Regeneration — Sequestering Carbon for a Brighter Tomorrow.

Siddhant Pavagadhi
16 min readMay 5, 2022

--

Optimizing genetic edits that can be performed on trees to help sequester more carbon by using Deep Reinforcement Learning.

Depiction of water levels on the rise in New York (Source)

You’re someone who loves to go on morning walks in the bustling New York environment. Every day you wake up, grab your Starbucks coffee and explore the city. What if I told you that your morning walks wouldn’t be possible in 2100.

Sounds pretty exaggerative, doesn’t it?

Well, this is the truth if the temperature increase doesn’t stagnante. You might look down from the top of the Empire State Building one day and all you will see is water surrounding you. Now that sounds terrifying. In fact, your grandchildren or your children might not have the opportunity to go on these morning walks or even live in New York due to climate change.

Predicted by climate models and climate observations, scientists have predicted that if the temperature increase reaches 2°C, then major coastal cities such as Mumbai, London, and New York have the potential to be submerged in water.

A study published in the journal: Atmospheric Chemistry and Physics, has gathered a lot of data and schematics to show the effects of a 2°C increase in temperature. According to current predictions, global warming can increase sea levels by several feet.

Rising sea levels aren’t the only effect caused by global warming. Food quality and the access to food is also degrading. Climate change is shedding off 1% of global consumable calories from the top 10 crops (corn, rice, wheat, soybeans, oil palm, sugarcane, barley, canola, cassava, and sorghum).

Additionally, 83% of the global consumable calories come from these 10 crops. This 1% decrease in consumable calories correlates to around 35 trillion calories each year. This is enough to feed 50 million individuals with a caloric intake of over 1 800. With the increase in trends regarding temperature, it’s possible for this number to grow, putting others in a worse situation.

Climate change has also worsened air quality, which in turn diagnoses people with diseases which leads to families going into poverty due to utilizing their income to help the person overcome the sickness. According to Our World in Data, the number of people who’ve died due to indoor and outdoor pollution in India has increased by 340 000 from 1990 to 2019. As someone who also originates from India, I have witnessed the air quality in large cities such as New Delhi being devastating for so many lives.

Even though there are significant methods in place to curb climate change, it simply isn’t enough due to difficulties with scalability, efficiency, and cost.

This has to end.

The Carbon Cycle (Source)

What causes all of the effects as mentioned previously in the article is the excess of greenhouse gases in the atmosphere. The trapped gases in the atmosphere contribute to global warming. The 2 biggest industries that contribute to climate change are electricity generation and food and land use. 25% of the effects of climate change are caused by electricity generation while 24% is generated by food and land usage.

The carbon cycle is a process that shows how carbon is released into the atmosphere. In fact, carbon accounts for 76% of total greenhouse gas emissions.

An essential component to remember is that carbon is needed for natural processes to function properly. Plants use carbon and organic carbon that is found in the soil to help make the soil more fertile. Below are the 3 main components of the carbon cycle :

1) Plant Respiration

The first part is through photosynthesis and plants. Trees and other plants use photosynthesis to grow although when these trees or plants die off, they release the carbon that they stored, back into the environment. Any of the carbon or metals that are absorbed by roots are a part of the soil allowing for natural processes to combat soil degradation.

2) Animal Respiration

Animals also release methane into the environment which contributes to climate change. In farming techniques, animals are usually overfed and through their digestive systems, they release methane that gets trapped in the air.

The organic waste they create is also fed into the soil which is distributed with dead organisms. These waste products and dead organisms then transform into fossil fuels that are mined.

3) Emissions Generated by Factories

These then go to factories or power plants where power is generated. The emissions created by the power plants and other industrial activities result in the accumulation of carbon in the atmosphere. Various other industrial factors other than power plants contribute to the accumulation of carbon. This can include industries such as transportation and agriculture.

Now what? What’s the method to combat this? To be straight up, there really haven’t been many technological advances in this field until now with carbon capture.

What’s up with Carbon Capture?

Process of how a DAC system works (Source)

The most effective method to reduce the effects of climate change is currently carbon capture. Due to the promise held in this solution, there are various solutions being researched. The main methods include:

  • Direct Air Capture (DAC)
  • Carbon Capture Storage (CCS)
  • Carbon Capture Pre-Combustion
  • Carbon Capture Post-Combustion
  • Carbon Capture Oxyfuel

Direct Air Capture at a high level refers to capturing carbon dioxide directly from the air. In turn, this creates a stream of CO2 which can be used for carbon-neutral fuel and windgas. However, direct air capture costs $250 to $600 per device. At scale, this isn’t economically viable.

