AI vs. rare diseases: fixing the talent bottleneck before the biology
Modern biotech can edit genomes and fold proteins in silico, yet most of the ~10,000 known rare diseases still have no approved therapy. The limiting factor is no longer just lab equipment or capital – it’s people. There simply aren’t enough experts to explore every niche indication or test every possible molecule.
According to reporting from TechCrunch on Insilico Medicine and GenEditBio, AI is now being positioned as a multiplier for scarce scientific talent, not a replacement. That shift matters: if it works, the biggest impact of AI in health may be in diseases pharma has historically ignored because the human labour simply didn’t scale.
In this piece we’ll unpack what’s actually new here, who stands to gain or lose, and where Europe fits in.
The news in brief
As reported by TechCrunch, two biotech startups used Web Summit Qatar to showcase how they’re applying AI to rare disease R&D.
Insilico Medicine’s CEO Alex Aliper described a push toward what the company calls “pharmaceutical superintelligence”. Insilico has launched an “MMAI Gym” environment aimed at training general-purpose large language models – the same family of systems as ChatGPT and Gemini – to handle multiple drug discovery tasks at once. The platform ingests biological, chemical and clinical data to propose disease targets, design candidate molecules and even suggest ways to repurpose existing drugs, including for conditions such as ALS.
GenEditBio, led by co‑founder Tian Zhu, is part of the second generation of CRISPR players focused on editing directly inside the body. The company uses AI to mine a large library of non‑viral nanoparticles and "engineered protein delivery vehicles" to deliver gene-editing tools precisely to organs like the eye or liver. Its NanoGalaxy system closes the loop between high‑throughput in vivo experiments and machine‑learning models. The firm recently received FDA clearance to start trials of an in vivo CRISPR therapy for a rare corneal disease.
Both companies frame AI as a way to compensate for a shortage of qualified scientists and to generate the large, high‑quality datasets that current models still lack.
Why this matters
Drug discovery has always been constrained by human bandwidth. Each promising target or chemical series requires years of effort from multidisciplinary teams – medicinal chemists, structural biologists, clinicians, regulatory experts. For common diseases with huge markets, that overhead is justified. For rare diseases affecting a few thousand patients globally, it often isn’t.
That’s the crux of the “labour issue” highlighted in the TechCrunch piece. AI doesn’t just automate tasks; it changes the economics of what is worth pursuing. If systems like Insilico’s can explore vast chemical spaces and propose high‑quality candidates with less human iteration, the fixed cost per indication drops. Suddenly, a niche metabolic disorder or an ultra‑rare retinal disease might clear the business-case hurdle.
The immediate winners are obvious: patients with rare or neglected conditions, and the clinicians who advocate for them. Also benefiting are small biotechs and academic labs that can license or partner on these platforms instead of building AI teams from scratch.
But there are potential losers. Contract research organisations built around brute‑force screening risk being displaced by algorithmic pre‑filtering. Mid‑tier pharma companies that fail to modernise their discovery pipelines may find themselves paying a premium to access tools their competitors helped build earlier.
Risks run deeper than market share. If models are trained primarily on Western, well‑phenotyped cohorts – as Insilico itself warns – they may systematically underserve patients in the Global South or in under‑represented European populations. Overconfidence in “superhuman” models could channel investment into targets that later fail in heterogeneous real‑world populations.
So yes, AI can relieve the labour bottleneck. But it also concentrates power with whoever owns the best data and models – and raises the cost of being wrong.
The bigger picture
This story fits into a broader trend: biology is becoming a data‑centric, software‑like discipline.
Google DeepMind’s AlphaFold and newer models for variant interpretation have shown that neural networks can learn fundamental rules of life from large datasets. Companies like Recursion and Exscientia are building closed‑loop systems where robots run experiments, AI analyses the readouts and designs the next batch, and the cycle repeats. Insilico and GenEditBio sit squarely in this movement from one‑off projects to "platformised" pharma.
We’ve seen a version of this before. In the 1990s, combinatorial chemistry and high‑throughput screening promised to flood pharma with hits. The bottleneck simply moved downstream to validation and understanding mechanism of action. Today’s AI‑native platforms risk a similar fate if they generate candidates faster than biology and clinical trials can absorb them.
