted KDs in mice. Future experiments should aim at differentiating the effects of KD timing versus tumor location, since external evidence is currently consistent with an influence of both of these factors.

While the first study assessing the effect of diet on cancer dates back to 1909 [

KDs have shown anti-tumor potential in many, but not all mouse studies. The reason for this discrepancy is not yet clear. In some studies, calorie restriction was required to elicit a potent anti-tumor effect [

In addition, a collective interpretation of murine tumor models is hampered by the large variety of experiment setups and the small number of animals used in most studies. Thus there remains some uncertainty concerning the anti-tumor effects of a KD in current preclinical models. We therefore conducted a systematic review of the literature to ascertain the effects of a KD on tumor growth, and to determine possible factors that may account for heterogeneity in response to the KD.

The inclusion criteria for this meta-analysis were defined a priori as follows:

Studies investigating tumor growth in a murine cancer model.

Studies testing the effects of an

Endpoint defined as reaching a pre-defined tumor volume or other sign of disease progression with no termination of the experiment at a pre-defined time interval.

Conduction of a survival analysis with the specified endpoint, so that in principle either a hazard ratio (HR) or a mean survival time ratio (MR) between the KD and the control diet groups could be calculated.

Studies not fulfilling all of the above four inclusion criteria were excluded from the analysis. No registered protocol existed for this study.

Potentially relevant studies were searched January 5, 2016 in the PubMed database using the search terms “ketogenic diet” AND “cancer”. References of selected articles and review articles on this subject were searched for additional studies.

Data from each study and risk of bias were extracted independently by two authors (RJK and CEC) using a preset form. In case of discrepancies between extracted data, consent was found by discussion between these two reviewers. For each study, we recorded the year it was published, the first author’s name, the tumor model used, the number of animals in each diet group, the time when the diet intervention was started (prior to/at the same day of/after tumor implantation), the ketogenic ratio of the chow and whether body weight under the KD increased, decreased, or remained unchanged compared to the control regimen (

Publication year | Study | Tumor model | Model details | Location | N_{KD}+N_{SD} |
Diet initiation | Ketogenic ratio | Ketosis | Glycemia | Body weight | Comment |
---|---|---|---|---|---|---|---|---|---|---|---|

2007 | Zhou | S | CT-2A brain tumor i.c., C57BL/6J mice | Brain | 9+7 | after | 4:1 | + | 0 | 0 | This study had two separate experiments. High risk of reporting bias (no HR/MR given). |

X | U87 glioma s.c., C57BL/6J mice | Brain | 7+11 | after | 4:1 | + | 0 | 0 | |||

2008 | Freedland | X | LNCaP prostate s.c. | s.c. | 25+25 | prior | 2.1:1 | + | + | 0 | SD defined as the Western diet. KD mice heavier than controls at tumor implantation, but this was accounted for in HR computation. |

2008 | Otto | X | 23132/87 gastric cancer s.c., NMRI mice | s.c. | 12+12 | day 0 | 2.7:1 | + | 0 | 0 | High risk of selection, performance and other bias (KD mice lighter than controls at tumor implantation; individual who performed the experiments also analyzed the data; conflicts of interest). |

2009 | Mavropoulos | X | LAPC-4 prostate s.c., SCID mice | s.c. | 48+41 | prior | 2.1:1 | + | 0 | 0 | SD defined as the Western diet. High risk of selection bias (KD mice heavier than controls at tumor implantation). |

2010 | Stafford | S | GL261 glioma i.c., C57BL/6 mice | Brain | 5+5 | after | 6:1 | + | NA | NA | High risk of reporting bias (no body weight trends reported) |

2011 | Maurer | X | LNT-229 glioma i.c., athymic Foxn1nu mice | Brain | 12+12 | after | 2.7:1 | + | 0 | 0 | High risk of reporting bias (no HR/MR given). Four mice in the SD group and two in the KD group were censored and not considered for mean survival time computation |

