Our Take
Turbine's platform finds real dependencies DepMap doesn't, but the company conflates in silico prediction with clinical proof; two biomarkers in trials is validation, not scalable discovery.
Why it matters
Precision oncology teams rely on DepMap for target discovery, but synthetic lethality and complex perturbations require richer data. Turbine's results show the gaps are real and fillable—but only if the predictions translate to patient response consistently.
Do this week
Oncology data teams: review your current target-ranking pipeline against Turbine's synthetic lethality framework before the next investment committee cycle so you can identify which dependencies your screening is missing.
Two new biomarkers entered clinical trials after virtual screening
Turbine, a computational biology platform built on machine learning, completed a 30,000-combination perturbation screen across 62 patient-representative cell lines spanning seven cancer types. The goal: identify why roughly 40% of PARP inhibitor-treated patients with BRCA mutations don't respond to treatment.
The screen surfaced 13 biomarker candidates. Eight were validated in vitro. Two entered clinical panels and are now used to guide patient selection in multiple PARP inhibitor trials (company-reported). One of these biomarkers was unknown to DepMap's public database at the time of discovery.
In a separate experiment, Turbine's DNA damage repair-specific model identified NEK1 as a dependency in PARP-resistant medulloblastoma and head and neck carcinoma cell lines. The NEK1 finding was later validated in vivo in HSC-2 cell lines and was not visible in DepMap screening alone.
The platform ingests multiple data modalities: high-throughput screening data, clinical outcomes, chemical compound libraries, and immune research datasets. It can run 50 million in silico experiments per day and has been accessible to clients since 2025.
DepMap has gaps; Turbine's gaps are different
DepMap and the Open Targets Platform are foundational tools in oncology research. Since 2018, DepMap has tested gene knockouts in over 900 cancer cell lines using CRISPR-Cas9 and RNA interference, making the data freely available. Both platforms are widely used for target identification and biomarker discovery.
Both platforms have documented limitations. They lack mechanistic insight into why a gene is essential. Neither covers complex perturbations or synthetic lethal pairs—dependencies that only emerge in specific molecular contexts, like the BRCA/PARP example. As tumor biology understanding deepens, those gaps matter more: 60% of PARP inhibitor-treated patients eventually develop resistance through restoration of homologous recombination repair, suggesting simple gene essentiality is insufficient for predicting drug response.
Turbine's approach trains on richer datasets and uses machine learning to predict not just which genes are essential, but which genetic combinations change drug response. That is a different—and narrower—claim than "finding all cancer dependencies." It succeeds on resistance prediction but does not claim to replace DepMap's scope.
Validate predictions before scaling
Two biomarkers in clinical trials is real validation, but it represents a small fraction of the 13 candidates the screen identified. The in vitro validation rate (8 of 13) does not guarantee clinical utility. The vivo validation of NEK1 is more encouraging, though the sample size and clinical translation remain unstated.
Precision medicine teams should treat virtual assay outputs as high-confidence hypotheses, not certainties. Turbine's strength is in ranking which combinations warrant bench testing, not eliminating bench testing. The economics only improve if in silico screening materially reduces failed experiments or accelerates time-to-lead, neither of which is quantified in the article.
For companies building internal oncology target pipelines, the question is whether integrating Turbine's model predictions into your workflow improves hit rates or reduces cycle time compared to current screening regimens. Two biomarkers in trials is promising. It is not yet proof that the platform systematically outperforms existing approaches across disease areas.