Do LLM Proposers Improve Budgeted Search for Ionizable Lipids?

In the race to deliver life-saving medicine into the human body, scientists rely on tiny, fat-like bubbles called lipid nanoparticles. These bubbles act as protective vehicles, carrying fragile genetic instructions to cells that need them. The key to making these vehicles work lies in a specific ingredient: ionizable lipids. These are special fat molecules that can change their electrical charge depending on their environment, allowing them to open up and release their cargo once inside a cell. Designing these lipids is a complex puzzle. Chemists can mix and match different chemical building blocks to create millions of potential candidates, but testing each one in a laboratory is slow, expensive, and limited by how many experiments can be run at once. The central challenge is not just finding a good lipid, but finding the best one before the budget runs out.

For years, researchers have hoped that artificial intelligence, specifically large language models, could solve this puzzle faster than traditional methods. These computer programs, trained on vast amounts of text and data, are excellent at spotting patterns and generating new ideas. The idea was that an AI could act as a creative partner, proposing novel lipid structures that a human might never think to try, thereby speeding up the discovery of better medicines. However, until now, there was no rigorous way to test if these AI systems were actually better at the specific task of searching for new molecules under strict time and resource limits. A new study introduces a framework called LEAF to answer this question directly, pitting artificial intelligence against established computer algorithms in a controlled, fair fight to see who could find the most effective lipids first.

The researchers set up a digital laboratory to simulate the discovery process. They created a vast virtual space containing over 30,000 possible lipid structures, built from a specific set of chemical building blocks. To mimic the real-world cost of testing, they imposed a strict budget: the search could only "ask" for the performance of 40 candidates. In this simulation, the true performance of each lipid was determined by a frozen computer model trained on real experimental data, acting as a perfect, unchanging judge. The goal was to see which method—artificial intelligence or classical computer search—could identify the highest-performing lipids within those 40 tries. The study tested four different large language models, ranging from smaller to larger sizes, and compared them against two standard baselines: a random search, which picks candidates by chance, and a genetic algorithm, a classic computer technique that evolves solutions over time by mixing and mutating the best candidates found so far.

The results were clear and unexpected. Despite the hype surrounding artificial intelligence, none of the language models outperformed the traditional methods. In fact, the AI models struggled significantly with a specific inefficiency: they kept suggesting lipids that had already been tested or were nearly identical to ones in their memory. For one of the smaller models, nearly 80% of its suggestions were exact duplicates of molecules it had already seen. This meant that a large portion of the precious budget was wasted on repeating the same work. Even when the researchers tried to fix this by forcing the models to avoid duplicates, the AI still failed to find better solutions than the simpler genetic algorithm or random search. The larger models made fewer mistakes with duplicates, but they did not become better at finding the best lipids; they simply stopped repeating themselves without gaining any real advantage.

The study went deeper to understand why the AI failed. The researchers checked if the models were actually using the information provided to them, such as the performance scores of previous lipids, to guide their next suggestions. They found that while the models could be prompted to look at this data, doing so did not improve their search results. In some tests, the models were given false information about which lipids were good, and they still failed to adjust their strategy effectively. The AI seemed to be engaging with the data in a superficial way, generating reasoning text that looked intelligent but did not translate into better chemical discoveries. The study also examined the best lipids found by each method. While the AI did find some promising structures, they were not consistently better than those found by the older, simpler algorithms. The genetic algorithm and random search were just as good at finding high-performing lipids within the tight budget.

This research does not suggest that artificial intelligence has no place in chemistry, but it does provide a concrete diagnosis of where current zero-shot models fall short in this specific task. The failure was not due to a lack of intelligence in the models, but rather a lack of efficiency in how they used their limited resources. The study highlights that in a field where every experiment counts, simply having a powerful AI is not enough; the system must be able to propose truly new ideas without wasting time on repeats. The researchers concluded that under controlled conditions with a fixed budget, the tested language models did not offer an improvement over classical methods. The most efficient path to finding better lipids, at least for now, remains the proven, traditional algorithms that have been refined over decades, rather than the newest generation of generative AI.