One cancer mutation.
One pocket it created.
Oncivra turns the TP53 Y163C mutant structure, its surface lesion and internal scoring models into a diverse shortlist of testable molecules — designed to hold the mutant protein in its working shape, and to leave the normal protein alone.
1.65 Å
High-resolution mutant crystal structure
4
Engines running end to end
0
Clinical programmes on this mutation
Pipeline
From one target to a small set of testable molecules.
Six molecules have been computationally prioritised for experimental validation against TP53 Y163C. The candidates were selected through the current Oncivra discovery workflow.
Final candidates
Six computationally prioritised molecules carried forward for experimental validation.
Current position: Prioritise complete · Lab not yet started · Validated hit not yet started
Loading the current shortlist…
The current campaign has completed computational prioritisation against the Y163C target structure. Potency estimates are withheld from this public view; they are visible inside the platform with full provenance.
No compound on the Oncivra platform currently has a measured binding value against TP53 Y163C. All experimental testing is still pending. None is a validated binder, lead, drug or clinical programme.
Standards of accuracy
Honesty over manufactured confidence.
Programmes fail at the bench when a system is allowed to mark its own work. Oncivra runs on a discipline of strict integrity, enforced in the record rather than in the copy.
No self-grading
Internal estimates are never allowed to tune the ranking that produced them. Only measured bench outcomes recalibrate scoring.
Sealed before testing
Every hypothesis — predicted binding mode, selectivity profile and developability parameters — is cryptographically sealed before any experiment begins. The goalposts cannot move retrospectively.
Mechanisms kept apart
Covalent and non-covalent chemistry follow fundamentally different physics. They are evaluated and ranked in separate pipelines and never compared against one another.
Missing is missing
A value that was never established is reported as not established. Nothing is filled in with a placeholder, and no figure is presented with more precision than its source supports.
High-resolution structural grounding
The pocket is not inferred. The campaign is built directly on the experimentally deposited 1.65 Å crystal structure of TP53 Y163C, mapping the exact cavity the somatic mutation creates.
Rigorous negative selection
Most of the work is rejection. Every molecule must survive layered selectivity barriers — pan-reactive chemistry filtered out, counter-screens against wild-type p53 and related variants, and stringent synthetic accessibility limits — before it reaches the shortlist.
Audit-ready provenance and complete assay specs
Each shortlisted candidate is delivered with complete derivation tracking, transparent confidence bounds, and fully specified triage protocols — protein constructs, buffer conditions, incubation timecourses and intact-mass counter-screens — ready for immediate synthesis and validation.
1.65 Å
Deposited crystallographic resolution
100% sealed
Hypotheses and derivations locked prior to testing
Orthogonal gating
Wild-type and variant counter-screened
6 priority leads
Specified for direct biochemical triage
Predicted flags, before the bench.
Between docking and the shortlist, every candidate undergoes computational developability assessment. Predictions come from a reproducible notebook outside the platform and are imported with validation controls. They inform triage. They never auto-kill a candidate, never enter the measured assay panel, and never touch a prediction-lock hash. Kill is on measured properties only.
What the engine does today
Built end to end, honestly accounted.
Structure-based design
Generative chemistry from a 1.65 Å crystal structure of the exact mutant, not a homology model. Pocket residues computed from deposited coordinates.
Hosted generative models
One model builds new molecules from fragments; another optimises existing hits. Both run as hosted models, with confirmation before every run.
Docking and affinity
Hosted models generate poses and predict affinity. Each estimate is the median of repeated runs, never a single pass. Covalent and non-covalent candidates are ranked separately and never compared.
Predicted ADMET screen
Eight advisory flags per candidate, validated against known-drug controls. Advisory only — never a kill trigger.
Structural verification
A structural check compares predicted poses against the known DNA-contact geometry. It is display-only computational evidence and never enters the learning weights.
Pre-registered kill criteria
Kill criteria are intended to be pre-registered before testing. None are registered yet for the current shortlist.
Learn engine
Measured assays alone recalibrate the scoring weights. Model estimates never grade themselves. Until the bench speaks, weights hold at defaults.
Selectivity, not assumed
Selectivity is assessed computationally and remains subject to experimental validation.
Experiment handoff
A diverse shortlist leaves the platform with full provenance: structures, scores, ADMET summary, and the package a laboratory needs to test it.
A second starting point
Alongside newly designed molecules, the platform also evaluates compounds already approved for other uses as possible starting points — checked against public records, tested in the same way, and never given special credit for their history.
Independent evidence streams
Large public cell studies can be imported and looked up live, kept strictly separate from the platform's own predictions. Observing a response in a cell is never counted as proof a molecule bound its target.
Failures that teach
When a candidate fails, the platform records exactly why — in plain structural terms — and turns that into written constraints for the next design round. Each new round stays linked to the failure that prompted it, so the reasoning is always auditable.
The target
Why this mutation.
The platform scores the target from the evidence it actually holds — structures, pocket residues, pocket druggability, citations and known chemical matter — and reports the terms that score badly alongside the ones that score well.
The score is computed live from the evidence records, not fixed here. Full scorecard inside the platform.
A pocket the mutation made
The Y163C change alters the surface of the p53 DNA-binding domain, creating a cavity that the normal protein does not have. That difference is the whole basis for selectivity.
A real structure, not a model
A 1.65 Å crystal structure of this exact mutant exists in the public record. The pocket residues used here are computed from those deposited coordinates.
A handle for chemistry
The altered surface gives chemistry a distinct feature to target. Whether that feature is actually usable is carried as a hypothesis, labelled as such — not assumed.
Selectivity is not automatic
The target protein carries other features that a reactive molecule could hit instead of the intended one. Selectivity has to come from fitting the pocket precisely, not from the reactive chemistry alone. We treat this as an unsolved medicinal chemistry problem, not a solved one.
An open field
The neighbouring Y220C mutation already has a clinical-stage stabiliser, which proves this class of pocket is druggable. Nothing is in the clinic for Y163C.
Platform
Four engines, running together.
The discovery funnel runs target → design → verify → learn. Three engines are complete for this campaign; the fourth waits on the bench.
Find
Which target deserves the campaign. Genomic and variant reasoning, human evidence, druggability, competition.
Design
What could attack it. Parallel generative campaigns around the chosen pocket, filtered for real chemistry.
Verify
Which molecules our internal scoring models agree on. Pose generation, affinity prediction, selectivity, developability.
Learn
What the bench taught us. Measured results recalibrate the models and steer the next generation.
Evidence
Everything traces back to a source.
Structures, assays, compounds and clinical programmes are drawn from public scientific records, and every candidate carries the models that supported it.
- Mutant structure
Public mutant crystal structure — p53 Y163C, 1.65 Å
- Selectivity
Y220C counter-target structures
- Structure
Wild-type reference structures
- Sequence
UniProt P04637
- Mutation frequency
cBioPortal, GDC, OncoKB
- Target and chemistry
Open Targets, ChEMBL, PubChem
About
A discovery engine, honestly described.
Oncivra uses generative and structural models to design and prioritise small molecules against a pocket created by a recurrent cancer mutation, for experimental testing.
It does not claim a finished medicine. The campaign ends with a small, diverse set of candidates and the package a laboratory needs to test them.