Today we are speaking to Dr Layla Hosseini-Gerami, co-founder of Ignota Labs - an exciting new start-up company aimed at rescuing distressed pharmaceutical assets using Explainable AI .
Just this December (8th December, 2023), Ignota Labs claimed the OxbridgeAI challenge award among exceptional competition this year. Tell us how you feel about this win?
We feel so honoured to have won the OxbridgeAI Challenge Award this year, especially considering the high calibre of our fellow competitors. This is a great recognition for all the hard work we have been undertaking over the last year or so. We are looking forward to putting the prize money towards obtaining lab validation for our AI models.
Can you tell me who you are, your background and how you got your current role?
I’m Layla, Chief Data Science Officer at Ignota Labs, and my background is in chemistry, particularly computational chemistry. I have a PhD in machine learning for drug discovery, with specific emphasis on industrial and pharmaceutical applications.
Following my PhD, I was contacted by Jordan Lane (one of the other co-founders), who said he was launching a startup and wanted me to come on board. He wanted to use AI to improve the efficiency of drug discovery by predicting toxicity early on in the drug discovery process. I was fascinated by this angle and agreed – and the rest, as they say, is history!
Can you give me a little bit of insight into drug discovery and some of its greatest challenges?
Drug discovery is a very long and expensive process, which works like a funnel: you start with all these different possibilities for a treatment drug and you filter down to just a couple of options to take to clinical trials. Candidate drugs are assessed on how well they bind to a required disease target (efficacy), their ability to enter the cell, whether they can cross the blood-brain barrier (if the drug intended for the central nervous system), things like that. But what isn’t being tested until much later is toxicity…. This is because toxicity tests are themselves challenging and costly, so they are done in the very late stages of drug discovery when there are significantly fewer compounds left to assess. Furthermore, toxicity tests (such as cell or animal studies) are not translatable to human, in vivo toxicity. Unfortunately, there is always a risk all candidate drugs would turn out to be toxic which is especially damaging in biotech, where there are less funds available to make these kinds of mistakes. Ignota labs aims to rectify this problem and rescue drugs from failure by predicting, understanding and mitigating toxicity issues.
What is the main idea behind your approach?
Our technology integrates rich, heterogeneous biological datasets (including cell morphology data, gene expression data, pathway information and compound structures) and uses graph-based AI to predict and understand the risk of toxicities in human or preclinical species, such as hepatotoxicity or cardiotoxicity. We can interpret our predictions to identify key molecular drivers of toxicity findings, helping us to design new drugs which are safe, reducing project abandonment and saving projects from being discarded.
What is the tool behind your technology?
We called it SAFEPATH, and it’s our Knowledge Graph engine which integrates multiple data sources and predictions of discrete in vitro endpoints (e.g, mitochondrial toxicity, cardiac ion channel inhibition) and PK/ADME (Pharmacokinetic and Absorption, Distribution, Metabolism and Excretion) parameters.
How do you know your algorithm works?
We used large databases containing information about the chemical structure and endpoint measurements of previously tested compounds. We split this data into two sets, one to train our machine learning algorithms, and another, ‘unseen’ part, set aside for testing. We made sure that the compounds for testing were very different from each other to optimise the model’s ‘generalisation’ capabilities and allow it to assess completely novel structures.
We are also currently doing prospective validation, where we let the algorithms make some predictions on new data and then check in the lab if the predictions were correct. We have successfully carried out two prospective validation studies, achieving over 80% accuracy in both cases. We are looking forward to validating novel toxicity mechanisms that we have discovered, using advanced lab techniques such as human-on-a-chip technology.
What makes you stand out from other competitors?
AI, as transformative as it's been for many different Industries, is often met with scepticism, especially from medicinal chemists. You can recommend which drugs are less likely to be toxic, but would people trust what comes out of a computer? To solve this problem, we built our algorithms to be explainable, showing why the result is what it is.
In fact, we are not a software company so much as a consultancy service. We have a human in the loop, a toxicologist who looks at the outputs of the machine learning algorithms and checks the outcome. If the output looks bit funny, this toxicologist alerts us and we address the problem before running the test again. Moreover, we provide confidence scores, context (e.g. what concentration will evoke toxicity) and show which structures drive toxicity so that chemists can make conclusions or even work on changing the chemical structure of the drug to avoid toxicity.
When it comes to competitors, I also have to mention that while there are some, they tend to use public data without necessarily assessing it or its source. There are models out there built on data that we've considered ourselves but rejected due to too much noise. The public domain data we use is cleaned and contextualised by in-house scientists that have ensured what we're feeding the algorithms is high quality. We also use our own proprietary data sources to boost the chemical space that our models can cover, including an upcoming mitochondrial toxicity dataset screened through an Innovate UK SMART grant.
Are you planning to branch out in any other areas beyond this?
Other industries face similar problems in terms of toxicity of chemicals. We're focusing a lot on the pharmaceutical industry but there's the agrochemical industry, consumer goods industry, etc. We are already thinking about how our software can be used as well in those domains. We are also looking to expand our capabilities to be applicable to biologics and other drug modalities, apart from small molecule drugs.
Find out more about Ignota Labs at https://ignotalabs.ai/
