Language models reconstructed biochemical networks but struggled to predict the effects of interventions
Language models reconstructed biochemical networks but struggled to predict the effects of interventions
On July 31, eLife published a study in which three large language models reconstructed biochemical networks from the scientific literature. They recovered 24–65% of the reactions in three signaling networks, but programs built from these maps often failed to predict how a cell would respond when a protein was altered.
A signaling network describes a sequence of events inside a cell: a signal binds to a receptor, the receptor activates proteins, and those proteins change gene activity and cell behavior. Systems biologists manually compile these connections from research papers, convert the resulting map into a program, and then test what happens when a specific protein is activated or inhibited.
The authors gave GPT-5.2 Pro, Gemini 3.0 Pro-Preview, and Claude Opus 4.6 sets of genes from three validated models: cardiac muscle cell growth, fibroblast activation, and the cellular response to mechanical stretch. The models recovered between 24% and 65% of the reactions in the manually curated signaling networks. In the central metabolism of Escherichia coli, they recovered 64–91% of the reactions.
The models performed best at identifying connections frequently described in the literature, including ligand binding to receptors, the PI3K–Akt–mTOR cascade, and calcium regulation. Errors became more common further downstream, where the same signal can activate different genes or alter the extracellular matrix in different cell types.
The authors then converted the maps into logical models, in which each connection defines a condition for the next step. They compared the predictions with experimental findings reported in the literature. In the cardiac hypertrophy network, the manually curated model correctly described 94,74% of 114 tests, while the models assembled by the language systems produced 6,14–33,33% correct predictions.
Identifying plausible connections and predicting the outcome of an intervention are different tasks. The second task requires the correct sequence of connections in a specific cell, with the correct direction and activation conditions. Aging involves overlapping cellular networks, so AI systems used to identify therapies will need to pass the same test: which target should be selected, and what cellular response should be expected? This test measures precisely that ability.