Guilherme Frainer

Gross dissection

This technique is the oldest scientific discipline of medicine as the first human Histologydissection is documented in Alexandria in the third century B.C. During my masters and PhD, I have basically used gross dissections to investigate the sound-generating structures of some dolphin species. The first species I have studied was the Franciscana dolphin, which I described the development of the main structures involved in sound production, and the general differences between adults and calves. A dissection is basically removing tissues ‘layer by layer, keeping adjacent structures intact as much as possible’ as stated by Schenkkan (1972) while investigating the broadest study comparing the nasal cavity of several dolphin species. The descriptions are based on notes and photographs made during the dissections.

Centro de Estudos Costeiros Limnológicos e Marinhos - CECLIMAR, Imbé, Rio Grande do Sul, Brasil. Photo: I Moreno.

Although I have focused my studies on small dolphins (that for me are the most easy to be studied this way), I have also the opportunity to dissect, literally, from small freshwater fishes to sperm whales. For small fishes, I clear and stained (c&s) some specimens to analyze the skeleton morphology based on Taylor and Van Dyke (1985). With this technique, we remove soft tissues while coloring bones and cartilages with blue and pink, respectively. For big whales, we need a bit more infrastructure. The first time I dissected a whale with all the tools available was in Büsum, Germany, where we had a truck to hold the big pieces of blubber while we were accessing its internal organs. No, we did not could see the lens of the sperm whale, maybe it is just visible using imaging analysis such as CT scans.

Dissecting a fresh, 12m long sperm whale (Physeter macrocephalus) with Dr. Stefan Huggenberger. Büsum, Germany. Photo: N Serpa.

Literature Cited
EJ Schenkkan (1972) On the nasal tract complex of Pontoporia blainvillei Gervais and d'Orbigny 1844 (Cetacea, Platanistidae). Investigations on Cetacea

• Top

Histology and Anatomic Geometry

I have used histological techniques to assess the morphology of the nasal tract in fetal specimens of some dolphin species using hematoxylin-eosin staining. I have prepared the head of a Franciscana dolphin fetus, and compared to other species already prepared and stored in scientific collections. To demonstrate the morphology of the structures, I have digitized a series of 50 of these slices (i.e., spaced 40 µm each other) and used this image sequence as a input for 3D reconstructions using Slicer3D. In that case I have even converted the image sequence to DICOM format, but you can also use jpeg images as well. The results can be seen in this paper.

3D reconstructions from histological slices of the nasal cartilages and loose connective tissue in late foetuses of Pontoporia blainvillei, Phocoena phocoena, and Lagenorhynchus albirostris in dorsal view. For more details see Frainer et al. (2019).

I grouped these two approaches here just because I have somehow combined them in terms of my research questions, which was to demonstrate the development of soft-tissue structures throughout the development. The difference is how the sequence of images are generated and how to process them. For late fetuses and very young individuals it is possible to use conventional Magnetic Resonance Imaging (MRI) scans which takes a long time. However, computed tomography (CT) scans demonstrated to be useful to define the fat structures and other connective tissues. The CT-scan is quicker than the MRI, which makes the logistics more productive and less stressful because you always depend on the patience and disponibility of an operator that saves time to scan some dolphins' heads. Since the best scans are placed in human clinics, I put a lot of effort on trying to get access to those ones. The first time I scanned a dolphin I passed through the waiting room carrying a dead dolphin the size of a nine years old kid.

Computed Tomography on a adult male Franciscana dolphin (Pontoporia blainvillei). For more details see Frainer et al. (2015). Porto Alegre, Rio Grande do Sul, Brasil.

The images can be analyzed, segmented and build 3D of target structures using Slicer3D. Additionally, you can also measure your models with high precision, and prepare it for printing. Not only the 3D models can be stored in digital repositories, but printing these models could be useful for science outreach.

