Translational Intelligence
Field guide · Updated July 20, 2026

The Day-One Target Safety Triage

How to use AI to pull a first-pass target safety read in a day. Not to replace the expert assessment, but to surface a disqualifying signal, for a human to verify, before you spend a scientist's month on a doomed target.

What a target safety assessment is, and why day one matters

Before a program commits real money to a target, someone has to ask the safety question the whole thing rides on. If we antagonise or attenuate this target, what breaks. A target safety assessment (TSA) answers it at target nomination, by pulling together the target’s biology, its gene and protein expression across tissues, human and animal genetic evidence, and what happened to competitors who touched it. Done well, it is expert, slow, and decisive: it confirms unavoidable on-target liabilities and surfaces the ones you could design around, so a program goes forward, pivots, or dies before it burns a preclinical budget.

The cost of getting it late is enormous, and the cost of getting a first read early is now almost nothing. That gap is the opportunity.

The one idea

AI can give you a first-pass target safety read on day one. Not the assessment. A first pass: the collection, the summarization, and the initial analysis, assembled in a day instead of the weeks it takes a human to gather it by hand.

The value is not that the machine writes your TSA. It is that a scientist’s judgment is the most expensive input in discovery, and you should not spend a month of it on a target that a day of automated collection would have flagged. If AI surfaces an embryonic-lethal knockout, a broad and essential expression profile, or a human genetic association with exactly the adverse phenotype you fear, you have a high-priority lead to verify before the expert deep-dive starts. You did not replace the assessment. You made sure the expensive one only ever runs on targets not yet ruled out.

The day-one target-safety triage, end to end From a stated target and perturbation hypothesis, AI pulls the five public evidence signals and assembles a first-pass synthesis in a day. A disqualifying signal is verified at the source and can stop the program cheaply. If nothing is found, a human safety scientist begins the expert assessment, unproven. Target + hypothesis what you plan to do Pull five signals the public record First-pass synthesis assembled in a day signal Verify, then stop cheaply, on evidence nothing The expert TSA begins unproven, not cleared
The day-one pass collects and synthesizes; a human verifies, interprets, and decides. It ends where the expert assessment begins.

State the perturbation hypothesis first

“Is this target safe” is not yet a question a machine, or a person, can answer. Safe how, and against what intervention? A target that is dangerous to knock out constitutively in an embryo may be perfectly tractable to inhibit partially, reversibly, in one adult tissue, for a fixed course. Before you run the pass, write down the perturbation you actually intend, because every signal below has to be read against it:

  • Direction. Are you inhibiting, activating, degrading, or replacing the target?
  • Modality. Small molecule, antibody, oligonucleotide, gene therapy? Each engages the target differently, and the genetics map onto each differently.
  • Magnitude and reversibility. Full or partial knockdown, and does it reverse when the drug stops, or is it permanent?
  • Tissue and exposure. Systemic, or restricted to the diseased tissue? Where will the drug actually reach the target?
  • Population and duration. Who is dosed, how sick, and for how long? An eight-week course in late-stage disease and a lifelong prophylactic carry different safety bars.
  • Developmental versus adult biology. Most knockout evidence is constitutive and developmental. Your drug almost never is.

You are not answering these yet. You are stating them, so that when the pass returns a signal, you can ask the only question that matters: does this signal apply to what I actually plan to do?

What the day-one pass collects

The raw material of a TSA already lives in public, machine-readable knowledge bases. The work AI removes is the retrieval and the reading, not the interpreting. A day-one pass pulls, for the target:

  • Expression breadth. Where is it expressed, and how broadly, from GTEx and the Expression Atlas. A target restricted to the diseased tissue is a very different safety story than one expressed everywhere and essentially.
  • Loss-of-function tolerance. How constrained is the gene in humans, from gnomAD (LOEUF and pLI). A gene humans cannot tolerate losing is a warning about attenuating it with a drug.
  • Knockout phenotypes. What happens when you delete it in a mouse, from the International Mouse Phenotyping Consortium. Lethal or severe phenotypes are the loudest early signal there is.
  • Human genetic associations. Do variants that reduce the target’s function associate with adverse phenotypes, or with the protection you are hoping for. Genetic support cuts both ways, and it is among the strongest early predictors available: mechanisms with human genetic support are more than twice as likely to reach approval, about 2.6-fold in the refined estimate (Nelson et al., Nature Genetics, 2015; refined by King et al., PLOS Genetics, 2019 and Minikel et al., Nature, 2024), with the effect varying by therapy area, development phase, and confidence in the causal gene.
  • Precedence and prior liabilities. Known drugs against the target, known on-target toxicities, and competitor programs that failed on safety.

Most of this is already integrated and scored in one place, the Open Targets Platform, which aggregates genetics, expression, mouse phenotypes, tractability, and safety into a single target view. Buy the record: the databases are the shared infrastructure, and you win nothing by rebuilding them. Build the intelligence: the day-one pass is your prompt, your retrieval-and-verify workflow, and your target-specific judgment on top.

Read every signal in context

Every signal above is a lead, not a verdict, and each one is context-dependent in a way the machine will not caveat for you. Read each against the perturbation hypothesis you wrote down.

  • A lethal knockout is not automatically a dead target. Most knockout evidence is constitutive and developmental. An embryonic-lethal gene deletion says little, on its own, about inhibiting the same target partially and reversibly in an adult tissue for a fixed course. It is a reason to look hard, not a reason to stop by itself.
  • Broad RNA expression is not broad risk. Transcript abundance across tissues is not the same as protein abundance, target engagement, or toxicity. Where the drug actually reaches the target matters more than where the message is transcribed.
  • Loss-of-function constraint depends on your modality. A high pLI or a low LOEUF warns against removing the protein fully and permanently. It reads differently for a partial, reversible inhibitor than for a knockdown or a gene therapy, and differently again for agonism versus antagonism.
  • Genetic support raises the odds; it does not prove safety. It is the most predictive early signal you have, but it speaks to probability of success, not to a clean safety bill, and its size varies by therapy area, by phase, and by how confident you are in the causal gene. Powerful evidence, not a context-free decision rule.

The pattern is the same every time: the pass tells you where to look; your hypothesis and a human tell you what it means.

What it flags

The day-one pass is not scored to bless a target. It is scored to surface a reason to stop, cheaply. It reads the collected evidence for the disqualifying signals: broad and essential expression, a lethal or severe knockout phenotype, strong human loss-of-function constraint, a genetic association pointing at the toxicity you are worried about, or a graveyard of competitor programs that died on the same on-target liability. Any one of those, surfaced on day one, is worth more than a beautiful summary, because it changes what you do next.

What it does not do

This is the line, and in a safety context it is not negotiable.

The day-one pass is not the target safety assessment, and it is never the decision. A human safety scientist is the author of record and owns every claim in the final TSA, exactly as there is no AI work product, only AI-assisted human work product. A confident, fluent, wrong summary of a knockout phenotype is not a small error here; it is a safety liability, so every extracted fact is verified against its source before it informs anything. The AI cleared the collection so the scientist could spend their scarce attention on interpretation, which is the whole point of not automating the struggle: you reach the hard judgment sooner, you do not skip it.

And the evidence has limits the machine will not caveat for you. Genetic support raises the odds; it does not guarantee safety or success. Absence of a mouse phenotype is not absence of human risk. A first pass built from public databases is exactly that, a first pass, and it does not replace the expert assessment, the bespoke de-risking assays, or the nonclinical safety package that regulators will expect. Treat the day-one read as triage, and confirm the safety strategy with your own toxicology and regulatory functions.

Monday morning

Take the next target on your list and run the day-one pass before you schedule the expert TSA. Pull its expression breadth, its constraint, its knockout phenotype, its genetic associations, and its precedent, have AI assemble the first-pass memo, and then verify every line against the source. If a disqualifying signal is sitting there, you found it in a day and saved a month. If nothing is, your safety scientist starts their real assessment already past the grind, on a target that still requires the full assessment.

Sources and further reading