2012 | Abdelwahab | S | GL261 glioma i.c., C57BL/6 mice | Brain | 20+19 | after | 4:1 | + | - | 0 | One mouse in the KD group was cured and not considered for mean survival time computation. High risk of performance bias (individual who performed the experiments also analyzed the data; conflicts of interest). |

2013 | Poff | S | VM-M3 metastatic cancer, s.c., VM/Dk mice | s.c. | 8+13 | day 0 | 4:1 | 0 | - | - | Ketone bodies on KD elevated, but not significantly. High risk of performance bias (individual who performed the experiments also analyzed the data). |

2014 | Rieger | X | U87MG glioma cells i.c., athymic Foxn1nu mice | Brain | 8+8 | after | 3.1:1 | + | 0 | 0 | High risk of reporting and other bias (no HR/MR given; conflicts of interest). |

2015 | Hao | X | HCT116 colorectal s.c., BALBc/J SCID male | s.c. | 24+12 | day 0 | 3:1 | + | 0 | 0 | Two KDs used (MKD and LKD); both groups pooled together. |

2015 | Dang | S | Spontaneous murine medulloblastoma, genetically engineered Ptch1+/- Trp53-/- mice on C57Bl/6:129SV 0background | Brain | 4+4 | after | 4:1 | + | NA | + | High risk of reporting, performance and other forms of bias (no HR/MR given; individual who conducted the experiment also analyzed the data; no ketone body measurements reported). |

2015 | Martuscello | X | Patient-derived L0 glioblastoma cells i.c., NOD/SCID mice | Brain | 10+11 | after | 6:1 | + | - | - | Two ketogenic diets used (KD and sHFLC) but only KD considered due to its high ketogenic ratio. High risk of selection, reporting and other forms of bias (time from tumor implantation until KD initiation differed by up to 4 days; no HR/MR given). |

Diet initiation refers to “day 0” which is the day of tumor implantation. S: syngeneic; X: xenogeneic. Ketosis and glycemia are coded such that 0 indicates that no statistically significant differences between both groups were found at any measurement (p>0.05), while the + and - signs indicate that there was at least one measurement in which ketosis or blood glucose levels in the treatment group were significantly higher (+) or lower (-), respectively, compared to the control mice.

If no mean survival times or uncertainty estimates were provided in the article, the corresponding study author was contacted by one of us (RJK) to obtain this information. One study [

Risk of bias was assessed by using the Systematic Review Centre for Laboratory animal Experimentation (SYRCLE) tool which consists of 10 items for which judged based on a number of signaling questions [

Finally, one of us (UK) extracted approximate blood concentrations of ketone bodies and glucose from figures and data, which was possible for 10 studies. As crude estimates, these were treated with care and only used to get an idea of the range of ketosis and blood glucose levels in the mice.

We conducted a Bayesian meta-analysis. Compared to the classical approach this has several advantages such as obtaining direct probability distributions for the parameters of interest, naturally accounting for the full uncertainty in the parameters and allowing each individual study “borrowing strength”, i.e., utilizing information from all other studies for estimating the “true” study treatment effect [

We anticipated different, yet similar, effects of the KD intervention between the studies, so that a random effects model was used [

Here _{i} and _{i} denote the outcome [ln(MR) or ln(HR)] and its SE in the _{i} are assumed to be exchangeable [

Heterogeneity was assessed by the between study variance τ^{2} which was modeled using four different prior distributions [^{2}, which is close to being uniform on log(

Finally, Bayesian meta-regression [

Here, _{i} is the covariate (also called moderator) for study

All analysis was conducted with R version 3.1.3 with the BRugs package and OpenBugs version 3.2.2. Two Markov chains were individually initialized and the first 10000 Markov chain Monte Carlo samples discarded. For the next 25000 iterations every fifth sample was kept to obtain the posterior parameter distribution for the parameters of interest. The median was taken as the parameter estimate and parameters considered “significant” when their 95% highest posterior density interval (HPDI) excluded zero.

