AINEWS
AI can sketch a drug candidate in hours. The decade long, billion dollar pipeline now stalls where the model meets physical data.
The numbers have not moved. Bringing a new drug to market still takes 10 to 15 years and somewhere between $1 billion and $2.5 billion, with more than 90% of candidates failing along the way. The curve has a name: Eroom's Law — the cost of developing a new pharmaceutical has roughly doubled every nine years since the 1950s.
A recent MIT Technology Review piece, authored in partnership with Cytiva (a global lab-tools vendor), is explicit about both the promise and the new ceiling. The piece is sponsored content, so the framing aligns with Cytiva's pitch. Even so, the piece is clear about where the constraint has moved.
AI shows up first in hit identification. In this step of the pipeline, labs screen vast molecular libraries against a disease target, often a protein, looking for something that binds. The work is empirical, slow, and expensive. AI flips the order: a model proposes candidate molecules from scratch and predicts which will engage the target before any glassware is touched.
Cytiva's director of protein research strategy, Paul Belcher, describes the value in cost terms: "The main cost in drug discovery is still the clinical phase, so trying to reduce risk and increase your success rates there is obviously hugely beneficial." The bet is that better candidate triage at the front of the pipeline sends fewer duds into the expensive middle.
The compression at the front of the pipeline is real, but the constraint moves rather than vanishes. Three things now decide whether a designed molecule becomes a drug.
The first is the data the model sees. Most pharmaceutical training data is fragmented, proprietary, and shaped by the labs that own it. Belcher points at the next problem: "And it can help eliminate low-quality candidates before you have to physically test them, saving time and resources." That only works if the model is trained on high-quality experimental data, not just the published record.
The second is the physical lab behind the model. Predictive design does not eliminate wet-lab work; it changes which wet-lab work is needed and demands tighter integration between the algorithm and the bench. A core loop the piece describes connects assay output and the next model run. Cytiva sells the protein-research tools this loop depends on — a framing that naturally centers on lab integration given the company's position.
The third is clinical prediction. The MIT Technology Review piece suggests AI cannot yet reliably predict whether a candidate that binds a target in a dish will work in a human. The 10-to-15-year timeline and the bulk of the cost sit in clinical trials, which AI currently touches mostly as a planning and patient-stratification tool. The failure rate has not moved.
The hard part is whose data the model sees, how that data gets generated, and which wet-lab capacity can verify what the model proposes.
Which pharma pipelines feed proprietary data into training sets, and on what terms. Whether regulators accept AI-designed candidates without the conventional empirical-screening trail. And whether the cost curve bends once the first wave of AI-designed candidates reaches phase II.
For now, the headline figure is the one that has not changed: a new drug still takes a decade and more than a billion dollars. AI has not yet broken Eroom's Law. It has, at most, redrawn where the slope gets tested.