Lucía Nadal

Organization of complex morphological diversity in ‘robust’ australopithecines and implications for intraspecific variability

The collection of mandibular fossils attributed to Paranthropus boisei holds a striking variability that has been explained as being the result of either sexual dimorphism, taphonomic damage, chronological change, or potential taxonomic heterogeneity. This talk presents new research characterizing the three-dimensional shape variability of this hypodigm using unsupervised machine learning algorithms to test for various contributing factors and describe patterns of morphological variability in ‘robust’ australopithecines.

FULL TRANSCRIPT 

So a huge honor to be here today presenting this. I'm particularly very excited about being able to present some of the research that I did for my PhD project with my supervisor Marta Mirazon Lahr. And I thought I would start by giving you a little bit of context about how this project came to be. I had done my master's thesis on Paranthropus teeth so by the time I met Richard Leaky, I had developed a healthy Paranthropus obsession. And I also was very ready to look at something other than teeth. And it was through many conversations with Richard and Meave and Marta and Rob that the research questions that we wanted to ask really started to take shape and that the inspiration for this project really came to be. I'll also say that Richard made sure that he let us know in every conversation that all of the results coming out of this new analysis were things that he had known all along from 72. 

And I think in many cases he was absolutely right. I'll also say that, so this research is really built on the shoulders of giants, and that's not only because of Richard, but it's also because of Bernard Wood whose incredible work on Paranthropus and really detailed descriptions of all of this collection of fossils make this conversation possible. Okay, so as I've mentioned, I do have obsession with Paranthropus. I think it's an absolutely fascinating hominin taxon. It's obviously an iconic hominin, and I think as many other fossil taxa, its history of discovery has influenced the interpretations that we have of this species, and we believe that, so one of those implications is that you have masked some of the variability within that collection of fossils. 

This variability is something that I think many people can recognize in the two most complete cranial specimens that have been ascribed to this species. KNM-ER 406 and OH 5. However, analyzing this variability in the cranial hypodigm of the species becomes a bit complicated since we don't really have the sample size to do so. So, in order to have a meaningful statistical analysis of variability in Paranthropus boisei, we have to look elsewhere. So, we decided to look at the mandibular hypodigm of Paranthropus boisei and this hypodigm is on the one hand large, it's a large collection of fossils. I think it is the largest mandibular collection of fossils assigned to a single hominin. Though I am not sure if naledi has changed that. 

And on the other hand, it's a hypodigm that contains high levels of polymorphism. And so you have the coexistence of very large mandibular corpus and relatively smaller mandibular corpus as seen here. So, there's been many hypotheses and explanations to try to understand what's going on in this collection of fossils. The most popular one has been sexual dimorphism, and this has either taken the shape of larger specimens or males, smaller specimens, or females. But there's also a lot of research that has been done on how the mandibular corpus proportions at M1 reflect a signature of the sexual dimorphism. Another factor that has been proposed is taphonomic damage and in this particular case, it's the idea that some of the specimens have matrix field cracks that have expanded the corpus and enlarged it, and some others have erosion to a degree that the proportions of the fossil have decreased. And of course, we have to consider chronology. This is a taxon that spans over a million years. So it is inevitable that we will be subsuming a large level of variation that has a correlation or could have a correlation with age. And then of course there's the elephant in the room, which is taxonomic heterogeneity. 

And this is just simply the idea that we might have a mix of taxa within this collection of fossils. What does this collection actually look like? These are some of the specimens that are part of this collection. Most of these come from Koobi Fora and one of the main challenges that you have when looking at this collection is that it's highly fragmentary. So, in order to make an analysis where we could compare specimens between them, we had to standardize certain things. So we chose to focus on the right side of the mandibular corpus, so we mirrored the things that had to be mirrored. And then we also did standardized virtual reconstructions where it was reasonable to do so filling cracks, things like that. And then we wanted to maximize both the number of fossil specimens that we could include in the analysis, but also the amount of morphology that we were going to be looking into. So, we chose the section from M2 to P4 of the right side of the corpus, which also includes the proportion of M1 that is we think has a signature of sexual dimorphism. So, we included a total of 21 boisei specimens and parallel to that, we also built a comparative sample library using six species of extant hominoids, including modern humans, all from collections that were sexed. 