The PubMed search for “ketogenic diet” AND “cancer” resulted in a total of 72 articles of which 23 were studies investigating the effects of a KD on tumor growth in a mouse model (

We were able to obtain sufficient data to compute either a MR or a HR from 12 of the 13 selected studies. The study supplying insufficient information was excluded (

The general design and results of the included 12 studies is given in

Publication year | Study | T_{KD} [days] |
T_{SD} [days] |
MR | MR 95% CI | HR | HR 95% CI | Data source |
---|---|---|---|---|---|---|---|---|

2007 | Zhou | 19.7±0.9 | 16.7±1.4 | 0.85 | [0.69,1.01] | NA | NA | Mean survival times provided by author |

18.7±0.9 | 22.5±1.8 | 1.20 | [0.98,1.42] | NA | NA | |||

2008 | Freedland | NA | NA | NA | NA | 0.48 | [0.27,0.86] | Publication |

2008 | Otto | 34.2±2.5 | 23.3±1.1 | 0.68 | [0.57,0.80] | 0.16 | [0.05,0.53] | Individual survival times provided by author |

2009 | Mavropoulos | NA | NA | NA | NA | 0.59 | [0.37,0.93] | Publication |

2010 | Stafford | 24±1.1 | 19±0.7 | 0.79 | [0.70,0.89] | 0.07 | [0.01,0.63] | Individual survival times provided by author |

2011 | Maurer | 82.4±1.2 | 94.9±1.3 | 1.15 | [0.89,1.49] | 1.65 | [0.65,4.21] | Individual survival times provided by author |

2012 | Abdelwahab | 28.8±1.5 | 23.3±1.1 | 0.81 | [0.70,0.92] | 0.35 | [0.17,0.71] | Individual survival times provided by author |

2013 | Poff | 48.9±4.4 | 31.2±4.4 | 0.64 | [0.43,0.85] | NA | NA | T_{KD} and T_{SD} taken from publication, standard errors computed from p-value (see text for details) |

2014 | Rieger | 35.6±0.7 | 33.9±1.6 | 0.95 | [0.85,1.05] | 0.79 | [0.28,2.24] | Individual survival times provided by author |

2015 | Hao | 34.5±10.1 | 24.8±3.1 | 0.72 | [0.27,1.17] | NA | NA | Publication |

2015 | Dang | 17.8±0.5 | 16.3±2.3 | 0.92 | [0.66,1.17] | 1.43 | [0.82,6.30] | Publication; individual survival times read off Kaplan-Meier plot |

2015 | Martuscello | 56±4.2 | 38±1.0 | 0.68 | [0.57,0.78] | NA | NA | Mean survival times provided by author |

T_{KD} and T_{SD} denote the mean survival times in the KD and SD groups, respective, and are given with their SE. These SE have been used to compute the 95% CI.

Not reporting MR and HR despite conducting a survival analysis was considered as evidence for reporting bias. By retrieving these measures from the study authors we eliminated the influence of this bias on the cumulative evidence. However, several other forms of bias were identified in all but one study (

When all studies were pooled together, a total of 192 mice were treated with a KD and 180 mice fed a SD. Mice receiving a KD had higher ketone body concentrations which was significant in all studies but one [

Eight of the 13 experiments found a significantly longer survival for mice receiving a KD compared to a SD. The result of the meta-analysis for the overall effect of a KD on the MR and HR is shown in Tables ^{2} were MR = 0.85 (95% HPDI = [0.73, 0.97]) and HR = 0.55 (95% HPDI = [0.26, 0.87]). Thus there was a significant overall benefit of the KD in terms of prolonged mean survival times and reduced odds of dying first. The effect measure estimates were not sensitive to the type of prior used for the between-study variance. The estimate of τ^{2}, however, was highly sensitive to its prior in the meta-analysis when HR was used as the outcome. This probably reflects the greater uncertainty associated with the small number of studies. With MR as the effect measure, estimates of τ^{2} were more uniform and reasonable, but the 95% HPDI supported both very small and substantial heterogeneity [

