Artificial intelligence and biology
I am interested in applying developments in the field of machine learning to biological problems. The remarkable successes of neural networks in solving many problems considered intractable by traditional computing, such as speech or facial recognition, promise expert-like automated species identification and better evolutionary inferences from molecular data.
As a postdoc at Arizona State University I teamed up with Dr Gabriele Valentini to mentor a group of Software Engineering students who were tasked with developing a proof-of-concept system for automated identification of North American ant genera. We harvested ca. 20,000 publicly available images of correctly identified ants that ranged in quality from highly standardized automontage photographs of museum specimens to phone camera-quality closeups of ants in the field. We supplemented this challenging data set with 3,000+ additional medium-quality images taken in the lab and trained the Google InceptionResNetV2 neural network architecture to distinguish among 65 of the most common ant genera found in Canada and the US. When tested on images never previously seen by the program, our model achieved overall 80% top1 and 92% top3 accuracy, meaning that the correct genus was in top three guesses the network had for 92% of images over all 65 genera. The network reached 96% top1 accuracy for the commonly encountered carpenter ants of the genus Camponotus for which there was the most training images.
This success of neural network approach motivated us to continue this work at University of Idaho. I am beginning work on a system for automated identification of pests and diseases of wheat, an important crop in the state of Idaho, and preparing a collaborative proposal to apply deep learning in phylogenetics and species delimitation.
Evolution of army ants and their kin
My dissertation research at UC Davis focused on building a taxonomic and phylogenetic framework for the research on army ant evolution. Although army ants include very charismatic species, they belong to a larger group, the subfamily Dorylinae. In addition to the army ants, dorylines comprise many cryptic ants whose biology and even taxonomy have been neglected. Partly as a result of this, even phylogenetic relationships of the army ants are not well-understood. The first step to advancing evolutionary research in the group was thus to examine the morphological diversity within this lineage. This resulted in a generic revision of the subfamily, published open-access in ZooKeys. Expertise gained during this work allowed me to design robust taxon sampling for a phylogeny of the dorylines based on next-generation sequencing data (ultraconserved elements or UCEs), published in Systematic Biology.
I am interested in developing tools that facilitate wrangling and analysis of genomic data in phylogenetics. A series of companion R scripts published along one of my papers allow other researchers to manipulate and extract information from trees and alignments. I also wrote AMAS (Alignment Manipulation And Summary), a program for fast and convenient handling of very large alignments. I am also working on a workflow that allows extraction of protein-coding sequences from UCEs. For more check out my GitHub profile.
Side-projects I was involved in in graduate school included a collaboration on the first transcriptome-based phylogeny of ants, bees, and wasps, and investigation of deep metazoan phylogeny using whole-genome data.
As an undergraduate student in Wroclaw, Poland I was a keen insect collector and published several notes on faunistics, taxonomy, and natural history.
The awe of natural world is my primary motivation and driving force behind my research.