Data is not the bottleneck. Interpretation is.

What you get

• An analysis report with publication-quality figures
• A reproducible pipeline with the full parameter record - you can run it again without us
• The processed data package in standard formats
• An interpretation session with your research team, because a PDF is not a conclusion

The AI capability

• A point-in-time feature store

  One curated, versioned source of target, asset and disease features that every model reads from, instead of each analysis re-deriving the same numbers slightly differently. Reads are point-in-time: a training pull can never see a value that post-dates its label. There is an explicit leakage test in the codebase, and it runs as part of validation rather than living in a doc as an intention.

• Knowledge-graph integration

  node2vec embeddings over the biomedical graph with a supervised edge classifier - Hadamard product of node embeddings plus common neighbours, Adamic–Adar, Jaccard, preferential attachment and random-walk-with-restart. Positives are held out before the embedding is trained, so the AUC we report is not the flattering number you get when the embedding has already walked the test edges.

• Semantic retrieval

  Across the integrated entity corpus, so a question phrased in your words finds the record filed under someone else’s.

What you get

• Validated biomarker panels with statistical performance packages - not candidate lists
• Per-marker evidence with the quoted sentence and character offset into the source paper
• An assay development route for the markers that survive
• Independent-cohort validation design
• Optional: full wet-lab validation and assay build through our laboratory partner

The AI capability

• Span-anchored evidence mining

  Our biomarker layer runs on 421 span-anchored evidence records carrying the biomarker_for predicate, across 184 targets and 86 diseases - every one with a quoted sentence and a character offset into a real paper. Not a pre-computed score you have to take on faith. The evidence itself.

• A closed 34-predicate relation schema

  Open-ended "extract any relationship" prompting produces a graph with thousands of near-synonymous edge types - inhibits, suppresses, blocks, downregulates, attenuates - that no query can traverse. Every extracted trigger maps onto one canonical predicate, and type signatures are enforced: a drug–disease pair cannot produce inhibits, because inhibits has domain asset and range target. That constraint alone removes a large class of nonsense.

• Assertion classification

  "PRG-1042 inhibited SHP2" and "PRG-1042 did not inhibit SHP2" differ by one token and by 180 degrees. A pipeline that ignores negation builds a graph that confidently asserts the opposite of the literature, and nothing downstream can detect it because the graph looks perfectly well-formed. We run a NegEx/ConText-style scope classifier over affirmed / negated / speculated / hypothetical, deliberately rule-based so a domain expert can inspect and debug it.

• Polarity is first-class

  Twelve papers may say X marks Y and three may say it does not. A scalar confidence value has destroyed that disagreement - and the disagreement is exactly what you need to see. We keep all fifteen records, and refuting evidence is retained as evidence.

What you get

• A variant interpretation workflow with defined turnaround times
• Per-variant evidence assembly: population frequency, clinical classification, predicted consequence, protein change and associated condition
• Gene and panel-level annotation packages
• Pipeline deployment into your environment with full parameter records
• Training for the team who will run it after we leave

The AI capability

• Variant evidence assembly

  ClinVar clinical classification with the review-status star rating, so a one-star single submitter is never presented as equivalent to a reviewed-by-expert-panel call - plus the associated condition, protein change and molecular consequence. gnomAD supplies dbSNP rsID, VEP consequence, HGVSp and combined exome + genome allele counts and frequency. Conditions are ontology-resolved through MONDO and EFO.

• Formal two-sample Mendelian randomisation

  Where the question is causal rather than descriptive: cis-eQTL instruments as exposure against disease GWAS as outcome, estimated by Wald ratio, IVW, MR-Egger and weighted median, with Cochran’s Q for heterogeneity and the Egger intercept for directional pleiotropy.

• Grounded natural-language querying

  Ask a question, get an answer with citations into the underlying records. Graph queries run from parameterised templates, never from model-written query strings - untrusted text in an ingested document must never become execution against the knowledge graph.

• Full reproducibility

  Every generated artefact records the corpus snapshot it was computed from. Evidence is append-only in spirit: re-extraction at a new model version writes new rows and marks the old ones superseded, so any past conclusion can be reconstructed exactly.