As written in the well-known textbook Developmental Biology, “One of the critical differences between you [as an embryo] and a machine is that the machine is never required to function until after it is built.” 

Yet, scientists are turning to machines to help them model and understand the very thing machines don’t have to do — develop and evolve their function.

Our understanding of human development continues to grow, but these findings often remain disparate and disconnected from the bigger picture of development. For one a variety of model systems help make discoveries, from mice to fruit flies, each focusing on a few specific cell types or a single organ system. 

To better understand this complexity, scientists have been turning to computational models. 

The Virtual Human Development Consortium (VHD) — an international group led by scientists at U of T and the University of British Columbia (UBC) — is contributing to this modelling effort by building “a computer-based simulator of human embryonic development.” 

The VHD, alongside other groups around the world, is attempting to bring together developmental findings and knowledge across many tissues, experimental models, and scales with the help of experimental, theoretical, and computational biologists. 

 The project was founded in 2021 by Maria Abou Chakra, a research associate in the Bader lab at U of T, and Nozomu Yachie and Nika Shakiba — a professor and assistant professor, respectively, from UBC. The VHD is still in the early stages but the group has ambitions of making the virtual simulator a reality with the help of this community.

Models: what are they good for?

Creating theoretical models of biological phenomena has a decades-long history, spanning disciplines from ecology to immunology. Nowadays, much of the theoretical work involves coding behind the computer, explains Abou Chakra, in an interview with The Varsity.

Modelling can take many forms and have many goals, but Abou Chakra explains her preferred approach: “We try to simplify things and try to capture [biological processes] with as few steps as possible. So if you can capture 90 per cent of the phenomenon with one or two rules, why would you add all the complexity that you think exists? […] Complexity doesn’t always mean accuracy.”

Abou Chakra doesn’t see these research models as the endpoint but rather a way of opening up new avenues of research. “It shouldn’t just answer a question and end there. Is it also generating a new question for us? […] Did it open up a new window that we have to explore?” This exploration can come in many forms, including collaborations with experimental biologists.

To illustrate this, Abou Chakra brings up her work on a digital cell model. She began by looking at the current literature on what defines a cell, settling, for the time being, on rules around cell death, cell size, and the cell cycle. But throughout this process, she discovered a gap in scientific knowledge: how does the rate of cell division change early in human development?

“What I couldn’t find was the rate of the slowdown [of cell division]: was it linear, gradual, abrupt, or exponential?” This eventually led to “proposing a whole new mechanism for the cell cycle, which not only controls cell numbers but […] also regulates the diversity of cell types,” explains Abou Chakra.

Once a model is created, scientists can make tweaks or disturb the system and observe the effects. Does a disturbance to one of the model’s rules result in a disease we see in humans? Bingo — there’s a new avenue of research to explore further. 

This feature is particularly helpful in ethically thorny areas of research, many of which exist in the world of development. For instance, there are guidelines and regulations on culturing human embryos in the lab for longer than two weeks for several reasons, including fears of the embryos gaining the ability to feel pain. This has earned the two- to four-week time period the name “the black box” of development. 

Models don’t share these ethical limitations and can shine some predictive light, at the very least, on this developmental time point.

Challenges that lie ahead

One of the biggest barriers to overcome with such an ambitious project — which spans three continents and nine countries — is to secure a thriving community with active participation. Building community and bridging gaps in communication between biologists and theoreticians from a variety of backgrounds has been a focus of the VHD from the beginning. 

In an interview with The Varsity, Sidhartha Goyal, an associate professor at the U of T Physics Department and member of the VHD, said he joined “mostly because [he has] good friends there and… good science happens… because we work with people we enjoy working with.”

Goyal’s lab applies physics to many biological contexts. In a recent project led by Mehrshad Sadria — currently a machine learning scientist at the American Altos Labs with senior authors Goyal and Gary Bader — a professor at U of T in the Department of Molecular Genetics and the principal investigator at the Bader Lab — they created the artificial intelligence (AI) model Fatecode. 

Using information about the genes that interact with a certain cell type, Fatecode can accurately predict which genes regulate the cell’s identity. Finding these regulatory genes traditionally requires a long and arduous process, which Fatecode could help simplify in the future. 

But with AI comes all sorts of questions, including those around explainability, which is scientists’ ability to explain why a model is making the predictions that it is. Goyal thinks our approach to explaining things may change altogether. 

As models become increasingly complex, it becomes harder for scientists to fully understand the reasoning behind their predictions. Some models may intentionally be kept simple, but others won’t. 

“As a physicist, as a scientist, I’m always obsessed [with] what can I explain? [But] what we call an explanation may genuinely change […] now that we can go from data to prediction without having the middle layer sorted out enough,” said Goyal.

We may find this lack of understanding acceptable in some situations, where the ultimate answer is all we care about. But beyond humans’ inherent want for explanations, not being able to understand why a model comes to a certain conclusion could cloud scientists’ ability to discern when the model’s conclusions are bogus.

Both Abou Chakra and Goyal also stressed that AI alone can’t replace modelling efforts and that modelling is more than just plugging data in and getting a prediction out. “It’s like saying, I’m going to pick one microscope to look at things, only one way to look at things,” says Goyal. 

AI is great at making predictions based on vast amounts of data, for instance, but if the data inputted into the model is faulty or biased, the predictions will be too. AI is a tool in the toolbox, not the toolbox itself.

Rather, the toolbox consists of a wide variety of computational and experimental tools and, more importantly, the expertise of the interdisciplinary team at the VHD who continue to tackle the unanswered questions of developmental biology. “It’s this united front that needs to happen,” as Abou Chakra aptly put it. 

At its core, the VHD is about removing barriers through bringing together disparate fields, isolated experiments, or sparking new collaborations. This electric mixture of sciences and people brings a whole new dimension of research to developmental biology and shows the importance of collaboration in science as a whole.