Competing-Risk Cumulative Incidence Functions as a Unified Protein Therapeutic Lifetime Predictor

Two very different fields are colliding here in an unexpectedly elegant way. The first is survival analysis — a branch of statistics originally developed by actuaries to model when and why people (or machines, or financial instruments) fail. It's been refined for over 200 years and is routinely used in clinical trials to track patient outcomes. The second field is brand new: AI-designed protein therapeutics. Scientists can now use tools like RFdiffusion to essentially 'dream up' entirely new protein drugs from scratch — proteins that don't exist in nature but are engineered to treat diseases. The problem? These custom proteins are fragile. Once injected into the human bloodstream, they face a gauntlet of ways to fail: they can clump together (aggregation), get chewed up by enzymes (proteolysis), unfold in the heat of the body, oxidize, or trigger an immune attack. Right now, researchers test for these failure modes somewhat separately, without a unified way to think about how they compete and interact. This hypothesis proposes borrowing a specific statistical tool — the Competing Risks Cumulative Incidence Function — to treat each of those five failure modes as 'competing causes of death' for the protein. The elegant insight is mathematical: because the probabilities of all failure modes must add up to no more than 100%, there's a built-in conservation law. If you engineer the protein to resist aggregation, you're implicitly asking: does that trade off against, say, increased oxidation risk? The framework makes those tradeoffs visible and quantifiable, rather than leaving them as educated guesses. Preliminary computational work suggests all five failure modes can happen on overlapping timescales — anywhere from 30 minutes to two weeks — making the competition between them real and practically important. The reason this matters is that designing a protein drug is currently a bit like building a car without understanding how different parts wear out together. You might fix the brakes only to discover the engine fails faster. This framework would give protein drug designers a single dashboard — a unified lifetime predictor — that shows not just when a protein is likely to fail, but which failure mode is likely to win, and what the engineering tradeoffs look like before expensive lab work begins.

This is an AI-generated summary. Read the full mechanism below for technical detail.