Prior on |
Uniform prior | Half-normal prior | Inverse gamma prior | DuMouchel prior |
---|---|---|---|---|

Prior distribution | ||||

Posterior median | 0.85 | 0.85 | 0.85 | 0.85 |

Standard deviation | 0.06 | 0.06 | 0.06 | 0.06 |

95% HPDI | [0.72,0.98] | [0.74,0.97] | [0.73,0.97] | [0.74,0.96] |

^{2} |
||||

Prior distribution | ||||

Posterior median | 0.0388 | 0.0334 | 0.0315 | 0.0290 |

Standard deviation | 0.0367 | 0.0266 | 0.0300 | 0.0260 |

95% HPDI | [0.0106,0.1446] | [0.0099,0.1078] | [0.0086,0.1165] | [0.0082,0.1022] |

Prior on |
Uniform prior | Half-normal prior | Inverse gamma prior | DuMouchel prior |
---|---|---|---|---|

Prior distribution | ||||

Posterior median | 0.54 | 0.55 | 0.55 | 0.55 |

Standard deviation | 0.20 | 0.11 | 0.17 | 0.16 |

95% HPDI | [0.25,1.04] | [0.38,0.80] | [0.27,0.87] | [0.30,0.92] |

^{2} |
||||

Prior distribution | ||||

Posterior median | 0.49 | 0.0649 | 0.1914 | 0.1886 |

Standard deviation | 0.7251 | 0.1111 | 0.658 | 0.5987 |

95% HPDI | [0.0198,2.842] | [0.0002,0.4019] | [0.0012,2.076] | [0.0004,1.822] |

The overall protective effect was still apparent after excluding three studies with high risk of bias due to financial conflicts of interest [

Figs ^{2}. Note how the Bayesian estimates of the true study effects of each trial are shifted towards the overall pooled effect and have decreased uncertainty by “borrowing strength” from all the other trials.

Values less than 1 indicated a beneficial effect of the KD. The observed effects

Values less than 1 indicated a beneficial effect of the KD.

There was a one-to-one correlation between time of diet initiation and tumor location, as all experiments with intracranial tumors started the diet a few days after tumor manifestation and vice versa. Accordingly, in meta-regression both the tumor location and the time at which the KD was initiated were able to account for 26% of the heterogeneity between studies measuring a MR. Brain tumors and the switch to the KD later than the day of tumor initiation were associated with less effectiveness of the KD with more than 90% probability (_{late diet initiation} = _{brain tumor} = 0.29, 90% HPDI = [0.003, 0.54]). In this subgroup of studies the MR estimate was 0.89 (95% HPDI = [0.76, 1.04]), still supporting a positive effect, albeit no longer significantly.

The ketogenic ratio was able to explain only 1% of the heterogeneity between experiments measuring a MR. An increase in the ketogenic ratio by 1 was thereby associated with a slight decrease in the MR of -0.055 (95% HPDI = [-0.19, 0.08]), although only the 65% HPDI excluded a zero effect (65% HPDI = [-0.11,-5.7×10^{−4}]). For tumor model (syngeneic/xenogeneic) or year of publication, no significant effects on the MR were found. None of the tested covariates were significantly correlated with a moderation of the overall HR, and accordingly no covariate was able to explain part of the heterogeneity between studies assessing a HR.

