Recursion Makes the Unknown Known in Rare Disease Drug Discovery

If one were to sketch the current landscape of AI-driven drug discovery, it might evoke Eugène Delacroix’s La Liberté guidant le peuple. It is not a romantic revolution in full blaze, but rather a long, hard-fought battle, where the first signs of victory are beginning to emerge. This is an industry navigating deflated valuations, heightened investor expectations, and a widening gap between computational ambition and clinical execution. The early passion surrounding AI in biotech has evolved into a more exacting demand: show the proof. For companies able to deliver that proof, the opportunity remains significant.

Facing this recalibration, leadership transitions carry symbolic weight. Two months ago, Dr. Najat Khan stepped into the role of CEO at Recursion, succeeding founder Chris Gibson at a moment when the sector is increasingly rewarded for demonstrating clinical progress. The question confronting AI drug discovery companies is no longer whether algorithms can generate hypotheses, but whether they can deliver clinical outcomes.

In a conversation with GeneOnline, she pointed to recent data from the TUPELO study, where REC-4881 demonstrated durable reductions in polyp burden in patients with a rare disease named Familial Adenomatous Polyposis (FAP). In a sector often defined by computational ambition and platform potential, such data offers a more concrete measure of clinical progress.

If Recursion is to separate itself from a crowded and chastened field, that distinction must rest on evidence. REC-4881 is where that case begins.

How Phenomics Reveals the “New Biology” of FAP

REC-4881 did not originate from a conventional, target-first screening campaign. It emerged from Recursion’s phenomics platform, which identified a previously underappreciated relationship between MEK1/2 inhibition and APC loss-of-function in FAP. Although MEK inhibitors are well established in oncology, their repositioning in FAP goes beyond simple repurposing. It marks a strategic shift from targeting established tumors to the chronic modulation of dysregulated signaling in a genetically defined premalignant condition. Khan describes the discovery process as “starting with the end in mind.”

Rather than beginning with a predefined molecular target, Recursion first engineered APC-deficient cells to computationally define the disease phenotype itself. High-content imaging captured subtle morphological changes invisible to the human eye, which were then translated into quantifiable disease signatures through proprietary computer vision systems and foundation models. Thousands of compounds were screened against this phenotype to determine which could restore cellular equilibrium.

“We selected diseases we would like to study, especially those without treatments such as FAP. The platform allowed us to see what human intuition could not,” Khan explained. “We weren’t looking for a specific pathway. We were asking which compound could actually correct the disease state.”

Unexpectedly, the highest-ranking compound in reversing the phenotype was a MEK1/2 inhibitor, an intervention without prior clinical precedent in FAP. For Khan, the finding highlighted the advantage of a hypothesis-free framework: beginning with observable disease consequences and allowing data, rather than bias, to identify therapeutic direction.

After in-licensing the previously shelved molecule from Takeda in 2020, the preliminary data from the ongoing Phase 1b/2 TUPELO study revealed two months ago that this phenomics-derived insight may translate into measurable clinical benefit. It includes durable reductions in polyp burden. Whether REC-4881 finally becomes a first-in-class therapy for FAP will depend on further validation. More broadly, however, its significance may lie in demonstrating that phenotype-first, AI-enabled discovery can systematically uncover therapeutic strategies where traditional target-centric models have plateaued. In an industry navigating the distance between digital promise and clinical proof, that distinction carries weight.

When Data Becomes the Control Arm

One of the most persistent structural constraints in rare disease drug development is data. Small patient populations, fragmented medical records, and poorly characterized natural history often render traditional trial designs statistically fragile and operationally inefficient.

Recursion’s response, its AI-enabled clinical development engine known as “ClinTech,” was built specifically to address these bottlenecks. “In rare diseases, the literature is often episodic,” Khan noted. “To bridge this gap, we leveraged a dataset of more than 300 million de-identified patient lives in the U.S.” From that foundation, a custom large language model scanned 256,000 physician notes linked to roughly 1,000 FAP patients, establishing real-world standards of care in near real time.

The effort was enhanced through collaboration with Amsterdam UMC, tapping one of the most comprehensive FAP registries. Applying clinical trial inclusion and exclusion criteria to this longitudinal dataset, Recursion created a robust natural history comparator. Anchoring trial outcomes against this context allows reductions in polyp burden to be interpreted against the disease’s expected trajectory. A critical consideration in single-arm studies common to rare disease research.

This data-centric approach also informs operational execution. “We don’t guess where the patients are; just like Google Maps, we have direction,” Khan explained. Using healthcare data aggregators, Recursion identifies hospital systems treating eligible patients in a de-identified manner, enabling faster site activation and more targeted enrollment. This leads to operational efficiency and clearer evidence for regulators and payers, ensuring observed benefits meaningfully exceed natural disease progression.

