Q&A with Daphna Joel
A Q&A with Daphna Joel on her new article, "Analysis of Human Brain Structure Reveals that the Brain 'Types' Typical of Males Are Also Typical of Females, and Vice Versa”
Brain sex-difference research usually focuses on average differences between men and women. Neuroscientist Daphna Joel challenges this approach by asking, “Wait! What exactly do you mean by ‘difference’?” In her 2015 paper, Joel queried whether individual brains of men and women are consistently “male” or “female” at the level of individual features. She found that they were not, giving rise to the hypothesis that most people are “mosaics” for so-called male and female brain traits. In this new paper, just out in Frontiers in Human Neuroscience, Joel asks whether a typical brain "type” for males is also typical of individual females, and vice versa. In this Q & A with the GenderSci Lab, Dr. Joel explains the methods and findings of this paper.
What hypothesis did you set out to test?
We contrasted two hypotheses – the hypothesis that the typical female brain is different from the typical male brain, and the hypothesis that the brain types typical or females are also typical of males, but differences exist in the frequency of rare brain types.
Importantly, according to the two hypotheses there are group-level sex/gender differences in single brain features. The hypotheses differ only in relation to what happens within an individual brain. If these differences typically ‘add-up’ within single brains, then the typical female brain should be different from the typical male brain. If, in contrast, the differences typically ‘mix-up’ within single brains, then the typical female brain should be similar to the typical male brain, as both would consist of ‘mosaics’ of female-typical and male-typical features.
What data did you use, and why did you choose that dataset?
We used three types of measures derived from the MRI data of 2176 individuals in two samples, the Connectomes+ and the Brain Genomics Superstruct Project. The latter dataset was more homogenous than the first dataset in terms of age, geographical region, and imaging parameters. We used several measures and two datasets to increase the generalizability of our results.
What is a brain ‘type’?
Brain ‘type’ loosely refers to a group of brains whose structure is similar in some mathematical sense (which depends on the algorithm employed and the method used for dimension reduction) and the word type is placed in quotations to emphasize that it most likely has no biological meaning.
You use computational methods - clustering algorithms and anomaly detection - to characterize human brain “types," determine the frequency of these different brain types in men and women, and analyze whether someone’s sex category predicts their brain structure. Can you briefly explain how these methods work? What are the strengths and weaknesses of these methods?
Anomaly detection was used to test whether the ‘types’ of brain typical of females are also typical of males, and vice versa. An anomaly detection algorithm builds a model of “normal” items so that it can detect an “abnormal” item when it appears, without having a priori knowledge of the characteristics of the “abnormal” item or of what distinguishes it from the “normal” items.
In our study, the algorithm was applied to examples of brains from a single sex category (say, females) to create a model of brains of this sex category, and then the model was used to identify for every new brain (i.e., from females who were not included in the training set and from males) whether it belongs to this group of brains (‘normal’) or does not (‘anomalous’).
Unsupervised clustering divides brains into clusters that best describe the variability in the dataset, while supervised clustering attempts to create two clusters that best separate between females and males. We used unsupervised clustering to test whether brain ‘types’ common in humans would be similarly common in females and in males, but large sex/gender differences would be observed in some rare brain ‘types’. The algorithms divided all the brains in a given dataset into several clusters, and then we assessed the proportion of males and females within each cluster.
We used supervised clustering to test whether a classification model created on the Boston sample would accurately classify brains from other, test, samples from Tel-Aviv, Beijing and Cambridge, and whether the classifications of the Boston model would agree with the classification model created on the test sample. In other words, whether a brain classified as ‘male’ by the Boston model would also be classified as ‘male’ by the Tel-Aviv model.
The main problem with all these methods is that they relate to similarity in a mathematical sense (see below). But if you ignore this, the main strength of the study is the use of several datasets, several analytical approaches, several manipulations of the data, and several algorithms. Together these increase the generalizability of our conclusions.
Briefly, what did you find?
Anomaly detection revealed that the chances that a brain from a male would be classified as ‘a normal female brain’ were very similar to the chances that a brain from a female would be classified as ‘a normal female brain’, and the same was true following training on brains from males.
Unsupervised clustering analysis revealed a similar proportion of females and males in the large clusters as well as in some of the small clusters. Large sex/gender differences were found only in some small clusters. In some of these small clusters, the number of brains from one sex category was six times higher than the number of brains from the other sex category. Relatedly, we found that the chances that a female and a male are in the same cluster are very similar to the chances that two females or two males are in the same cluster. This means that knowing one’s sex category provides very little information regarding whether her/his brain ‘type’ is similar to or different from mine.
This pattern of results was obtained with all types of MRI-derived data tested, except when the absolute volume of brain regions was considered. With these data the algorithms better distinguished between brains from females and males. However, further inspection of the results revealed that the algorithms were actually distinguishing between large and small brains. In other words, we (like others before us) found that the main morphological difference between brains from females and from males is in total brain volume. Yet, our study shows that human females and males are highly similar in brain architecture, that is, in the relations between the size of different brain structures, with brain architectures common in one sex also common in the other, and large sex differences existing only in the frequency of some rare brain architectures
Supervised clustering revealed that you could use brain structure to predict whether someone is female or male with accuracy of ~80% – an accuracy similar to what other have found. Accuracy remained high when the model created on the Boston sample was applied to the Cambridge sample, but was lower, though still better than what you would achieve if you were simply guessing (50%), when it was applied to the Tel-Aviv and Beijing samples. But the more interesting finding was that in almost all cases, the classification of the Boston model and of the model created on the test sample were independent – that is, the two models did not agree on which brain belongs to a male and which to a female more than would be expected by chance. In other words, the classification of brains into ‘male’ and ‘female’ did not capture a core difference between human females and males, but was specific to the subpopulation of humans on which the classification model was built.
