It is a biologist's daily business to look for correlations in data, and even though correlation does not imply causation, it can provide a starting point for further experiments. If manipulation of one factor in a system changes the response, scientists assume there is causality. This well-tested approach starts to run into trouble when systems become more complex.
Imagine a fish swimming around a tank, changing directions and speed at will. A researcher tasked with reconstructing an image of the animal in the tank gets ten snapshot images. Because the fish shows complex and fast movements, reconstructing the body shape simply from averaging over the images will result in a blurry shape. "This is the problem with statistics; if the system is neither static nor in equilibrium, you will not get a clear picture. We often try to linearize problems for statistical analysis even if the underlying relationship is not in equilibrium, ending up with a blurry image," explains Prof. Gerald Pao.
Developing and utilizing methods analyzing datasets based on time series is the focus of Prof. Pao's Biological Nonlinear Dynamics Data Science Unit, which was established at OIST in the spring of 2023. The researchers in the unit developed a mathematical technique called causal compression to find causal relationships between things that are not correlated.
An example is the cell cycle, which relies on various input signals to run smoothly. The product of different genes can be necessary to keep a cell moving through the cycle; however, no gene alone is sufficient for the process. Classical statistical analysis at a given time would reveal no correlation between the expression of the genes and the cell entering the next stage of the cycle. "When recording a time series of the same genes instead, we found a causal connection that allowed us to make predictions about the cell cycle. Because of how scientists are trained classically, these things often get overlooked," Prof. Pao points out.
The unit used this approach and built a computer model to simulate the behavior of a fruit fly. The researchers trained the model on recordings of neuronal activities, and the virtual fly behaves like its real-world counterpart and mimics brain pattern activities astoundingly accurately.
When recording the fly's movements in the lab, the animal moves freely on a styrofoam ball and sometimes pauses during the experiment. When training their model, the researchers only used data recorded during the active movement periods of the fly, "even though we did not include any data from the pauses the fly took, our virtual fly showed the same behavioral pattern and took long breaks in between while moving on the ball," says Prof. Pao.
Using these time series allows the team to move away from a narrow slice of a problem and look at it during every possible stage. The researchers now want to expand their approach to downloading the human brain. However, getting the necessary data to build models for the human brain is difficult. Currently, functional magnetic resonance imaging (fMRI) is the state-of-the-art method to obtain functional neuronal data. This machine measures blood flow in the brain to determine which areas are active at a specific moment. For data scientists, this method has several downsides, including a low spatial and temporal resolution, no distinction between single neurons, and difficulties in interpreting if an area has an activating or inhibiting function.
To achieve their ambitious goal, the unit's scientists must therefore first find a way to get better data for their tasks. "Currently, we are working on making a mammalian brain transparent, which would allow us to measure activity on a single neuron level," Prof. Pao explains.
If an object scatters light, it becomes opaque. While water and air are transparent, clouds are not because the water and air inside the cloud have different optical densities, which the refraction index quantifies. "In a cell, the membranes have a different refractive index from the cytoplasm, which leads to light being scattered," says Prof. Pao. This way, both individual cells and entire organs become opaque. However, cells turn translucent if filled with a substance that reduces the difference between the refractive indices and reduces light scattering. Unfortunately, typical methods based on this principle, lead to the death of the cells. To allow the refraction index matching substances to enter the cell, its membranes have to be perforated, causing the cell to die.
With his team, Prof. Pao developed a new approach to make cells translucent, based on a protein found in the skin of squids called reflectins, that come from a structure named iridophore. "Out of these proteins, we developed nanoparticles that have a similar refractive index as the cell membrane," says Prof. Pao. Because the nanoparticles are smaller than the wavelength of visible light, they themselves do not scatter light. In principle, this characteristic makes them a suitable approach to turn living cells translucent.
As unusual as his research is, so was his path to becoming a scientist. After seeing a documentary on DNA replication in 5th grade, his appetite for science could never be satiated. After completing an undergraduate degree combining molecular biology with biophysics, Prof. Pao continued to earn an M.D. at the University of Washington and a PhD at the Salk Institute. Ultimately, a paper on theoretical ecology took his career in yet another unexpected direction.
George Sugihara showed in a paper that causal connections can be detected from time series data without experiments. "This changed how I think about the world," Prof. Pao says. After getting in touch with Prof. Sugihara, who currently works at Scripps Institution of Oceanography (the oldest oceanographic research institute in the world), the then newly graduated Dr. Pao joined his lab in applied mathematics as a postdoctoral researcher. Switching fields like this quickly humbled the ambitious young scientist, "I had to start from zero. I was a decent molecular biologist but had to start from scratch with mathematics. But my unconventional background has also helped me by giving me a different perspective and a set of tools that helped me find unique solutions," Prof. Pao recalls.
A maxim that made OIST a good fit for his new lab, where he established collaborations with several other units, including neuroscience and nanomaterial units. With their approach, they can get more results out of time series datasets than with classical statistics alone, "by collaborating among units, we will get results none of us can get alone," Prof. Pao says.