Recursion Operational Discipline Meets Ambition

Beyond rare disease evidence generation, the company applies the same data-centric mindset to its broader R&D operations. The platform supports compound design and optimization, where roughly 330 compounds are synthesized per advanced candidate, far fewer than the 5,000 or more typically seen in the industry. It also accelerates the journey from target identification to development candidate to approximately 17 months, compared with an industry average of 42 months.

Khan likened their strategy as an accurate model. “We simulate more and make less. We only press go on development of a compound once we are confident it is the right one. That is what our AI chemistry platform allows us to do: simulate more, make less, and lower the cost to IND.”

This operational discipline has extended to financial strategy. Since the end of 2024, Recursion has reduced its pro forma cash burn by roughly 35% without compromising programs or milestones. The integrated platform, Khan emphasizes, compounds benefits across decision-making, timelines, and cost, turning operational rigor into a sustainable competitive advantage.

Recursion strengthened its position through strategic collaborations with leading pharmaceutical and technology companies. It generated revenue through milestone payments and licensing fees tied to program advancement. Bayer, Roche and Genentech, Sanofi, and Merck KGaA are among its major collaborators. Also, in partnership with NVIDIA, Recursion built BioHive‑2, a supercomputer powered by 63 DGX H100 systems and 504 H100 Tensor Core GPUs, designed to accelerate AI-driven drug discovery. These reflect industry confidence in the scalability of its approach.

Rare disease programs such as FAP are therefore more than proofs of concept. With the platform’s ability to decode unknown biology, these programs illustrate how highly targeted populations, whether rare diseases or precision oncology subgroups, can be approached efficiently. In doing so, they open areas of unmet need while demonstrating the value of end-to-end integration from idea to clinic.

Making the Unknown Known: Illuminating Rare Diseases

Recursion’s integrated platform aims to do more than discover new molecules. It seeks to shorten the torching journey from a patient’s first symptom to effective treatment. Khan summarizes the mission succinctly and firmly: “Right diagnosis, right therapy, at the right time.”

For the 300 million people living with rare diseases worldwide, the biggest barrier is often the Diagnostic Odyssey. Accurate diagnosis can take four to five years and multiple physician visits, during which patients are frequently told “nothing is wrong” while their condition progresses. Khan drew on her prior experience at J&J in Pulmonary Arterial Hypertension (PAH), a disease that typically takes seven years to diagnose. By applying an AI algorithm to a standard ECG, her team worked with Anumana, part of nference, to reduce detection to six months. “That is what we need for rare diseases. Earlier detection enables earlier intervention,” she said.

Reflecting on the broader challenge, Khan asked, “All these drugs that we’ve painstakingly developed, why shouldn’t they be able to deliver more value? Why can’t they reach the patients who need them most?” She used REC-4881 in FAP as an example of how rethinking biology can unlock previously untapped potential.

The role of AI in drug discovery remains controversial, yet Khan frames it as a necessity. With fewer than 9% of rare diseases having an approved treatment, she points out, the traditional pharmaceutical model has delivered value but is not enough. “Rare disease innovation hasn’t changed significantly in the last 20 to 30 years,” she noted. “When people ask if we should even try AI, I ask: Is there really any other option? If you have a loved one suffering from a rare disease, you will try anything. What we’ve done so far hasn’t been sufficient.”

Underlying this effort is the pursuit of clarity—transforming unknown biology into actionable insight. Whether through computer vision that decodes subtle cellular signals or large language models that map the natural history of overlooked conditions, the objective remains consistent: make the unknown known.

Towards a Future Where No Patient Population Is Too Small for Industrialized Medicine

In an industry long optimized to treat the many by overlooking the few, Khan’s vision is reshaping how AI serves drug discovery. She positioned it not merely as a tool for efficiency, but also a means to decode the causal biology of rare diseases. “Rare disease is one of the hardest areas,” she concluded. “We aim to improve lives where no standard of care exists. It’s not just about incremental improvement. It’s about creating care where none exists.”

Indeed, the TechBio evolution signals a broader shift. When grounded in clinical reality, technology returns to its human purpose. For companies like Recursion, drug discovery is no longer simply a commercial frontier. It represents an effort to show that scientific and technological rigor, financial sustainability, and patient impact can enhance one another rather than compete.

By moving beyond theoretical abstraction into clinical execution, this model points toward a future where no patient population is too small for industrialized, high-precision medicine. Commercial viability does not diminish scientific responsibility; it enables it. In that equation, scale becomes a means of extending care instead of diluting it. If successful, even the rarest patients may no longer stand at the margins of innovation, but at its center.

Author

Steven Chung

紅杏枝頭春意鬧固然是眾人所追求的風景，但我願意說，青蘿拂行衣更能湧生感嘆。 願在海量晦澀的生醫資訊中，為您點亮一座心燈塔。鐘煌翔