How does this finding differ from the view that there are large and thoroughgoing differences between male and female brains?
Similar to other studies, we find differences between females and males in characteristics (e.g., volume, cortical thickness) of single brain regions. Where we differ from others is in going beyond differences in single features to considering the brain as a whole. What we find is that although there are group-level differences in some features, they do not add up to create two types of brains, one typical of females and the other typical of males, but rather the brain types typical of females are also typical of males, and vice versa, and large differences exist only in the frequency of some rare brain types.
Similar to others, we show that the existence of group-level sex/gender differences in brain structure is sufficient to predict, with accuracy above chance, one’s sex category on the basis of one’s brain structure. We differ from other studies in testing whether there is a universal “sex prediction model,” and showing there isn’t, and in discovering that even though sex-related variance in brain structure can be used to predict the sex category of the brain’s owner, it is not a major determinant of brain structure.
You emphasize that a mathematical difference is not necessarily a biological difference. Why is this important?
It is the actual brain morphology which determines a brain’s function. Brains which are similar in some mathematical sense may differ widely in their morphological details. For example, two brains may be mathematically similar in that they both have four regions in the female-typical form and three regions in the male-typical form, and we may predict and be right 80% of the time, that they belong to females. Yet, these two “female” brains may be very different morphologically, if the three regions that are in the male-typical form in the first brain are the ones in the female-typical form in the second brain, and the three regions that are in the male-typical form in the second brain are the ones in the female-typical form in the first brain.
How can it be that knowing the structure of a brain can often predict the sex of its owner, but knowing the sex of the owner is not very predictive of the structure of the brain?
An analogy may help – imagine that someone came from out of space and wanted to describe human clothes to their home planet. They may come up with categories like warm clothes versus light ones; large clothes (for adults) versus small clothes (for kids); clothes you put on your legs, on your upper body, or on your feet; or they may create categories depending on the color of clothes, etc, etc. If they are unaware of the social importance of sex category, they may not create a category of men’s clothes versus women’s clothes. However, if you ask them to classify clothes as belonging to a man or to a woman, they may be able to learn to do so, depending on features like colors that are more common in the clothes of one sex over the other (like pink), specific designs that appear mainly in clothes of one sex, or even the direction of the buttons in men versus women shirts. In other words, they can use sex-related variability in human clothes to predict whether a specific custom belongs to a man or to a woman, but this variability would not reflect the most important aspects of the variability of human clothes. For example, although knowing that someone is, say, male, is enough to predict that he most likely isn’t wearing a pink outfit, it gives no information on whether he wears warm or light clothes. Similarly, knowing that someone is a male is not informative about which brain structure he has, since sex is not a strong determinant of variability in human brain structure.
You do find sex differences in the frequency of “rare” brain types. What is a rare brain type, and why do you think you found a sex difference in their frequency?
The unsupervised clustering algorithms created clusters of varying size, from very large (that is, including a large proportion of the brains in the sample) to very small (including a very small proportion of the brains in the sample). The latter are referred to as rare brain types. According to our hypothesis there must be sex/gender differences in the frequency of some of these rare brain types. Without this, we would not be able to account for the observations that there are group-level sex/gender differences in single brain features and the brain types typical of females are also typical of males, and vice versa.
Gender as well as sex might influence brain structure. What might be the implications of this for the sort of results you found here?
In all our analyses we are not interested in the source of the sex/gender differences we observe. What we show is that regardless of the source of these differences, they do not add-up to create two types of brains, one typical of males and the other typical of females.
Scientists' work can sometimes be misrepresented or misunderstood. Have you encountered or do you anticipate encountering any misunderstandings of this study that we haven't discussed already? If so, what would be your response?
The main misunderstandings arise from the fact that brain structure can be used to predict, with accuracy above chance, whether someone from the same geographical region is female or male. This fact leads people to conclude, wrongly, that sex category is a major predictor of brain structure and that there is a universal way to predict sex category on the basis of brain structure.
Some also claim that we deny the existence of sex/gender differences in the brain. We do not. The mosaic hypothesis is framed in terms of group-level differences in single features.
Your work innovates new ways to examine brain-sex variation and to characterize similarity and difference in the brain. How might these methods be used by feminist scientists in other fields who are similarly looking to use existing data sets to re-examine sex difference claims?
Our work shows that the existence of sex/gender differences in features of a system (brain, immune system, cardiovascular system) is not enough to conclude that the system has two forms, one typical of females and the other typical of males. To conclude this, one has to specifically compare the frequency of internal consistency versus mixing of male-typical and female-typical features within individuals. I believe, but this belief has to be put to test, that the reproductive system is the exception, and that, similar to the brain, other systems affected by sex would be characterized by mosaic rather than by two distinct systems.
How to cite this blog
Richardson, Sarah. “A Q&A with Daphna Joel,” GenderSci Blog, October 29, 2018, https://projects.iq.harvard.edu/gendersci/joelqa.
Statement about intellectual labor
The GenderSci Lab is committed to an equity- and justice-based approach to scholarly research, which includes the practice of citations and intellectual attribution. As such, we’ve followed the CLEAR Lab’s (an innovative feminist, anti-colonial marine laboratory) guidance in determining author order. For this blog post, Richardson drafted the questions and gathered feedback on them from GenderSci Lab members. Joel answered the questions by email, and Richardson lightly edited the final draft.