The PHAROS preprint presents an algorithm that searches for drug sequences toward a desired cell state
The PHAROS preprint presents an algorithm that searches for drug sequences toward a desired cell state
On September 10, the PHAROS preprint appeared: the system takes the current and desired states of a cell population as input and searches for a sequence of drug treatments that, by its calculation, moves the first toward the second. It builds on STATE, a model trained to predict how a cell population changes after drug exposure.
A conventional cell-response model takes a drug name and predicts how cells will change. PHAROS starts from the source and target states of a cell population. The authors frame the problem this way:
"Therapy discovery more often poses the inverse question: which interventions are most likely to drive a heterogeneous cell population from its current state to a desired one?"
PHAROS applies drugs one at a time: after the first step it recalculates the state of the entire cell population, then models the second step. The order of treatments therefore matters; the system ranks sequences by how close the predicted outcome falls to the target.
In an open search the authors considered 379 compounds at three concentrations, yielding 1,289,358 ordered two-step combinations. On an external set of experiments with A549, a laboratory lung cancer cell line, across three positive controls where the correct pair was already known, the exact drug pair or a substitute sharing the same mechanism of action outperformed at least 99 out of 100 random pairs and reduced the computed distance from the source state to the target by 66%, 65%, and 74%. When the direction of the transition was reversed, the advantage disappeared, and summing individual effects did not improve on the best result. These controls verified that the ranking is tied to the specified transition between cell states.
Before running a search, PHAROS checks whether the source and target states resemble the examples STATE was trained on and whether the model can distinguish between the two cell populations. In the PerturbLDM evaluation on Tahoe-100M, six models only slightly outperformed the mean drug effect for a new cell line: response differences between lines reproduced poorly across experimental batches. The PHAROS check addresses this limitation: it separates queries that fall within the model's familiar domain from hypotheses intended for the next experiment.
In metastatic breast cancer samples, PHAROS ranked nine already approved combinations, and for basal cell carcinoma it suggested directions for new ones. The authors classified both analyses as hypothesis generation: the required cell states lay outside the domain supported by STATE. The output is a shortlist of combinations for validation in organoids (small tissue models), ex vivo cultures, and patient-derived models.