From a protein to a hillside, AI narrows where to look

Choosing what to test next can be valuable even when a model cannot make the final decision. Two peer-reviewed studies published October 31 show what that looks like at very different scales: a protein small enough to help deliver messenger RNA, and a district large enough to plan for monsoon damage. DeepDelivery screened human proteins and nominated 512 candidates. The researchers selected 15 for experimental testing. Twelve worked as scaffold components in an engineered system that formed particles and delivered mRNA. That is a wet-lab result, not just a score assigned by another algorithm. The denominator matters: twelve of fifteen selected for testing, not twelve of all 512, and not an 80% success rate in patients. A protein component with useful properties is not a finished treatment. In Nepal, a separate team combined terrain, recent rainfall and rainfall forecasts to estimate which districts might experience landslide impacts within a 14-day window. It trained its models on 2016–2022 records and tested against the 2023 and 2024 monsoons. The authors report F1 scores of at least 0.7. That number combines precision and recall: how often warnings correspond to recorded impacts and how many recorded impacts the model catches. It is not a 70% probability that a specific hillside will collapse. A laboratory can afford to try several candidates that fail. A district deciding where to position supplies faces a different cost for a missed warning or repeated false alarms. The common achievement is making the next decision more focused. The next test must belong to the people making that decision.