Carbon Capture Storage involves the capturing, transporting, and storing of greenhouse gases emitted from industrial facilities. This stops the emissions from ever reaching the atmosphere, ensuring that no further damage is done to the planet. Carbon capture storage is not perfect. Facilities that have the infrastructure to employ carbon capture storage are far too sparse. Furthermore, they aren’t effective and have yet to reach the carbon sequestering expectations.

Carbon Capture Pre-Combustion is when carbon is removed from fossil fuels before combustion is completed. Fossil fuels must undergo the pressure of high temperatures and steam to collect the energy, however, removing the carbon before it’s created stops carbon at its point of emission. Unfortunately, this process is extremely expensive at around $60 for each ton of CO2.

Carbon Capture Post-Combustion refers to the separation of CO2 from other gases emitted by the combustion of fossil fuels. The CO2 is collected using solvents, sorbents, membranes, or cryogenics. This method only separates around 20–30% of carbon emissions.

Carbon Capture Oxyfuel is when rather than burning fossil fuels with air, oxygen is utilized. To ensure the flame temperature is maintained, excess gas is recycled back into the furnace. Nearly all of the nitrogen emitted is removed from the process. Although this solution is one of the best in terms of efficiency, it requires a tremendous amount of energy.

Although all of the current solutions regarding carbon capture are viable, they all contain noticeable issues that are holding back carbon capture from its true form.

The 2 Directions to Tackle Climate Change

When we think about how we can decrease the amount of greenhouse gas in the atmosphere, there are 2 approaches that companies, research labs, or institutions work on; decreasing the greenhouse gases emissions, or capturing and storing greenhouse gases from the atmosphere.

The first approach is to create technologies that require fewer fossil fuels or don’t use any fossil fuels at all. Examples of such technologies are Electric Vehicles, generating electricity from renewable energy sources, biodegradable replacements for plastics, or the electrification of factories. They aim to reduce the number of greenhouse gases that are emitted into the atmosphere.

The second approach is to create technologies that sequester and store greenhouse gases that are present in the atmosphere as discussed before. Examples of such technologies are Carbon Capture and Storage, trees (not necessarily a technology although act as natural carbon sinks), and storing liquified carbon dioxide underground.

Sequestering carbon provides a great method to curb climate change due to the fact that if scalability is achieved then enough methods of carbon sequestration can reduce the amount of carbon in the air. Where the first method of slowing down climate change falls is with economic factors. The processes cost a lot and implementation/encouragement of these products aren’t in place.

An Alternative Approach…

Living Carbon growing genetically modified trees (Source)

Carbon capture can be expensive and resource-heavy to implement in certain scenarios. This makes the process less efficient. What if we could potentially transition to a more natural method, one that utilizes the foundational components of life to build upon?

Nature has the potential to adapt quickly to change and provides sustainability to the world. Trees are an essential component of ecosystems and help with carbon emissions due to their ability to capture carbon for energy. During the process of photosynthesis, the trees capture carbon and in exchange provide us with oxygen. The carbon that is used by trees combines with water and sunlight to create chemical compounds that are used for the growth of the tree. These chemical compounds are used to kill off a harmful toxic byproduct of photosynthesis known as phosphoglycolate. For this reason, trees are a fantastic method of carbon sequestration, when they die off, some of the carbon is stored in the soil

The opportunity with this conventional method lies within the fact that we could optimize the carbon sequestration capabilities of the trees. This allows us to use the same amount of land while sequestering more carbon.

Regeneration in a Bigger Picture

Could we actually optimize the carbon dioxide sequestration capabilities of a tree, and if we could how would it work? Why are we specifically targeting its carbon sequestration capabilities? Why is our solution specifically focused on trees? Well, introducing Regeneration.

Regeneration is working on building a solution that answers all of these questions and aims to improve the efficiency of carbon sequestration of different types of trees. Before we get into what Regeneration is specifically doing to answer these questions and how it reduces the amount of carbon dioxide in the atmosphere, let’s understand what the company actually does.

Regeneration is a company that uses Machine Learning, more specifically Reinforcement Learning to find the most optimal gene(s) to edit that would increase the carbon sequestration capabilities of a tree, and use computational models to simulate this edited tree in a real-life environment to validate whether it would actually sequester more carbon.

If the tests through the simulation find that the edit has a positive effect on the goal (maximize the amount of carbon dioxide sequester), then the model gets a positive reward proportionate to the impact else it will get a negative reward proportionate to the impact. This trains the Reinforcement Learning to make edits to current parts of the genes over many iterations by fine-tunes the model made and the feedback from the reward system.

Reinforcement Learning systems don’t have a threshold or ceiling that once it is passed the model will stop improving or the accuracies will stagnate, which means that as it is fine-tuned more over iterations it essentially becomes super-human at this task, and it would be exponentially better than any other alternative method in the market.

