Protein-Protein Interactions
Rank the interactions that actually matter
Tesorai turns noisy AP-MS and BioID data into a prioritized, high-confidence interactome — cutting through background to surface the biology worth following up.
A smarter way to rank interactions
AP-MS and proximity-labeling (BioID) experiments generate thousands of candidate interactions per experiment. Tesorai's ranking layer tells you which ones are real.
Source: pilot with a biotech partner (anonymized).
More signal to start with
Tesorai Search identifies about 8% more pulled-down proteins than FragPipe, giving every downstream ranking step more true signal to work with.
Doubling the known interactions you recover
In a pilot with a biotech partner, Tesorai Ranking improved per-bait interaction ranking (AUPRC) by 21% over the partner's prior SAINT/CompPASS approach — and combined with an orthogonal co-abundance signal, doubled the number of high-confidence known PPIs recovered in the top hits per bait.
+21% AUPRC over SAINT/CompPASS on DDA data
+11% AUPRC from an orthogonal co-abundance signal
2x more known PPIs recovered in the top-2-per-bait shortlist
Interaction ranking (AUPRC), indexed = 100
Source: pilot with a biotech partner (anonymized).
+21%
AUPRC lift on DDA
+20–27%
AUPRC lift on DIA
+10–15%
additional, DDA+DIA combined
Any acquisition mode
Gains hold up on DDA and DIA
The same ranking approach delivers a 20–27% AUPRC improvement over SAINT on DIA data, and combining DDA and DIA evidence into a single model adds a further 10–15% over either alone.
Rigor
Built with real generalization controls
Our sequence-based PPI predictor was stress-tested against a well-known pitfall: naive models can reach 93% AUROC just by memorizing which proteins appear often in interaction databases. After correcting for this with protein-level and temporal train/test splits, Tesorai's model reaches an honest ~76% AUROC — a number built to hold up under scrutiny.
1
Naive scoring hits 0.93 AUROC
by simply memorizing which proteins appear often in interaction databases
2
We removed the leakage
protein-level + temporal train/test splits, CD-HIT similarity filtering
3
Honest result: ~0.76 AUROC
on truly unseen proteins — a number built to hold up
Real Results
Sharper rankings, fewer false leads
+8%
more pulled-down proteins identified vs. FragPipe
+21%
AUPRC improvement in interaction ranking vs. a partner's prior SAINT/CompPASS approach
2x
known PPIs recovered in the top hits per bait, combining ranking + co-abundance signal
Results from a pilot engagement with a biotech partner (anonymized per agreement).
See what Tesorai can find in your interactome data
Talk to our team or try it on your own data.
.svg)
Insights leveraging every byte of raw data