AI-assisted molecular interpretation
Support for reading molecular findings in context — annotation, tiering, and surfacing the evidence behind a classification rather than only its label.
AI & cancer intelligence
Building cancer intelligence
Artificial intelligence in oncology should not simply generate text. It should help clinicians interpret complexity, connect evidence and retrieve what is relevant — and show its reasoning while doing so.
The interesting question is not what these systems can produce. It is which parts of oncological reasoning are genuinely bottlenecked.
Four bottlenecks recur: the volume of evidence, the amount of clinically important information that exists only as prose, the manual work of connecting molecular findings to disease context and evidence, and the difficulty of identifying genuinely relevant trials.
None of these is a shortage of generated text. Each is a retrieval, extraction or integration problem — which points the work somewhere quite specific.
A system whose reasoning cannot be interrogated must be either accepted on faith or ignored. Neither is a good outcome, which is why showing the evidence behind an output matters more than the fluency of the output itself.
Architecture
01
Data
02
Molecular knowledge
03
Computation
04
AI
05
Evidence
06
Clinical intelligence
Each layer constrains the one above it. A model cannot recover information the data layer never captured.
One base changes, and the biology changes with it
Reference DNA
ATGGAGGATCTGGTGGGAGCCGGT
Reference protein
MetGluAspLeuValGlyAlaGly
Variant DNA
ATGGAGGATCTGGAGGGAGCCGGT
Variant protein
MetGluAspLeuGluGlyAlaGly
A single substitution alters one amino acid. Whether that matters clinically depends on where it sits, what the protein does, and what the evidence says — which is the whole of the interpretive problem.
Categories
Research and technology areas
Eight areas where computational methods are applied to oncology. Each is chosen because a real bottleneck exists there, not because the technique is available.
Clinical natural language processing
Extracting structured, checkable information from pathology narrative, clinic letters and the biomedical literature, while preserving negation and uncertainty.
Patient–trial matching
Representing eligibility criteria in a form that can be reasoned over, so that trial search returns relevance rather than volume.
Biomedical knowledge systems
Connecting genes, variants, drugs, diseases and evidence into structures that can be queried and kept current as the literature moves.
Clinical decision support
Systems that make reasoning visible — evidence tiers, actionability categories and explicit statements of what remains uncertain.
Multi-omics modelling
Joint models across genomic, transcriptomic and other molecular layers, with attention to whether integration is adding signal or noise.
Genomic reasoning
Moving from variant lists to mechanistic argument: pathway context, functional consequence, and the strength of the inference at each step.
Longitudinal cancer intelligence
Treating a cancer case as a trajectory rather than a snapshot, so that molecular and clinical change over time is part of the analysis.
Project
ExplainMyMutations.AI
A project built around a specific failure: molecular reports that are technically complete and practically unreadable.
In development
Explaining molecular findings without flattening their uncertainty
Patients and clinicians increasingly hold molecular reports that are technically complete and difficult to interpret. The barrier is not access to the result; it is understanding what it means, how confident it is, and what question it raises next.
The design goal is explanation with its uncertainty intact — including saying plainly when the evidence behind a finding is weak or absent, which is precisely what a fluent summary tends to lose.
Project details will be published here as the work develops.
OnKommon
Technology and computational intelligence ecosystem
OnKommon is the technology and computational intelligence ecosystem. It is a separate entity from KPCIRC, which is the clinical and institutional ecosystem, and the two are not interchangeable.
Technology, AI/SaaS and molecular data-platform activities are maintained as separate institutional and technology functions, distinct from clinical services.