How Avalo Uses Machine Learning to Speed Up Crop Adaptation to Climate Change
Climate change is affecting agriculture around the world, and the solutions are rarely simple. What if you could plant crops that withstand heat, cold or drought, instead of moving thousands of kilometers again and again in search of a suitable climate? Avalo is helping make such plants a reality, using AI-based genome analysis that can cut the time and money needed to grow hardier plants for this hot century.
Avalo was founded by two friends who thought they would try their hand at a startup before committing to academic life. The company has a very straightforward value proposition, but understanding it requires a bit of science.
Large seed and agricultural companies have put a lot of effort into creating better versions of major crops. By making corn or rice slightly more resistant to heat, insects, drought or flooding, they can significantly increase yields and farmers' profits or, alternatively, make the plant viable to grow where it could not grow before.
“Yields in equatorial regions are falling sharply, and not because corn kernels are getting smaller. Farmers are moving to higher ground because saltwater intrusion is destroying their fields, but they face early spring frosts that kill their seedlings. Or they need rust-resistant wheat to survive fungal outbreaks in damp summers. We need to create new varieties if we want to adapt to this new environmental reality”
To systematically make these improvements, researchers highlight traits that already exist in the plant. It is not about splicing in a new gene, but about identifying qualities that are already there. In the past this was done by a simple method: growing several plants, comparing them, and planting seeds from the one that best shows the trait.
Today, however, researchers have sequenced the genomes of these plants and can be more precise. By learning which genes are active in plants with the desired trait, the enhanced expression of those genes can be targeted in future generations. The problem is that this takes a lot of time, practically a decade.
The complexity of the modern process stems from the fact that traits such as drought survival are not single genes. They can involve any number of complex interacting genes. Just as there is no single gene for becoming an Olympic gymnast, there is no single gene for becoming drought-tolerant rice. So when companies run so-called genome-wide association studies, they ultimately end up with hundreds of candidate genes that contribute to the trait, and then must laboriously test various combinations of them on living plants, which even on an industrial scale takes many years.
“The ability to simply find genes and then do something with them is actually quite limited as traits become more complex. Trying to improve the efficiency of an enzyme is easy: you just go into CRISPR and edit it, but when increasing corn yield, thousands or even millions of genes contribute. If you are a large agribusiness trying to produce drought-tolerant rice, you are looking at 15 years and $200 million. It is a long game”
This is where Avalo comes in. The company has built a platform for modeling the effects of changes in a plant's genome, which, according to the company, can cut a 15-year timeline to two or three, with costs falling in a similar ratio. The idea was to create a more realistic genome model that takes evolution into greater account.
That is, a system that models the genome and the genes on it incorporates more context from biology and evolution. With a more advanced platform, you get far fewer false positives among genes associated with the trait, because it filters out much more noise, unrelated genes, minor contributors and so on.
Collins gives the example of a cold-tolerant rice variety that one company was working on. A genome-wide study identified 566 "genes of interest," and investigating each one costs about $40,000 due to the required time, staff and materials. This means that studying this single trait could cost $20 million and take several years, which naturally limits both the players who can undertake such an operation and the crops into which they will invest time and money. If you expect a return on investment, you cannot spend that kind of money improving a niche crop for an alternative market.
“We are here to democratize this process. In the same dataset on rice cold tolerance, we found 32 genes of interest and, based on our modeling and retrospective studies, we know that all of them are truly causal. And we were able to grow 10 samples to confirm them”
From the start, the Avalo system excluded more than 90% of the genes that would have had to be studied individually. They were confident that these 32 genes were not just correlated but causal, having a real effect on the trait. This was confirmed by brief studies in which a specific gene was blocked and the effect was observed. Avalo calls its method "gene discovery through uninformative perturbations," or GDIP.
In part, this is an inherent capability of machine learning algorithms when it comes to separating signal from noise, but the researchers needed to approach the problem in a new way, letting the platform learn structures and relationships on its own. It was also important to them that the model be explainable, meaning its results do not simply emerge from a black box but have some rationale behind them.
The last problem is difficult, but they solved it by systematically replacing the genes of interest in repeated simulations with dummy versions that do not disrupt the trait but help the model learn how much each gene contributes.
“Using our technology, we will be able to develop a minimal set of predictive characteristics for identifying the traits of interest. You can create an ideal genotype in simulation, then carry out intensive breeding and track that genotype”
The company is working with several universities to accelerate the creation of other climate-resilient crops that would otherwise never see the light of day. These research groups have plenty of data but not many resources, which makes them excellent candidates for showcasing the company's capabilities.
