Why don't preclinical results translate, and how do you de-risk before in vivo?
Over 90% of agents that work in mice fail in the clinic, how do I check a candidate against human data before committing an expensive in vivo study?
More than 90% of agents that look good in mice fail in the clinic, because model physiology and patient biology diverge. We used Hydra to interpret a candidate against human expression, genomics, and outcomes before any in vivo study, synthesising ~190 analyses overnight and turning a dead-end synthetic-lethal pair into a specific, context-dependent hypothesis.
Who this is for: Preclinical and translational scientists who want a candidate interpreted against human clinical and omics evidence before committing wet-lab or in vivo resources.
- preclinical agents that fail in the clinic
- >90%
- analyses synthesised into one summary report
- ~190
- what took a bioinformatician days, run in parallel
- days → overnight
preclinical agents that fail in the clinic
analyses synthesised into one summary report
what took a bioinformatician days, run in parallel
The translation gap between a model and a patient
More than 90% of agents that succeed preclinically fail in the clinic. Mouse physiology and target homology differ from humans, efficacy criteria differ (a response that advances an agent in mice can be progressive disease in a patient), and a candidate that looks clean on paper may carry hidden liabilities, pan-essential genes, impossible co-deletions, or no real correlation in patient data.
The risk is committing an expensive in vivo study on a finding that was never going to translate. De-risking it means reading a single candidate against human expression, genomics, and clinical outcomes first, slow, manual work that's easy to skip under timeline pressure.
What teams in this space search for
- Why don't mouse results translate to humans?
- How do I validate a target against human data before an in vivo study?
- Which preclinical model reflects my patient population?
How we solved it with Hydra
“Take this candidate, a two-gene knockout pair, a tumour-suppressor and a second gene, and tell me whether it could be synthetically lethal and in which cancers. Check expression in tumour versus healthy tissue, dependencies, and genomic context before we commit a wet-lab study, and tell me which patient populations would actually benefit. Synthesise everything into a report I can verify in the morning.”
What Hydra ran
Hydra planned the bioinformatics and ran it in parallel across hundreds of skills: Ensembl for genomic context, DepMap for dependencies, TCGA and cBioPortal for tumour-versus-normal expression. In orchestrated mode a single run produced ~190 results, which it then synthesised into a summary paper so the volume was readable.
It checked the things that actually decide translation: is the target differentially expressed in tumour versus healthy tissue, is the proposed mechanism even possible, and where does it carry toxicity risk.
What it found
On the naive question the pair was a dead end: querying Ensembl, the two genes sit on different chromosomes (no cis-codeletion); DepMap showed the second gene is pan-essential (broad inhibition would be systemically toxic); and there was no real correlation between the two in patient data.
But the same chain surfaced a better angle: one cancer type frequently shows monosomy of the relevant chromosome alongside overexpression of that gene, leaving a single copy, which is a plausible, tumour-selective inhibition target. A negative result became a specific, testable, context-dependent hypothesis.
What we learned
The point isn't only validation, it's triage and direction. Caught before an in vivo study, a dead-end candidate either dies cheaply or is reframed into the precise context where it could work.
What used to take a bioinformatician days now runs overnight, so the scientist comes back in the morning to verify and judge, not to execute. That's where human expertise adds the most value.
What you get
- A candidate interpreted against human expression, genomics, and outcomes, overnight
- Subtype- and population-specific framing for a preclinical signal
- Dead-ends caught early (pan-essentiality, impossible co-deletions, no patient correlation)
- A sharper, human-anchored hypothesis to take into the lab
Data sources used
- DepMap (dependency & essentiality)
- TCGA (tumour vs normal expression)
- cBioPortal & Ensembl (alterations, genomic context)
- ClinicalTrials.gov & published outcomes (clinical alignment)
Figures reflect analyses PharosBio ran on public datasets and public benchmarks. Named competitors, collaborators, and logos are withheld at this stage; the methods and results shown are real and repointable to your own target.
Sources & methods
- Preclinical-to-clinical attrition: Mak et al., Am J Transl Res 2014 (>90% failure)
- Dependency & expression: DepMap; TCGA; cBioPortal; Ensembl
Frequently asked questions
Does this replace in vivo experiments?
No, it triages them. The aim is to interpret a candidate against human evidence first, so the in vivo studies you do commit are built around findings most likely to translate, in the right disease context.
What does 'verify, not execute' mean?
A single orchestrated run can produce ~190 analyses overnight and synthesise them into a summary. The scientist's morning is spent checking and judging that work, not manually running queries, which is where human expertise matters most.
Run this analysis on your question
Hydra plans, executes, and validates, so you reach a defensible answer in hours, not weeks.