This meta-analysis indicates that in mice a KD prolongs survival (MR<1) and reduces the risk of experiencing the pre-defined endpoint (HR<1) compared to a high-carbohydrate SD when used as a monotherapy. It is therefore in line with a previous review by Lv et al. [

The protective effect of the KD is most likely related to the state of ketosis, which was the most consistent covariate across studies. In particular, survival seems to be less dependent on weight loss in the KD group since most studies reported similar weight trends in both groups. Amongst several putative effect moderators only the time of KD initiation or alternatively tumor location were found to influence survival times and account for some of the between-study heterogeneity, as all brain tumor models included in the analysis for MR were also the ones using a late switch to the KD and vice versa. With more than 90% probability, the studies supported a survival-prolonging effect when the KD was started early (day of tumor cell injection) compared to at least one day after tumor cell injection or—alternatively—when a subcutaneous tumor instead of an intracranial one was used. Since it is currently not possible to differentiate both effects based on the studies evaluated in this meta-analysis, other evidence could be considered to reach a careful conclusion.

A protective role of the KD against early stages of tumorigenesis, but a much lesser effect when tumor growth has already been initiated, would be consistent with results from the largest rodent study on KD and cancer growth conducted to date. In this study, a total of 303 rats were used to investigate the effects of a carbohydrate-free diet started either before or concurrently with tumor transplantation [

On the other hand, Seyfried and coworkers have argued that unrestricted KDs are not effective against various brain tumor models [

Translated into the clinic, our result would imply at best a weak effect of KDs as the sole therapy against either already manifested tumors in general and/or brain tumors in particular. It is interesting that both hypotheses are consistent with the findings from human studies on glioblastoma multiforme, in which a KD as monotherapy seems ineffective in retarding tumor growth but more promising when combined with standard treatments [

Our meta-analysis has several limitations. First, all animal studies assessing the KD have a small sample size which leads to large uncertainties on the outcome measure. While this meta-analysis can therefore help to reach an overall conclusion with better precision, one must provide caution with definitive conclusions, especially as other non-random biases not accounted for can exist. Secondly, as expected, there was a moderate to large amount of heterogeneity present. Although tumor location and/or the time of KD initiation were able to account for roughly a quarter of this heterogeneity between studies using MR as the effect measure, much of it remains unexplained and probably relates to the large variety of mouse strains, tumor cell lines and endpoint criteria used. However, regardless the source of heterogeneity the results were highly robust against various a priori assumptions about the heterogeneity. Thirdly, the results might be sensitive to the various amounts of bias identified but also unidentified due to underreporting. We have shown that removing three studies with financial conflicts of interest still gave an overall protective effect of the KD, yet this was no longer significant. Thus we judge the uncertainties of the overall result as higher than estimated due to various forms of bias. Fourthly, the relations between individual blood glucose and ketone body levels on survival remain elusive. There is evidence for the importance of minimizing the ratio of glucose-to-ketone body concentrations for brain tumor control, which would indicate that additional calorie restriction could make KDs even more efficient [

Finally we note that several of the identified biases and other aspects of the tumor models provide caution when extrapolating the results to humans. For instance, in some studies, tumor cells are injected subcutaneously in the mouse prior to assessment of growth, as opposed to their native organ location [

In conclusion, we found that the published data thus far indicate that a KD impedes tumor growth in mice. Our analysis reveals that the primary moderators of this effect may be the tumor location (brain/subcutaneous) and time of diet initiation. However, the strong correlation between these two covariates in the studies makes the exact mechanism elusive. Furthermore, all studies suffered from various biases and underreporting of methods whose influence on our result also remains elusive. Future studies should therefore improve methodological reporting and evaluate the effects of early versus late KD initiation for both subcutaneous and intracranial tumors. Also the translationally most relevant setting of a KD initiation concurrent with standard therapies after cancer manifestation should be more frequently investigated. If the timing of the KD is of major importance this would imply a role of fasting and KDs as a prevention strategy in humans, but only a supportive role during cancer treatment which is consistent with the current available human data. Further studies in humans to test these hypotheses are warranted.

We thank Thomas Seyfried, Brent Reynolds, Adrienne Scheck and Gabriele Maurer for sending us relevant study data. We especially appreciate the provision of complete survival time data which for a scientist cannot be taken for granted.