How do you find synthetic lethal targets without screening everything?

Reasoning layer: to test whether an LLM can navigate this space, we used a problem with a known answer. We took Olivieri et al.'s 2020 CRISPR screen - which gene knockouts sensitise cells to olaparib - as ground truth, then gave the same ~10,000 pairs to several open-weight LLMs and asked each to score every pair from −1 to +1 (−1 rescue, 0 no effect, +1 lethal). To rule out memorisation, we re-ran the whole thing on a CRISPR screen released after the models' training cutoff.

Scaling layer: we widened to ~400,000 pairs of clinically relevant genes (in an RPE1 TP53-deficient background) and layered two more agent-assigned scores - novelty (an agent with literature access checks whether a similar pair was already tested) and feasibility (pan-essential genes, for instance, are harder to validate). Filtering to the top 100 by synthetic-lethality score with novelty and feasibility both above 5 left 11 pairs.

Validation layer: each surviving pair goes to Hydra - our general agentic tool. Give it a hypothesis and optional data and it plans the bioinformatics, runs it across hundreds of skills (DepMap, TCGA, cBioPortal, Ensembl…), and returns a ready-to-read report on whether the hit is worth testing and which patients would benefit. The 11 pairs are now prioritised for experimental validation.

The best untrained model reached 73.8% accuracy (AUC 0.768) - beating the purpose-built tools SLant and MAGICAL - despite never being trained to predict CRISPR knockouts. On the unseen, post-cutoff screen it held ~60%: a modest drop that points to real generalisation, not data leakage. It echoes coding agents - no one trained them to spot concurrency or security bugs; the capability emerged. Above-random performance on screens the model never saw hints at something similar happening in biology.

Scaling collapsed ~400,000 pairs to 11 testable hypotheses - impossible to surface by hand. But 11 knockout pairs are a starting point, not an answer: are the genes differentially expressed in tumour versus healthy tissue, in which tumour types, for which patients? That is what Hydra answers. On the TP53 + TUBGCP2 pair it queried Ensembl (the genes sit on different chromosomes, so cis-codeletion is impossible), pulled DepMap (TUBGCP2 is pan-essential - broad inhibition would be systemically toxic), and found no real TP53-TUBGCP2 correlation. Yet the same chain surfaced a better angle: glioblastoma frequently shows chromosome-10 monosomy alongside TUBGCP2 overexpression, leaving a single copy - a plausible, tumour-selective target.

An LLM score is a powerful prior, not a verdict: specificity is low (false positives), and the models show a bimodal preference for particular scores - so survivors must be grounded in real data before anyone touches a pipette. That is the division of labour: the scoring module ranks the space, Hydra validates the shortlist on omics, and only then is a pair worth testing experimentally. The same reason → analyse → validate loop applies anywhere combinatorics outrun lab capacity.

Hydra turns a one-line prompt into hundreds of analyses. A single TP53/TUBGCP2 run produced over 190 results in orchestrated mode and synthesised them into a summary paper, so the volume stays readable. What used to take a bioinformatician days now runs overnight across dozens of programs - the scientist comes back in the morning to verify and judge, not to execute.