Enzyme design: the AI way

Challenges in using the power of computers to create better proteins

You can find it in phones, cars, and even people’s houses! Nowadays, artificial intelligence seems to be spreading to all our everyday lives, which includes the lab work of researchers. Bioinformatic tools are increasingly becoming essential pieces for many areas. In our field of protein engineering, AI is helping researchers create better enzymes for specific purposes. According to experts, machine learning is really moving the field forward, however, it’s not all smooth sailing yet. We talked to Dr. Bas Vroling from Bio-Prodict, our expert partners in this area, to find out more about the obstacles to overcome on the way. 

A curious conundrum 

One of the main challenges of protein engineering is that “we want to find the best proteins that perform a certain activity as good (and as soon) as possible”. However, we don’t have enough money to carry out endless experiments – and this is where AI can be put to good use. Artificial intelligence can make this process less expensive by learning from experimental data to predict finer proteins without a great deal of additional tests. Well-curated models can give us much better proteins than we could have ever created by trial and error – what’s not to like? 

There’s no denying that using AI in protein engineering holds great promise, but when we set out to work, we come across a particular paradox. “Machine learning models require enormous amounts of data to make good predictions, which are not always easy (or affordable) to obtain.” 

But if we had enough resources to perform a gazillion experiments, these prediction models wouldn’t be necessary, right? We could just experiment until we get our desired protein. As Dr. Vroling explains, “when we opt for computational modelling, we wish to reduce the amount of data that we need, so we can spend less money and time on conducting experiments […] The challenge is to get to something predictive while learning from a rather limited dataset”.  

An arduous learning process

Computers work with numbers. According to Dr. Vroling, “for them to work with protein sequences, we first need to translate them, making precise numeric descriptions that we can then introduce into our model”. The better the input, the better the output: accurate translations help generate good machine learning models, which then make detailed protein predictions. 

However, although new approaches come out every day, making these protein descriptions is not always as easy as it seems… “When we work together with a partner, we get from them a set of protein sequences and a set of measurements. Then, we obtain a model that’s specific to these proteins, which helps us make suggestions to make the protein work better”. Thus, to ensure we attain the best results, we need to have a well-taught model, which depends specifically on the descriptions that we make. Models become tailored to specific translations of specific datasets and perform poorly out of their context. And this complexity only increases when considering proteins with different biological functions and environments. 

Navigating endless possibilities

Many experts agree that AI will continue to get better and better in the future. “Protein engineering, which is a pretty expensive undertaking, will get a lot cheaper, or at least, it will take a lot less time and a lot less work to get to the same quality of results”. Other technologies, such as de novo protein generation, are still “on their first steps”, but given the great progress in this area, this kind of protein design might be ready sooner than we think. 

Notwithstanding, even with the help of AI, specialists in the field are clear about something: “most things will fail”. Surely, this technology allows us to explore countless possible protein modifications, but the truth is, most of them will probably not work! That’s why collaborative work is so important – no progress can be made without feedback between bioinformaticians and lab researchers.