Understanding the Details with Regeneration

To further understand our moonshot company, it’s important to recognize the small details on how Regeneration functions, in other words — the technology behind the moonshot.

Why Regeneration Uses Machine Learning

One of the biggest challenges with working on this problem in a research lab would be that it’s incredibly time-consuming and has a high probability to fail which makes the process even more inefficient. It would likely take researchers weeks or months to study the biology of a given tree, and figure out what part of the genome they want to edit. After all of this work, it’s highly likely that the edit would have an adverse impact on the tree’s ability to sequester more carbon than before the edit.

Even after multiple iterations of this process, assuming that an edited tree has a greater amount of carbon sequestration capabilities. The impact may end up being minimal and would end up taking a long period of time, money, and resources.

Instead, Machine Learning would be able to quickly reiterate its previous results and improve on them much quicker, using much fewer resources and money. This would allow it to come up with many more different unique edits that it can make to the genome of the tree, than if it was done in a research lab.

The more solutions there are, the more solutions there are to pick from for companies and research facilities that plan to implement the most optimized edited tree.

Rengeration’s Deep Reinforcement Learning System

a) What is Deep Reinforcement Learning

Diagram on Deep Reinforcement Learning (Source)

Deep reinforcement learning incorporates an extra layer to the conventional reinforcement learning systems. It includes a neural network that helps give the Agent a sense of direction. In the diagram above, the system works by observing the state or position the agent is in. By using data, the neural network helps the agent recognize some correlations between data. The agent then takes an action which directly impacts the artificial environment. Depending on the reward function that is created, the agent would receive positive rewards for a certain action and negative for another. A way to understand deep reinforcement learning systems is literally by thinking of your life.

Every time you take a certain action in your environment, you might either get a reward or a consequence depending on the action. Based on the outcome, you would know whether to repeat a certain action in the future or not.

b) How Does Regeneration’s Deep Reinforcement Learning System Create a Genetically Edited Tree

In DRL, the agent is observing/learning information that it needs for its specific domain. For example, in the context of genetically editing trees to sequester more carbon dioxide, the agent needs to understand the correlation/pattern between the given type of tree, its genetic information, and the carbon that is able to sequester over a certain period of time. In our Deep Reinforcement Learning system that is going to happen through a DNN (Deep Neural Network). Once it learns these patterns, it will know about how certain genes of a given genome have a relation to the amount of carbon sequestered, and what the effects of changing certain genes would be.

The objective function for a DRL system (Source)

For the agent to understand the objective that it has to complete (for example in this case optimizing for carbon sequestration), there would be an objective function. The objective function would help the model recognize its goal. The formula above is an example of a potential objective function. The way summation works is with the bottom expression representing the starting point or lower limit of the summation while the expression on the top of the sign represents the upper limit of the summation.

T, in this case, stands for time steps. Time steps are essentially every iteration, so the model would try an action and receive a reward ; 1 iteration. The y describes the discount factor. The discount factor is a function used to help eliminate delayed gratification in DRL systems. It helps remove a certain value from future rewards. DRL systems can predict future rewards by looking at the current state, the previous action, and future actions (through simple predictions). This would be the typical element as part of the summation. X is the state that’s present at the time step and a is the action that is performed by the model in the state.

Based on this information it can make edits to parts of the genome of a tree optimizing its ability to sequester more carbon dioxide which is the agent taking action in the environment. Although, since the data that the DNN is trained on is not enough for it to learn all of the different correlations, and since the genome of a tree is incredibly complex interactions that happen between genes of the tree and what its effects are.

Most of the learning will happen as the RL agent plays around with parts of the genome in the environment and looks at whether it had a positive impact on making progress toward the end goal, or having a negative impact by doing the opposite. The environment would be a simulation where the agent edits certain genes of the tree through mathematical representations of the genome and interactions that happen between the genes and agent.

c) How Is the Genetically Edited Tree Validated to See if It Is Able to Sequester More Carbon

This genetically edited tree is then put in a simulation that simulates a real-life scenario where a tree might be situated. The goal is to see if it’s able to sequester more carbon dioxide compared to the pre-edited tree, or does it not have the desired reaction with the environment.

An example of this could be if there was a reduction in adenosine triphosphate (ATP) in the metabolic process which would disrupt the supply and transportation of energy in the cells of trees. This would have a negative effect on the photosynthesis process.

The goal of the Reinforcement Learning system would be to create a genetically edited tree that is consistently improving the amount of carbon it can sequester.

