Artificial intelligence is reshaping how new therapies are found, but rare disease drug discovery poses a particular challenge: small patient populations, limited and heterogeneous data, and few validated drug targets. Within the DREAMS project, AI is being combined with iPSC-derived muscle models, omics data and phenotypic screening to identify new therapeutic hypotheses for rare neuromuscular disorders. Ahead of an upcoming DREAMS Expert Article exploring how these different forms of evidence can be brought together, we spoke with Nicolas Maignan, Chief Operating Officer at Kantify, about what it takes to build AI that is genuinely useful for biomedical research.
1. AI can generate an enormous number of predictions in drug discovery. What makes a prediction genuinely useful to a researcher?
In the drug discovery field, I like to present AI predictions as hypotheses. It’s never a crystal ball, as a prediction only becomes useful the moment someone can act on it and validate it experimentally. Technically, it is easy to produce thousands of plausible target-disease or compound-target associations. One of the hardest parts in AI drug discovery is producing a short list that has a high chance of working in a human-relevant model and, eventually, of working clinically. Especially when you don’t have the opportunity to validate in high throughput.
In my experience, three things make that difference. First, the prediction has to be easy to translate into a decision. If a researcher has to reinterpret the output of an AI model before knowing what to do with it, it is not yet a useful prediction. Second, it has to come with a sense of confidence, including a clear view of what the model is, or is not, confident about. Third, it has to be testable, ideally with a hypothesis attached that states what should happen experimentally if the prediction is right and, just as importantly, what would falsify it.
“Technically, it is easy to produce thousands of plausible predictions. The hard part is producing a short list of valuable hypotheses that will make science materially move forward.”
Nicolas Maignan, COO at Kantify
2. There is a great deal of hype around AI in drug discovery. What makes Kantify’s Sapian platform different?
There is a real AI drug discovery jungle right now, and I observe that for outsiders, it’s hard to differentiate between the different approaches. So I would rather talk about what is different with our AI technology, Sapian.
Sapian can be seen as a Rosetta Stone to answer to the tough questions of drug discovery: what are the targets for this disease? What are the promising and new molecules that have a chance to work? Sapian also stands out in its capacity to work throughout very different target families: ion channels, transcription factors, GPCRs, we are growing our set of validations, and have in vivo evidence in a growing number of indications. In that way, Sapian is disease agnostic and target agnostic.
Early on, we have tried to tackle drug discovery in a diferent way, not for the sake of difference, but because we believed that the status quo sometimes deserved to be questioned. This is why we do not systematically start a drug discovery project from a list of fashionable or “hot” proteins or from a pathway that a field has already agreed on. And it works at both ends of the problem: AI-driven target identification, and identifying the small molecules that act on those targets, whether novel compounds or repurposed drugs. So far, very few approaches genuinely do both and can produce truly novel and highly reliable hypotheses. This is our sweet spot, to call it that way.
We have invested considerable time to validate Sapian’s predictions. For a rare neuromuscular disorder that we worked on before DREAMS, we used Sapian to predict the likely targets. The targets Sapian predicted, and that we then validated experimentally, were completely outside the scope of what neuromuscular experts were considering. They were not variations on known biology, but proteins that were simply not on anyone’s list.
I cannot disclose much yet, but in DREAMS the hit rates we are beginning to see are of a different order from what conventional screening approaches would deliver. That is why we are genuinely excited to bring Sapian to more rare neuromuscular disorders, where patients are waiting for better treatment options.
3. DREAMS brings together iPSC-derived muscle models, omics data, phenotypic screening and AI. What becomes possible when these different types of evidence are considered together?
It is worth distinguishing between AI in general and what we are doing here. Most AI approaches in drug discovery are, at heart, very good at connecting data that already exists. They surface relationships within what the field has already measured and published. That is useful, but it tends to keep you inside known biology.
Sapian is designed to do something different: to generate genuinely novel hypotheses, including drug targets that have been overlooked or never associated with the disease, and then to identify the small molecule, new or repurposed, that acts on that target.
