SynthyraHost-pathogen interaction prediction
Predict which host proteins a pathogen acts on.
Infection starts when a pathogen's proteins bind the host's. Atlas predicts those contacts for every protein in a pathogen against every protein in the host, and ranks the ones a laboratory should test first. A response team starts with a short list on day one.
What it does
A ranked list of where a pathogen touches its host.
Atlas predicts protein interactions. This service applies it across species, with the pathogen on one side and the host on the other.
01
Target and countermeasure mapping
Find the host proteins a pathogen protein is predicted to engage, and the cellular systems those proteins belong to, so countermeasure work starts where the contact is.
02
Threat characterization
Characterize a new pathogen by the host systems it is predicted to reach, before experimental work exists.
03
Experiment prioritization
Turn a proteome-scale question into a ranked list a laboratory can test, with functional annotation and pathway enrichment beside each protein.
How it works
Every pathogen protein against every host protein.
Atlas embeds each protein once, from its sequence alone. Scoring a pathogen proteome against a host proteome is then a single matrix operation, so proteome-scale inter-actomes come back in seconds instead of one pair at a time.
Reference proteomes for human and common model organisms are pre-embedded. A query is a pathogen proteome or a set of pathogen proteins, and the result is a ranked set of host partners with functional annotation and pathway enrichment, in the web app or through the API.
A prediction is a hypothesis to test. It narrows the search, and the experiment confirms it.
See it
A real prediction you can explore.
Pick a pathogen. Each view comes from a real run of the model: the human proteome as a map, the pathogen's proteins ranked by the partners they are predicted to reach, and the contacts drawn between them. Hover to read any prediction.
Seventeen SARS-CoV-2 proteins against the human proteome, from a model that never saw SARS-CoV-2 in training. ORF7b is predicted to reach the olfactory receptors. Spike is shown as a specificity control.
SARS-CoV-2 proteins
Ranked by predicted human partners
Loading the inter-actome.
Every point is a prediction from a real run. Hover to read it.
Why it is hard
Why cross-species prediction is hard.
A model can look accurate by learning the wrong thing, listed below. The test we hold ourselves to is whether the model transfers to species and pathogen families it never saw, under splits that remove those shortcuts.
If it does not transfer, we narrow the claim to the organisms where it holds and report the limit.
01
Negatives that reveal the species
If non-interacting pairs are drawn differently from interacting ones, a model can score well by recognizing the organism and never learn the interaction.
02
Hub proteins
A few proteins appear in a large share of known interactions. A model can rank by how often a protein is seen instead of what it does.
03
Homology between training and test
Near-identical sequences on both sides of a split let a model recall instead of generalize.
04
Pathogens the model never saw
The case that matters in a new outbreak is the one no training example resembles.
Where it stands
Cross-species scoring is available today through Atlas. We are extending it to more pathogen families, and each expansion waits for the result of the one before it.
Use of the platform is bound by our terms, which prohibit designing, optimizing, or testing agents intended to harm human health, agriculture, or the environment.
Working on a pathogen or a host question? Tell us which one.
A scientist replies within two working days.