3D reconstructions of the main fat pathways (posterior dorsal bursae, blue; anterior dorsal bursae, red; melon and the branches of the melon, yellow) and the skull of Pontoporia blainvillei (A, B), Phocoena phocoena (C, D) and Tursiops spp. (E, F) in dorsal view. For more details see Frainer et al. (2019).

3D model of the sound generating structures of the Lahille's bottlenose dolphin (Tursiops gephyreus). Cologne, Germany.

• Top

Phylogenetics

This is a very broad topic and represents the base for all the studies I have done so far, or at least the way I see the world. Systematics and classification are extremely important to understand biodiversity since we can only assess and conserve what we know. I have used the phylogenetic approach to determine the phylogenetic relationship of small freshwater fishes found in scientific collections, which did not fit in any descriptions for other species in the family. The family Characidae is the largest family of freshwater fishes and I have included the specimens I had in hand into this matrix. In total, I have used 520 morphological characters already described and created five more, and also used three molecular sequences (16S, CytB, COI). I ran the matrix using parsimony and extended implied weighting (Goloboff, 2014). All the analysis was done using TNT software, but I like to sort the matrix using Mesquite.

Phylogenetic relationships of the two new species of the new genus, Dinotopterygium. For more details see Frainer et al. (2021).

• Top

Morphometric Geometrics

There are so many questions to answer using this approach but I will mention a bit just what I have done. This is the most used approach to investigate shape variation in a series of specimens stored in scientific collections. Shape variation is based on the distance and angles formed by several anatomical landmarks, which are homologous points on the anatomical structures such as bone junctions. There are basically two ways of sampling shape information: (1) based on photographs (2D), where you set the landmarks for each view you have photographed the individuals; and (2) using a 3D digitizer such as a MicroScribe, where the landmarks are taken and stored in a 3D environment.

I have used 3D morphometrics to investigate the development of the skull in some key dolphin species that exhibit distinct patterns of sound production, feeding habits and skull morphology. All the analyses were performed in R, using mainly two packages: geomorph, for basic multivariate analysis; and RRPP, for a trajectory analysis (focused on the development).

• Top

Macroevolutionary studies

Please, don’t tell anyone but this is the topic I most like to think about. Although it is easy to combine morphometrics and macroevolutionary studies, I will need maybe another life to do everything I wanted to do. The good thing is that there is a lot of information out there, and they are all published. I got the chance to use Phylogenetic Generalized Least Squares (PGLS) to demonstrate that raptorial dolphins with longer rostrum produce clicks with higher centroid frequency than other raptorial species, indicating the importance of the rostrum morphology for these animals in sound production. The analysis was performed in R using caper package The idea behind is that since some raptorial species such as the Franciscana dolphin(Pontoporia blainvillei) and the Indian Ocean humpback dolphin(Sousa plumbea) depend on the postnatal ontogeny to mature their biosonar apparatus, younger animals would be more susceptible to not perceive fishing or shark nets. For more information see Frainer et al. (2021).

The number of tooth pairs in the mandible (TP) as a function of the residuals from the linear models (a) CF ~log(Body mass) and (b) SL ~log(Body mass) relative to feeding strategy using phylogenetic generalized least squares (PGLS).

A have also created what I called the ‘phylosoundspace’, which came from the original name phylomorphospace (Sidlauskas, 2008), which I wanted to compare the frequency and intensity emitted by all odontocetes related to size. It is interesting to see narrow-band high frequency (NBHF) dolphins grouped together, beaked whales very distant from other toothed whales except for some delphinids, and an isolated sperm whale.

I am now working on a analysis of ancestral character reconstruction to investigate signature whistles evolution within Delphinida (the why of studying at ‘infraorder’ level is half of the results and discussion). For now, I can share here some graphs I want to have with me everywhere. I have been using ape and phytools packages in R.

Comparison on the centroid frequency (kHz) and source level (SLpp dB re 1µPa) among toothed whales. Tree from McGowen et al. (2020)

Comparison on the centroid frequency (kHz) and the number of teeth in the lower jaw.

