UbiquiTx
Editing the proteome
to unlock biology
An AI-native platform for scaled parallel protein perturbation that learns from every experiment to accelerate drug discovery.
Problem 1 of 2
Knockout answers one question. Rarely the one that matters.
You need
The clinical mechanism
Degrade, stabilize or relocalize — not a gene-knockout stand-in.
You need
An in-cell readout
Pathway response, directly linked to the modulation.
You need
Protein-state confidence
Which state you moved, and that it was the right target.
~52%of Phase 2 failures are lack of efficacy — the target hypothesis is where that beginsCook et al., Nat Rev Drug Discov, 2014
$200–500Mto carry one wrong target that far
Problem 2 of 2
AI discovery models are limited by two data gaps.
Peptide binder models
High hit rates. Low functional yield.
Trained on
Public binding databases — positives only
Missing
Failed binders. Whether binding did anything.
Proteome data
Blind to the layer that governs behaviour.
Trained on
Transcriptomics, bulk abundance, PPI maps
Missing
Modification state. What it does downstream.
The construct
CLIP — a Chimeric Ligand for Induced Proximity.
An AI-designed peptideEngages flat surfaces on the native protein — no binding pocket required
Genetically encoded~200 amino acids, written as DNA or mRNA — one gene per construct
So it can be pooledThe gene is its own barcode: thousands screened in one dish, read out by sequencing
What it does
Swap the effector, swap the outcome.
Degrade
Remove the target
Route the protein to the cell's own disposal machinery.
Stabilize
Protect the target
Hold a protein that disease is otherwise destroying.
Relocalize
Move the target
Change where in the cell the protein does its work.
The same guide peptide drives opposite pathway outcomes by effector class.
The discovery engine
Two coupled screens. Every result trains the next design.
What it answers
Target hypothesis confirmation: is this target worth a chemistry campaign?
Question 1
Did we engage the target — and only the target?
Binding and selectivity, measured together.
Question 2
Did the protein move, and how far?
Degraded or stabilized, on the native protein.
Question 3
Did the pathway respond, and which way?
The phenotypic answer the hypothesis turns on.
Three answers, on every construct — not a hit list.
Screen one · multiplex
~2,000 guide peptides against each of ~10 targets, at once.
variants, one screen0
Every guide is built into a complete CLIP with a fixed effector and barcoded, so one pooled screen covers a whole target panel.
Sort on the reporter, sequence the barcodes.
What comes out is a functional hit — not a binder. The guide bound, and the effector it carries changed the target.
Screen two · combinatorial, optional
Then ask what a different modification would do.
Take
The winning guides
Binders already proven functional in screen one.
Pair
~100 effectors
Ligases, deubiquitinases, kinases, phosphatases — degrade, stabilize, phosphorylate, dephosphorylate.
Read
One target, many directions
Which modification moves the pathway, and which way — on the same native protein.
Linker length is screened alongside, tuning the geometry between guide and effector.
What you receive
A decision, a construct, and the data behind both.
01
A decision
Go or no-go on the target, before a multi-year chemistry campaign is funded.
02
A validated construct
The CLIP that answered the question — a therapeutic starting point, not a tool compound.
03
The full dataset
Every variant measured on all three axes, failures included.
~4–6 months from target to answer.
How we compare
Matched to the drug you intend to build.
| CLIPs | CRISPR / siRNA | Tool compounds | |
|---|---|---|---|
| Modality | Degrade · stabilize · relocalize | Loss of function only | Pocket-dependent |
| Resolution | PTM-resolved, in cell | Gene or RNA level | Indirect |
| What you end with | A validated construct | A phenotype, then a multi-year campaign | A probe, if one exists |
Knockdown reaches a phenotype sooner; CLIPs return a modality-matched answer and the construct that produced it.
What compounds
The data is the asset.
Peptide design
A design model trained on our own screens
Binders and non-binders, functional outcomes and failures — the results public databases never hold. Each campaign starts ahead of the last.
Virtual cell models
Functional perturbation data
What actually happens to a pathway when one protein is moved — the input virtual cell models lack and cannot generate.
Business model
One platform. Three monetization engines.
Engine 1
Target hypothesis confirmation
Target and modality confirmation for partner pipelines.
Engine 2
CLIP library + data asset
Functional, PTM-resolved outcomes — including the negatives no public database holds.
Engine 3
Therapeutics
Lead CLIPs advance as biologic assets against targets we have already confirmed.
Each engine feeds the next: every campaign deepens the data asset, and the data asset sharpens the next campaign.
Here, and where we go next
The methodology works. What remains is scale.
Where we are
Multiplex screen running end to end
Pooled, barcoded, sorted, deconvoluted.
A lead therapeutic candidate
A degrader CLIP against cytosolic β-catenin.
NIH and NSF awards
Non-dilutive funding behind the platform.
Where we are going
Scale the screen
Full target panel, combinatorial layer on top.
Demonstrate the data loop
Retrain on our own screens, measure the gain.
First pharma partnerships
Engagements on partner-nominated targets.
Raising a $10M seed round
Ask the proteome
a better question.
Target hypothesis confirmation on the native protein, with the mechanism you intend to drug — and the construct that answered it.
mathew@ubiquitx.com
·
ubiquitx.com
Full validation packet available under NDA.