And we wanted to look at shape. We wanted to do a geometric morphometric analysis, and we designed a very dense landmark atlas that included fixed landmarks, surface landmarks, and curved equidistance landmarks. And we followed a pretty straightforward GMM workflow with landmark placement patching and sliding and general procrustes analysis. Okay, then we wanted to look at how we could actually evaluate this hypothesis and this factors that had been suggested before as explanations. So, for taphonomic variable, we use Bernard Woods' 1991 variable where he classifies specimens based on their level of erosion and the level of matrix field cracks. For taxonomic heterogeneity we looked at Collin Grove's 1989 work where he proposes that there are two taxa contained within this hypodigm, a large tooth of taxon and a small tooth taxon. And then of course we wanted to look at geological age to see what was happening with the chronology. And we wanted to look at size because that's obviously something that might be affecting what's going on. For our sexual dimorphism, things got a bit trickier because we analyzed how sexual dimorphism actually looked like between closely related species of extant hominoids. And it very clear that there was a high variability in the phenotypic expression of sexual dimorphism even amongst closely related species like this two orangutans. And so that obviously creates a problem because that means that there is no single best analog to contextualize a fossil species. 

A work around this is to try to build a model that is based on multiple hominoid species and that is able to distill a global or general signature of sexual dimorphism. So how do you test if that model actually works? Well, you take out a species, you don't let the model see that, and then you test the accuracy of the model with a species that had not seen before. So, we built several linear models using this approach, and linear models really didn't do that well at all, I should say. All of these validations had really, really low accuracy, so they weren't really doing well at predicting a species that it had not been trained with. So we turned to machine learning as everybody does these days, and built a neural network that was trained again with a multiple species approach and leaving one species out. 

And after multiple rounds of training and optimization, we acquire really, really high accuracies for unseen species. And indeed, we were able to retrieve a general signature of global dimorphism in hominoids. And when we ask the model to predict what's going on with Paranthropus boisei, we get some, I think, very interesting results. And I think what's interesting is that many of these predictions actually match previous assessments that have been done before, and we only actually have three that don't match that. So, this was the way of building a way of controlling for sexual dimorphism and for testing if that was affecting the structure of the variability. But how does that variability actually look like? So, we contextualize that variability against all of the comparative sample of extant hominoids and some interesting things happen. It turns out that size at least measured in centroid size in this section of the mandible is not as variable in Paranthropus as it is in orangutans. But when you look at shape and you look at procrustes variances, Paranthropus boisei emerges as a significantly variable taxon in comparison to other hominins. 

So, when we actually take those procrustes distances and that shape variability and we map it out in a dendogram and do a hierarchical clustering analysis of that shape, we got something like this, which is just a visual representation of who looks more like who and who does not look like who. And you have a heat map of the first principle components just to give you an idea of how that's dispersed. So once we had our dendogram and had this idea of who was more similar to who we wanted to see if we could actually statistically recognize clusterings and groups within that structure. And so we ran an average silhouette analysis, which basically looks at the cohesion of clusters based on distance, and we get an optimal number of two clusters contained within this structure. We validated that using a gap statistic, which looks at the dispersion and looks at the intra cluster, some square errors, and we get an identical result with two clusters that are identical to the average silhouette. 

How do those clusters actually look like in terms of our little piece of mandible? Something like this. So these are hypothetical shapes or average shapes for each cluster, and several things really become very notable. So, there's a significant differences in the breadth of the mandibles in the shape and size of the extra molar sulcus. And then there's also differences in the buccal-lingual and mesial-distal diameters of this teeth. When we saw this, we thought we need to go back to the fossils and see if there was something outside of this piece of mandible from M2 to P4 that we have not seen, and if there was a pattern that was also there. And so we looked at several things. We definitely recognize differences in the symphysial morphology, but I won't get much onto that because there's not enough time. But I will say that we also looked at the anatomy of the P3 roots, which are sort of classically derived P.boisei phenotype. And what we see is really interesting. So, in cluster two we see a derived P3 root morphology with two flattened mesial and distal roots and in the two specimens where root morphology is distinguishable from cluster one, we see the primitive version of this, which is a two roots, but one is mesial-buccal and the other one is distal. So, we thought this was really interesting because as I said, this is a trait that is sort of almost diagnostic of boisei. 

And then we wanted to test this hypothesis, and I will go a bit fast on this, but just to say, I mean we tested sexual dimorphism with our neural network variable. We tested it for both shape and size, and we see no significance in the structure of that variability. We also tested it against our cluster structure that also comes back as non-significant. We tested for taxonomic damage, again, for shape and size, it turns out to be non-significant, for the cluster structure again there's nothing. And then we looked at chronological variation and the possibility that there were some trends happening with the shape. 

