Dairkee, S. H., Moore, D. H., Luciani, M. G., et al. (2023). Reduction of daily-use parabens and phthalates reverses accumulation of cancer-associated phenotypes within disease-free breast tissue of study subjects. Chemosphere, 322, 138014.
Briefing note & Study Breakdown by Lyle Burgoon PhD ATS, President of Raptor Pharm & Tox, Ltd
Briefing Note
Main takeaway: This study should not be a cause for concern regarding cosmetic product safety. Anyone who tells you otherwise very clearly didn’t critically read the study.
The data doesn’t support the conclusions of the paper. Anyone who relies just on the abstract or a news release will be misled. There is no data here to support the idea that any of these subjects had a “cancer phenotype” or gene expression that was driving these people to develop “breast cancer.” The reality is that using summary statistics can be a misleading way of presenting and basing an argument. We need to look at the full distribution and uncertainty in the data first. In this case, the same genes were up in some people, down in others, and thus all over the place. The authors said it best – interindividual variability is high such that people tended to cluster with themselves in the gene expression. The fold-changes of gene expression before and after using the personal care products were all over the place. Means suck to summarize noisy data as they are susceptible to outliers. Looking at the full distribution and the variance is always necessary.
Most importantly – this is a study of just 41 people total, and only 14 for the gene expression conclusions. We can’t even tell if a coin or dice is fair with only 14 flips or throws. How are we supposed to tell what is happening with people, who have far greater variance, with just 14 people?
I’ve asked the authors for their gene expression data so I can do my own analyses. The authors have not responded to my request yet.
Key study issues
- This study is a classic example of selection bias – I would technically call it sampling bias. What does that mean? It means that there are too few people in the study to actually draw any kind of conclusions. There’s a reason why clinical trials need thousands of people. And this study has 41 subjects at the start – but they don’t get data from all 41 (see Table 2 where they get measurements for at most 36 (what happened to the other 5?), and sometimes only 22 (what happened to the other 19?)). For the gene expression study, they only looked at 12 people. We don’t know why data is missing for various women. All we read is that they were only able to obtain paired fine needle aspiration samples for 36 women.
- Then the study takes a turn for the weird. When they are looking at the gene expression data, they randomly choose 2 of the 14 subjects to be “controls” by using a random number generator. Here’s what’s so odd about this – what are these 2 women being used to control for? Usually we use controls as an untreated group to compare our treatment to. This is part of the basis of the counterfactual required for causal analysis to occur in experiments. The control is the counterfactual compared to the treatment – meaning, the control is identical to the treatment in all ways except for the treatment application. But that’s not what’s happening here. These 2 control women are randomly chosen from the group of 14 to serve as “controls”, but we don’t know what they’re being used to control for.
- What we do see is that the fold-changes in gene expression (Figure 2B) of each individual follows no pattern whatsowever for the “differentially expressed genes”. This is a classic example of noise masquerading as signal. What this means is that in some people, these genes go up, while for other people, the same genes go down. I find it curious that the random number generator chose the most extreme individual (sn128) by chance – that is simply remarkable! Also remarkable is that the second most extreme person who has the most similar profile to the individual with the most extreme profile was also randomly chosen by chance to be a control (sn126). What is even more interesting is that sn126, the one that is most similar to sn128, has a very different gene expression profile to sn128, despite being the closest.
- What I find even more interesting is that the authors state that the upper region of Figure 2B is the “upregulated” region and the lower region is “downregulated” despite the fact that 9/14 of the individuals actually show a mostly downregulated pattern in the “upper” region, and a mostly upregulated pattern in the “lower” region of the figure. It would appear that the authors labeled Figure 2B based solely on individual sn128.
- Also, this is all gene expression – it’s not protein. So none of these changes are actually biologically meaningful in and of themselves.
However, it is odd that the authors conclude “[b]ased on high-dimensional gene expression data of matched FNA pairs of study subjects, we demonstrate a striking reversal of cancer-associated phenotypes, including the PI3K-AKT/mTOR pathway, autophagy, and apoptotic signaling networks within breast cells of REDUXE compliant subjects.” That’s not at all what the data show. Look at IRS1 which they call out in Table 3. It has a fold-change of 0.2 (this is no longer log2 transformed it appears) – that would be about a 5x change – and it appears this is a mean. The problem is, by representing this value solely as a mean, the authors are not being transparent about the fact that the full distribution includes values that are clearly upregulation and values that are clearly downregulation (see Figure 2B). As a result, the authors are being very misleading. This same pattern exists for all of the genes in their table.
What does this mean? Anyone who reads beyond the abstract will see that this paper has numerous flaws that do not support the conclusions.
In response to the slew of “experts” who’ve shared this study online. Any expert that wants to talk about a paper has an obligation to actually read the full text of the paper (even if they have to pay for it because it is behind a paywall), to see if the data actually support what is in the abstract. And once someone does that, in a critical way (which is how all reputable scientists approach papers), they will see that the conclusions in the abstract (and the paper) do not match up with the data that are being presented. In fact, it’s clear that the results are not backed up by the data being presented.