Platform

Two engines.
One experimental learning loop.

KritRNA is developing a coupled computational platform: one engine designs and ranks suppressor-tRNA candidates; the second models how each candidate perturbs the wider translation system. Selected candidates must then be tested experimentally.

Understand the core biology →Explore Engine 02
Dual-engine architecture

Design the molecule. Model the system.

The two engines are intended to exchange constraints. A sequence that looks viable in isolation may still fail when termination competition, transcript surveillance or cellular stress are considered.

Engine 01 · Suppressor-tRNA design

From target mutation to testable shortlist

Candidate generation, biological filtering, structure and context evaluation, then explainable multi-objective ranking.

01 · Input definition
Gene, nonsense variant, stop codon, local sequence context and intended amino-acid restoration.
02 · Candidate generation
Explore de-novo and mutation-based suppressor-tRNA candidates rather than relying on a single anticodon edit.
03 · Biological constraint gates
Processing, folding, aminoacylation identity, structural viability and normal-stop risk.
04 · Context and kinetic modelling
Termination competition, codon neighbourhood, NMD context, ribosome dwell and cellular state.
05 · Explainable ranking
Multi-objective scoring and Pareto selection to identify candidates worth experimental testing.
Engine 02 · Translation small-world

From candidate molecule to system consequence

A mechanistic network intended to connect translation dynamics, surveillance, stress signalling and protein output.

01 · Translation network
Represent initiation, elongation, collision, termination, recycling and quality-control relationships.
02 · Perturbation model
Introduce a candidate suppressor tRNA as a network intervention rather than an isolated molecule.
03 · Systems outcomes
Estimate consequences for readthrough, NMD, ribosome traffic, ISR signalling and protein output.
04 · Feedback to design
Return system-level penalties and opportunities to the candidate-ranking engine.
Defensible platform

The value is in the integration

KritRNA’s intended defensibility is not one public algorithm or one database. It is the evolving combination of biological constraints, candidate-generation logic, assay-normalised evidence, mechanistic models and wet-lab feedback.

Design grammar

Rules defining viable search space without disclosing exact generation logic.

Constraint stack

Coupled sequence, structure, charging, context, kinetics and safety filters.

Evidence memory

Versioned candidate and assay records, including failures and uncertainty.

Closed-loop learning

Experimental outcomes progressively reshape prioritisation and calibration.

Public pages describe the architecture and scientific rationale. Proprietary model weights, feature transformations, ranking functions, training corpora and candidate sequences are not published.
Kinetic and translation features

A candidate enters a race at the ribosome

Readthrough depends on timing and context. The model therefore needs more than a static tRNA structure score.

Release-factor competition

Model competition between suppressor-tRNA accommodation and eRF1/eRF3-mediated termination.

Ribosome dwell

Estimate how sequence context and decoding behaviour affect residence time at the premature stop.

Local codon context

Consider the stop codon, the +4 nucleotide and surrounding sequence features that influence termination.

Aminoacylation

Assess whether a design is likely to preserve productive recognition by its intended aminoacyl-tRNA synthetase.

NMD and EJC geometry

Represent transcript context that can influence nonsense-mediated decay and available message abundance.

Stress-state effects

Account for cellular stress, tRNA modification state, ribosome collisions and integrated-stress-response risk.

Machine-learning layer

Learn from biology without flattening it

Readthrough measurements from different assays are related but not automatically commensurable. The platform is designed to preserve experimental context rather than pooling every number into one misleading target.

01

Assay-aware learning

Keep experimental readouts tied to assay type, cell context, expression level and measurement method instead of treating every percentage as equivalent.

02

Structured biological features

Combine sequence, structure, thermodynamic, evolutionary, kinetic and transcript-context features.

03

Explainability

Use interpretable feature contributions so a shortlist can be challenged scientifically before it reaches the bench.

04

Active learning

Use each wet-lab round to prioritise the next experiments with the highest expected information value.

05

Uncertainty

Separate model confidence from biological promise and avoid presenting a score as evidence of therapeutic efficacy.

06

Evidence memory

Preserve candidate history, assay provenance and versioned outcomes so the platform learns without erasing failed designs.

Design–test–learn

The bench is the evidence gate

01
Design
Generate a biologically constrained candidate set and document why each candidate survives the filters.
02
Test
Measure readthrough, protein restoration, amino-acid identity, normal-stop effects and relevant cellular responses.
03
Learn
Return assay-labelled results to the platform, update uncertainty and select the next experiments.
Ribosome structure representing the translation system

KritRNA is at an early, preclinical research stage. These pages describe a platform under active development, not a validated therapeutic product. Experimental validation remains the next major evidence gate.