When we look at the sample as a whole, it turns out that this relationship is not significant. However, when we run the same analysis for the individual clusters, trends emerge. So we have two clusters that have a high significance with variability and time. And then we wanted to evaluate how Grove’s hypothesis matched up with our cluster structure. And this was really exciting because when we map our cluster structure with Grove’s hypothesis, we got almost a perfect match except for, oh yeah, it it's hard KNM-ER 818 up there, which is classified as part of the large tooth taxon in Grove’s, and we classify it as cluster two. So, we decided to explore this a bit further, and we built a linear discriminant function based only on Grove’s hypothesis, I should say, that Grove did not include all of the samples that we were looking at, but we regardless wanted to build a function that was only based on that. 

So we only used the samples that he had described. And when we asked that function to predict the whole set of the 21 mandibles that we were working, that creates the exact match of clusters. And so, 818 is reclassified as part of the smaller-toothed species. So we thought this was really, really exciting and really interesting and that obviously there was a relationship with tooth size that was something that we wanted to explore a bit more. And we wondered if there was a correlation with something in diet and that if that was something that we could look into. So we decided to look at isotopes. And the problem with the isotopic sample is that even though there is a lot of data out there of isotope data that is assigned to boisei, that isotope data comes mostly from isolated dental remains. And so that becomes a problem when you have a mandibular hypothesis. 

So, we wanted to see if we could extrapolate our cluster structure to the dental isolated hypodigm using mesial-distal and buccal-lingual diameters to do that. And we went first with the linear models again and got really low accuracies with that. So we ended up with a K and N that basically just looks at distances to classify samples. And we built several classifiers of this type for several teeth. And when we map this onto the cluster structure, what we got was really interesting. So there is an overlap, but there is a statistically significant, significant differences between the two clusters in terms of delta C 13 values. 

And when you map this with time, other interesting things arise. So, we have cluster one that seems to be consuming less proportion of C4 resources, and we have cluster two that seems to be on the higher end of that spectrum. We also see that there seems to be a trend towards consumption of C4 resources with time, which is something that has been noted before. And we see this trend in both clusters. So this made us think, okay, what we're seeing here, then maybe it's two solutions for robusticity or two solutions for this dietary adaptation. And we wanted to circle back to the cranial variation and the cranial polymorphism with this. 

So obviously the differences between KNM-ER 406 and OH 5 are well-documented. You have 406 that has a short facial height that has a large bizygomatic breath, and then you have things like a bar-like supraorbital taurus. And then you compare that to things like OH 5 where you have a very large facial height, a shorter bizygomatic breath, and a very different sloped supraorbital taurus. And this is differences on the frontal view of this specimens. When you look at the basal view of the specimens, differences also arise. I'll just go through the most obvious ones, but you have a high degree of prognathism in OH 5 versus 406 as shown here by the root of the zygomatic, you have differences in the positioning of the temporomandibular joint so in 406, this is a lot more laterally placed with respect to the tooth row. 

You have differences in the shape of the foramen magnum, and you have larger mesial-distal diameters in OH 5 compared to approximations that you got from the alveoli in 406. So this to us suggest the idea that you indeed have two solutions to robusticity, that the masticatory apparatus is obviously sort of at the center of this whole thing; and that because mandibles don't really just happen together and they're a part of apparatus that we would expect to see the variation that we see in the mandibular clusters and the mandibular analysis reflected in the upper masticatory structure of boisei. And so, our idea is that you have things like 406 in cluster one where you are seeing lower mesial-distal diameters, very, very derived traits and very, very extreme consumption of C4 resources. And then you have things like OH 5 that don't seem to survive that much in time, that have larger mesial-distal diameters and that seemed to be a bit less adapted to chewing than things like 406.

So, in summary, boisei has been called many things sexually dimorphic, iconic, idiosyncratic, derived, static and I think many of these things are true if you look at the big picture. But they also seem to obfuscate looking at the details of the photo and the variability that's going on there. We see indeed a high level of morphological variability and that its structure does not reflect or not independent of sexual dimorphism of taphonomy or chronology. And we recognize two very distinct morphotypes that reveal that at least there's a more complex evolutionary history to robust australopithecines in East Africa. And lastly, that reassessing all of those collections that are sitting there in museums through the light of new technology and new methods like machine learning like, 3D analysis can really, really increase the resolution that we have in our interpretations of these fossils. And with that, thank you.

View all Human Evolution videos