The reward function is optimized in a way that doesn’t have a threshold or ceiling to which the reward function will be applied to the agent, this would mean that the agent would continue to improve indefinitely. Over time, its edits will become so effective and creative that the system will be orders of magnitude better than any other competing approach in the market.

d) What Method Will Be Used to Genetically Edit Trees

Diagram of how agrobacterium-mediated transformation works (Source)

Currently, there is progress being made on the number of methods to genetically modify a plant. The most common method is through utilizing agrobacterium-mediated transformation. The process works by taking agrobacterium (a type of bacteria that causes tumors on plants) and editing the plasmids within agrobacterium using restriction enzymes.

These restriction enzymes are essentially just scissors that will cut the foreign genes and use a process known as DNA ligation. This process basically helps “glue” back the DNA to create a single strand of DNA without any breaks. The plasmid is then reinserted into the agrobacterium. Explants from the plant would be used and the agrobacterium would be added. When the agrobacterium detects small molecules that are released from plant cells, the bacterium sends a message to the plasmid.

This process triggers transcription for the virulence genes. These genes are used to help make proteins for the T-DNA (transfer DNA that needs to be inserted into the plant cell) and create an entryway. One of the proteins created will help initiate the process for the T-DNA to transfer to the plant cell’s chromosome DNA.

These plant cells are further cultured and a transgenic plant is grown through this process. Companies such as Living Carbon use this method as it is a less damaging method when compared to a gene gun. Living Carbon has also seen 53% increase in biomass which indicates greater carbon sequestration alongside the increase in height when compared to a non genetically modified tree.

Next Steps/Expanding Regeneration’s Impact

Rough outline of next steps after the DRL system portion

After the DRL system creates an edited tree that has a substantial improvement in its ability to sequester more Co2 compared to the original tree validated by the computational model that runs the simulation, it can be further testing in research labs with our partners and real-life environments to ensure that it is effectively able to sequester more carbon as its validation that reinforces the predictions of the DRL system.

Our partnership would also allow our company and other research institutions/universities to come up with ways of making the gene edits on real trees that are more efficient.

Once we have validated that our DRL system has a successful result when editing certain trees to increase their carbon sequestration capabilities, we can partner with organizations and initiatives to scale up the process of genetically editing trees.

For example, Regeneration can partner with the UN for its Climate Action initiative that supports the use of nature-based methods for the reduction of carbon emissions by starting an initiative to genetically edit certain types of trees in different geographic areas, as well as figuring out how we can at large-scale genetically edit trees. As an example, this could be figuring out how to scale up the production of gene guns to edit certain trees in that area.

Potential Outline for Partnerships

a) Part 1

The first step is to get researchers from a gene-sequencing lab to sequence the genomes of trees and send the data collected to Regeneration’s dataset. They would also get the amount of carbon dioxide sequestration capabilities over a certain period of time (e.g. 3–4 weeks), and the class of the tree. This data would then be passed through the latest version of the trained Deep Reinforcement Learning model to get the resulting edited genomes of those trees which would then be sent to the simulator to test their validity.

Duration: 3–4 weeks depending on the number of trees to sequence, the amount of data, and the analysis timeframe.

b) Part 2

The second step would be to test the edits on a small number of trees in that geographic area and monitor them for 3–4 weeks to see if they are able to thrive in the environment and if they are able to sequester more carbon dioxide compared to the tree prior to the edit over the same timer period.

Duration: 3–4 weeks depending on the circumstances of the environment.

c) Part 3

The third and last step would be to work with other volunteers and employees working at Climate Action Initiative to figure out how to genetically edit the trees according to the DRL model and the tests at a large scale if the previous steps are successful.

Duration: 4–8 weeks depending on factors like the geographic size of the location, workforce size, or problems faced in the development of a large-scale implementation.

Regeneration isn’t just a company with people working on their own thing. It’s a community where our passion and vision for a better future help us work every day on this very ambitious project.

Creating diverse regions with the optimized genetically modified trees to indirectly help save some of the world’s biggest cities from being submerged in water to helping provide nutritious food for individuals by eliminating the harsh effects of climate change.

We also imagine a world without as many carbon emissions. By creating an effective business model, we hope to attract various individuals/organizations that are willing to join us in our vision to help create a greener future.

At Regeneration, it’s not about thinking short term, it’s about dreaming big while maintaining feasibility to solve one of the world’s most pressing problems. All of these dreams will come alive with 1 thing — the next generation of DRL systems that will help in providing the world with optimized genetically modified trees.

Regeneration. Regenerating the future for future generations.

--

--

Siddhant Pavagadhi
Siddhant Pavagadhi

Written by Siddhant Pavagadhi

16 y/o working on researching mathematical concepts related to multivariable calculus, abstract algebra, and complex analysis.

Responses (1)