AI is set to transform the study of evolutionary morphology – an interview with Yichen He

Recently, IOB published an excellent introduction and overview of the current and potential future applications of AI to evolutionary morphology.

Opportunities and Challenges in Applying AI to Evolutionary Morphology

Y He, J M Mulqueeney, E C Watt, A Salili-James, N S Barber, M Camaiti, E S E Hunt, O Kippax-Chui, A Knapp, A Lanzetti . G Rangel-de Lázaro, J K McMinn, J Minus, A V Mohan, L E Roberts, D Adhami, E Grisan, Q Gu, V Herridge, S T S Poon, T West, A Goswami

Fig. 1 Broad definitions, relationships, and differences between artificial intelligence, machine learning, and deep learning, the sequential development of each successive subset, and their broad introductions over time (Carbonell et al. 1983; Goodfellow et al. 2016).

The submitting/lead author , Yichen He took time out to elaborate on this paper and the concepts it highlights.

When was your very first interaction with any type of AI in your academic journey?

My first interaction with AI in my academic journey dates back to my undergraduate studies, where I learned about neural networks and various AI algorithms. However, the first application of AI for an academic project came in 2017 during my PhD, when deep learning and deep neural networks were gaining widespread attention. I applied AI models to assist in measuring digitized natural history specimens, which marked the beginning of integrating AI into my biological research.

Why do you feel that there is still so much controversy in certain areas of science and academia at large about the use of AI?

One crucial step in using AI effectively is to thoroughly review whether it is suitable for the task at hand. While it’s beneficial to explore off-the-shelf, user-friendly AI methods for analyzing datasets or solving problems, researchers need to assess these tools to ensure they align with their specific research objectives. Many scientists remain cautious because implementing AI can be complex and often requires specialized knowledge, which not everyone may possess. Moreover, concerns about data quality, potential biases, and the ‘black box’ nature of some AI models—limiting interpretability—contribute to this hesitation. To address these challenges, researchers should begin with reliable, easy-to-use methods and conduct small-scale, proof-of-concept studies before fully committing to AI for their projects.

This IOB paper opens with “The rapid proliferation of tools using artificial intelligence (AI) has highlighted both its immense potential and the numerous challenges its implementation faces in biological sciences.” Could you detail a bit for us about which of the challenges you feel are creating the biggest hurdles to people engaging more with AI for scientific purposes and why you feel those are the biggest?

We consider the biggest challenges are data quality, technical expertise, and interpretability. AI needs high-quality, well-labeled data to perform well, but such datasets are often limited in biological sciences. Additionally, using AI requires specialized skills, which not all researchers have. Finally, the ‘black box’ nature of many AI models makes it difficult to interpret results, which can reduce trust and hinder adoption in scientific research.

Part of this paper focused on 10 case studies in which AI can benefit evolutionary morphological studies. Were there any other case studies that did not make the cut that are of interest to you ? Can you detail for us the decision to leave this one (or more than one) out and why you ended up selecting the 10 for the paper?

One case study we considered but ultimately decided to leave out involved using AI models as automatic extractors and descriptors of morphological traits through metric learning. This approach trains AI models to create a latent space that represents morphological traits, which contrasts with the more common AI-in-morphology methods that predict annotations (e.g., segmentation) and then use predefined approaches to extract traits for projection into color spaces or morphospaces.

Although this technique shows considerable promise, it has only recently begun to be applied in the field, with current research mostly centered on 2D images focused on color and color patterns. Given its novelty and the need for further testing to validate its effectiveness, we decided not to include it in the paper. However, it has significant potential; hopefully, we can see more studies on metric learning in evolutionary morphology..

How has your work on this paper and with AI in general springboarded you into other areas of focus for research?

Working on this paper and with AI has opened up new research areas for the team, especially in developing AI-driven tools for measuring phenotypic traits and exploring AI in phylogenetic studies. It’s also inspired us to apply AI techniques to more biological datasets, pushing forward methods that can uncover complex patterns in evolutionary biology.

What are some recommendations you would have for those in academia (the sciences) who are wanting to be more convinced of the pros of using AI?

I would recommend first identifying the specific task you want AI to perform, such as predicting segmentations or recognizing species. Focus on applications where AI can save time or do tasks that traditional methods may not capture. Next, choose AI methods that match your skill level, ranging from user-friendly tools to full-code implementations. Build a pipeline, starting with a small pilot dataset to see if it produces good results. If the pilot is successful, proceed to the full dataset, iterating and optimizing the pipeline to enhance performance and efficiency.