• Top

Acoustics

I have started working with acoustics properly when I did a training course with Seiche to work as Passive Acoustic Monitoring (PAM) operator. But it was definitely in the postdoc I started digging deep into the topic. I have learned how to set and deploy SoundTraps, Audiomoths, HTI’s attached to a TASCAM recorder and CPOD’s.

Preparing 24 recording devices to be deployed in Mossel Bay, South Africa, to investigate the endangered Indian Ocean humpback dolphin (Sousa plumbea) together with Sea Search Research and Conservation. Photo: Unknown.

Besides the basic modules used for seismic surveys, I have been using PAMGuard for many things such as for localization with four array hydrophones in the boat, whistle detection and classification and Long-Term Spectral Average. I have prepared an algorithm in Python to automate the PAMGuard outputs such as the ‘whistle and moan detector’ and ‘whistle classifier’ modules. The idea behind is to apply a post-processing step to eliminate potential false-positives and determine what I called ‘acoustic encounters’. This code is part of a big project investigating the acoustics and movements of the Indian Ocean humpback dolphin(Sousa plumbea) in South African waters. This is the first step for what I think is ideal to start thinking on automatically identify dolphin signature whistles by its definition (see Janik et al. 2013). Another software I use a lot is Raven, which is useful to annotate sounds and it is free (sometimes it crashes in my laptop, but maybe it is because of my laptop). Very useful!

Analyzing an audio recording of a huge pod of common dolphins (Delphinus delphis) in Raven, just noticed this signature whistle that its last loop looks very similar to my dad’s signature that starts with the letter ‘N’. Delphinus nivaldii :-).

I am now diggind a bit more signal processing in Python, then using the Scipy and Librosa libraries. I have found an algorithm to perform the same LTSA as Beluga does, but I have found an error and am still not confident on sharing that. For now I am mostly interested in building sound libraries for sound classification using machine learning, so will not go to far for now.

Which kind of species is that? I bet you cannot figure it out.

• Top

Machine Learning

This is the thing that is keeping my life busy at the moment. But I mean: busy. To be honest, one of the most insane creations of the human being I have ever seen. Sometimes I think biologists would love to be a computer engineer, and vice-versa. You can try to do both, but it is definitely a tough way. Anyway, this project arose from the challenge of monitoring the most endangered dolphin from South Africa, the Indian-Ocean humpback dolphin(Sousa plumbea). There are few applications available to find dolphin sounds in long-term recordings and classify those at species level, which is essential for population size estimation using individually distinctive signature whistles as source for spatial mark-recapture models (Longden et al. 2020. However, the low accuracy for tonal calls (i.e., whistles) identification and species classification preclude further studies on target species. Imagine that it is commonly found Indian Ocean bottlenose dolphins (Tursiops aduncus), common dolphins (Delphinus delphis), and killer whales (Orcinus orca). And that’s it: we aim to build an open-source application that will help the long-term acoustic monitoring of whales and dolphins by identifying and classifying their sounds at species level using convolutional neural networks (CNNs) for audio (spectrograms) recognition. The good news is that the code is ready and I started running it at Microsoft Azure through AI for Earth grants. For now I could share partial results and what the output looks like: for the species classification, the best result gave me 97% accuracy.

Output of a random model for the species classification, built with a small amount of data.

I have also adventured myself in the world of machine learning and, by advice of my collaborator Dr. Emmanuel Dufourq, started digging into Variational Autoencoders (VAE) to generate artificial dolphin sounds. This approach will not only be useful for training the CNNs models, but also to differentiate them as it seems it would be very useful to categorize different call types for a single species (or multi-species, depending on the question, I don’t know). The thing is that you can literally see how the machine ‘learned’ and distinguished distinct types of killer whales (Orcinus orca) sounds. This is the first time this approach is done for cetaceans, as far as I am informed.

Output of the latent space created with Variational Autoencoder, which is based on the mean and variance of the training library. Each image represents 50 epochs. One epoch means training your model with all your training libraries one time.

• Top