What is different now is the integration across layers: target discovery, molecule design, delivery engineering and – soon – digital twins and virtual trials. When GenEditBio talks about AI‑designed delivery vehicles for in vivo CRISPR, it’s part of a shift from editing cells in a dish to rewriting instructions directly inside the body. If this becomes semi‑standardised – “off‑the‑shelf” delivery shells plus custom payloads – gene editing starts to look more like software deployment.
Competitively, the likely endgame is a small number of horizontal platforms powering many vertical programmes. Big Pharma will continue buying or partnering with these capabilities rather than trying to recreate them internally at full scale, much like banks now rely on cloud providers instead of running their own data centres.
The risk, again, is concentration: a handful of AI‑bio platforms could become gatekeepers for which diseases get serious attention.
The European and regional angle
Europe is unusually well positioned – and constrained – in this field.
On the one hand, the EU has some of the richest rare‑disease infrastructures: European Reference Networks, strong registries for conditions like cystic fibrosis, and a long‑standing Orphan Medicinal Products Regulation that created incentives for niche indications. Countries such as Finland and Estonia run population‑scale biobanks that could be perfect fuel for AI.
On the other hand, GDPR, national data protection laws and the (still‑forming) European Health Data Space create a complex compliance puzzle. For an Insilico‑style platform, being able to integrate genomic, clinical and imaging data from multiple member states is a competitive advantage – but negotiating data access and consent frameworks across 27 jurisdictions is slow and expensive.
Culturally, European societies – particularly in DACH and parts of Southern Europe – are more cautious about gene editing and data‑driven medicine than the US. That may slow deployment of in vivo CRISPR like GenEditBio’s, but it also pushes companies to build stronger safety, transparency and governance practices from day one.
European players aren’t starting from zero. BenevolentAI, Owkin, Evotec, BioNTech and a long tail of startups from Berlin to Barcelona are building their own AI‑first discovery engines. For smaller health systems in Central and Eastern Europe facing clinician shortages and emigration, AI that stretches limited specialist capacity could be a lifeline – if they can participate as equal partners in data partnerships, not just as raw‑data suppliers.
Looking ahead
Over the next three to five years, expect AI’s impact in rare diseases to be felt in three concrete areas.
First, target identification and molecule design will continue to accelerate. We’ll see more early‑stage assets where the story is explicitly "AI‑designed", especially in neurology and ophthalmology where unmet need is high and trials can be smaller.
Second, drug repurposing for rare indications will move from opportunistic to systematic. Platforms like Insilico’s can scan existing pharmacopeias for overlooked matches. If even a small fraction of these hypotheses hold up, regulators and payers will be under pressure to create faster, cheaper trial pathways for generics in new rare‑disease settings.
Third, delivery engineering – GenEditBio’s focus – will likely be where the first spectacular failures and successes occur. The safety bar for in vivo CRISPR is extremely high. Any serious off‑target effects will trigger backlash. Conversely, a well‑tolerated "one‑and‑done" edit for a disabling condition would reset expectations for what is treatable.
Looking 10–15 years out, realistic digital twins are still aspirational, but we may see regulatory pilots where virtual cohorts are used to refine dose ranges or stratify patients before physical trials. The EU’s AI Act and emerging FDA guidance will probably demand detailed documentation of how models were trained, validated and monitored – a compliance burden, but also a competitive moat for those who get it right early.
For readers, the key signal to watch is not flashy benchmarks or "superintelligence" branding, but clinical readouts: Do AI‑designed therapies reach phase 2 and 3 with better success rates, and do they reach neglected patient groups?
The bottom line
AI will not magically cure rare diseases, but it could finally make them economically and logistically tractable by relieving the human talent bottleneck. The combination of algorithmic design, automated labs and smarter delivery systems reshapes which diseases are worth pursuing and who gets a say.
For Europe in particular, the open question is whether we choose to be merely the world’s most regulated data donor, or a genuine hub of AI‑driven drug creation. The answer will depend less on our models than on how boldly we share data, fund platforms and update our rules.
