Last week was a busy week for SciBite! Along with our first ever US-based User Group Meeting (more on that in the next post), we attended the Pistoia Artificial Intelligence (AI) / Machine Learning (ML) workshop and annual conference.
The workshop featured a number of interesting presentations from both consumers and producers of AI/ML tools. While this covered a range of different use-cases within the industry (focusing on the well-established areas of image recognition and genomics analysis), there was one clear message that came through time and time again…
That while many within the industry are subject to the hype associated with AI/ML, it is not a magic bullet and it can’t work without proper scientific rigor. There was almost unanimous agreement that simply throwing the technique at a bucket of bad quality data and hoping “it’ll just work” was not the way forward and betrays the fundamental principles of science.
Much of the discussions in the meetings and over coffee focused on how to generate “clean” data, what that actually meant, and its relevance to prominent issues such as experimental reproducibility that are very pertinent right now.
This discussion brings us to a common question we’re often asked as we travel the world talking to potential customers and collaborators, ‘Are ontologies relevant in a machine learning-centric world? Can’t the AI just do it all?’ In fact, if SciBite is an “ontology-company” at heart, why do we need SciBite? As most data scientists will know, it is within standard practice to be able to train an algorithm to recognise different concept types, such as identifying potential adverse events (although this sometimes leads to bizarre consequences!).
However, this misses the key contribution that ontologies make – identifying what is known, in the context of an existing scientific framework. By annotating content with say, the MedDRA ontology, we know these concepts are adverse events, not just predictions of something that might be. While this may be obvious, in the past, ontology-based annotation was actually quite hard to achieve, both in a technical sense (finding software that could perform at scale using REST) and a linguistic one.
The work we have done to deliver ontology-based text annotation as a simple, scalable service is now a critical component in data-preprocessing/cleansing for many pharma companies.
But in a practical sense, what does this actually mean – how do ontologies make machine learning better? Let’s take the example of the gene “Insulin Like Growth Factor Binding Protein Acid Labile Subunit”. That’s quite a lengthy and definitely un-ambiguous name so you’ll see it referred to as IGFALS quite often.
If you look at this genes official synonym list, you can clearly see it’s also referred to as “ALS”. Within common biomedical discourse, ALS is used almost entirely as an abbreviation for the disease, Amyotrophic lateral sclerosis, but sometimes it doesn’t and it means this insulin-related gene.
Thus, there is a conflict as to what ALS actually means. Thus, if we just “leave it to the computer” we may find that the machine learning for the disease ALS, also mistakenly incorporates literature for IGFALS, giving an incorrect link between ALS and insulin signaling with potentially dangerous consequences for any generated model.
We see this with many more entities, such as MAP3K8, an important cancer/growth protein which is also known as “cot”. A quick search of PubMed shows that the vast majority of the uses of the word “cot” in the literature are nothing to do with this protein. Perhaps the most famous gene example is BRCA1, the well-known breast-cancer protein, which has an official synonym of “IRIS” which is of course, part of the eye.
Conversely, we also have a situation where multiple terms may be used for the same thing, with almost similar frequencies (e.g. Tylenol vs. Paracetamol or Viagra vs. Sildenafil). If the data concerning one of these terms is more skewed (e.g. “Tylenol” is more often used in the USA), then the models themselves will be biased and unreliable.
At SciBite we routinely use ontology-based data cleansing as a pre-processing step in our machine learning activities and have extensive evidence as to the value of this in critical real-world pharma use cases. By performing this step, instead of plain-text entering into machine learning models, we supply concept identifiers, which the algorithms can use to generate more reliable models by uniting different terms and eliminating the ambiguity of human language.
In turn, this aids reproducibility within in silico-based experiments, which was a further significant topic of debate at the Pistoia meeting. So, back to our question “are ontologies still relevant”? Hopefully this post demonstrates the essential contribution ontologies make when using text-based data within ML/AI activities. It was very clear from the Pistoia meeting that data-scientists are deeply concerned with the quality of data going into their models and so solutions to tackle this issue are very much required.
All in all, this was a great meeting, with some fantastic insights into how pharma are using ML/AI and we’re looking forward to being part of this community going forward.
In this post we explore this important but often overlooked topic and discuss its relevance to the work of SciBite and other groups such as FAIR and the Pistoia Alliance within the broader scientific community.Read
Get in touch with us to find out how we can transform your data
© SciBite Limited / Registered in England & Wales No. 07778456