What validation evidence looks like when graded
Target validation is a chain of claims of very different strengths: an epidemiological association, a knockdown rescue in a disease model, a measured binding thermodynamic. Treating them as interchangeable is how programs talk themselves into weak targets.
Atlas Bio's published evidence hierarchy separates them explicitly. L1 is correlational or observational — surveyed but never headlined. L2 is causal, meaning the effect is demonstrated in a model, and is described as the working evidence layer for recommendations. L3 is thermodynamic or first-principles and carries the strongest publishable assertions. The level travels with the claim.
Why genetic support deserves its own weight
Human genetics is the one strand with a measured effect on program outcomes. Nelson et al. (Nature Genetics 2015) found the proportion of drug mechanisms with direct genetic support rises from 2.0% at the preclinical stage to 8.2% among approved mechanisms, and estimated that selecting genetically supported targets could double clinical success.
The decade-later refinement is more precise. Minikel et al. (Nature 2024) estimate that 'the probability of success for drug mechanisms with genetic support is 2.6 times greater than those without', and report that this relative success improves with increasing confidence in the causal gene but is largely unaffected by genetic effect size, minor allele frequency or year of discovery. The implication for an automated pipeline is direct: variant-to-gene confidence must be modelled, not assumed.
What the agents add that a literature review does not
The gain is not reading speed, it is adversarial coverage. Atlas Bio runs every output through a reasoning audit whose nodes each attack a different failure mode: causal-chain integrity, internal contradiction, calibration of stated confidence, mechanistic plausibility, misapplied analogy, best-explanation reasoning against competing hypotheses, time-order violations, and a meta node that arbitrates when the others disagree and surfaces the disagreement to a human reviewer.
In practice the catches are mundane and expensive. A causal claim the underlying data cannot support. A conclusion that contradicts a reference cited three sections earlier. A mechanism asserted with no plausible route. These are the errors that survive a human skim and then anchor a two-year program.
- Enumerate competing hypotheses rather than defending the first one
- Separate association from demonstrated mechanism in the write-up
- Record which references drove each conclusion, so a reviewer can disagree with a specific one
- Surface node disagreement to the human reviewer instead of averaging it away
Where the human decision stays
An agentic pipeline can tell you that a target's support is mostly L1, that the causal gene assignment is uncertain, or that the mechanism rests on an analogy to a different tissue. It cannot tell you whether to spend the next two years on it.
Atlas Bio keeps that boundary formal: hypothesis prioritisation, compound selection and ejection, kill-criteria enforcement and the translational readiness gate are named non-delegable decisions held by a lead scientist, and every advance between evidence tiers requires a signed human review.
Atlas Bio grades every claim L1, L2 or L3 and runs each output through an eight-node IBC reasoning audit covering causal, contradiction, confidence, mechanistic, analogical, abductive, temporal and meta-reasoning checks before a human reviewer sees it.
Does genetic support guarantee a target will work?
No. Minikel et al. (2024) estimate success probability is 2.6 times greater with genetic support than without — a shift in odds across a portfolio, not a guarantee for any single mechanism, and it improves with confidence in the causal gene.
What stops the pipeline from confirming what we already believe?
Explicit enumeration of competing hypotheses, a dedicated contradiction check against earlier claims and cited references, and pre-registered kill criteria that a human is obliged to enforce when they trigger, regardless of sunk cost.
How fast is an agentic target review?
Fast enough that breadth stops being the constraint — the agents survey continuously. The pacing item becomes human review of flagged items, which is the part you want to keep slow and deliberate.
Talk to Atlas Bio.
Sponsors, researchers and clinicians: tell us the decision you are trying to make and we will show what the platform can and cannot answer.
Contact Us →