That is where the combination with the rest of DREAMS matters. Omics from patient-derived material tells you what is dysregulated. Phenotypic screening in an iPSC-derived muscle model tells you what actually changes cellular behaviour in a human context. When a novel computational hypothesis, molecular evidence and a cellular response point in the same direction, the arguments are largely independent, so the case for a target is far stronger than any single line of evidence. The disagreements are informative too, and often tell you something real about the biology or about the limits of the assay.
4. In biomedical research, how important is it to understand why an AI model has made a particular prediction?
For me, explainable AI is the entry point to a conversation. When we present predictions to a disease expert, the expert can immediately jump in or push back. They will tell you that a signal reflects something already known and uninteresting, or that an unexpected protein is actually plausible for reasons that never appear in the literature we trained on, or that the predicted mechanism is just surprising and worth testing. That exchange is where the real value sits. It is worth being clear that a model explanation is not a description of how the disease works, or at least not yet, and treating the two as the same thing is one of the fastest ways to waste experimental resources.
This is why we enjoy drug discovery partnerships where we collaborate with biologists, clinicians and drug development professionals who are extremely complementary to us. We have consistently found that the more open we are, the more we innovate and the faster we advance.
And I would actually add one thing that is easy to lose in the interpretability debate: what matters, in the end, is that it works. We should not lose sight of what we are seeking: better treatments, faster. So a prediction that is beautifully explained but does not translate into a human-relevant model is worth less than one that is harder to understand and does.
5. What is one misconception about AI in drug discovery that you would like to challenge?
There are many misconceptions, which I understand easily: each week we hear about a novel AI model in drug discovery. To answer your question, I think that the most common misconception is that the value of AI for drug discovery is in the discovery of drugs. Let me explain. An enormous share of attention, funding and talent has gone into molecule generation, property prediction and optimisation. That work is very valuable, and the field is becoming good at it. Our model for small molecule discovery has become remarkable, even for undrugged targets, which are a tough challenge for drug discovery, including rare diseases. But it addresses the second half of the problem. The original bottleneck sits upstream, in target discovery.
“The original bottleneck sits upstream, in target discovery.”
Nicolas Maignan, COO at Kantify
This matters because of where projects actually fail. A large proportion of clinical trial failures are failures of efficacy, not chemistry. The molecule did what it was designed to do, nevertheless it did not help patients, because the biological hypothesis behind the target was wrong, or only partially right in the relevant context.
Discovering a better compound against a poorly chosen target sadly simply lets you fail faster, and more expensively. Patients deserve better.
So the question I care about is not only whether AI can design a molecule more quickly, but whether AI can help us choose targets we are less likely to regret. That is a harder problem. The data are sparser, the ground truth is delayed by a decade, and success is difficult to benchmark. It is also where the largest gains for patients are still available. That is what Kantify’s mission, AI for human health, points us towards: not accelerating pipelines for their own sake, but improving the odds that what enters the clinic is worth testing.
“Designing a better compound against a poorly chosen target sadly lets you fail faster, and more expensively.”
Nicolas Maignan, COO at Kantify
6. Looking ahead, what scientific question are you most interested in answering through Kantify’s work in DREAMS?
The question we are answering is really challenging: how much biology is genuinely shared across rare muscle diseases, and how far that shared layer can be exploited.
Individually, these conditions offer very little data. But if distinct genetic causes converge on a limited number of downstream processes in muscle, then evidence from one rare disease becomes informative for another, and one molecule can be used as treatment for multiple indications. That would change the economics of rare disease drug discovery quite fundamentally.
This is not an abstract question for us as our work on Sapian began after a rare disease diagnosis affecting one of our own team members. Confronting how little was known, and how few groups were working on it, shaped how we approached the problem and drove us to move beyond the state of the art.
With DREAMS, we also aim at going beyond the start of the art and find novel answers which are truly useful and usable by patients and by the drug discovery community.
