Methodology
How we build a defensible target ranking
The evidence tiers, scoring model, and validation approach behind every Assaygrove output.
Evidence Architecture
Three evidence tiers, ordered by causal strength
Not all evidence for a target is equal. Genetic associations carry a different causal weight than functional phenotypes, which carry a different weight than literature co-mention frequency. Our scoring reflects that hierarchy.
Tier 1: Genetic Evidence
Genome-wide association study hits, rare variant burden analyses, and expression quantitative trait loci (eQTLs) constitute the strongest prior for target-disease causal relevance. A target with a significant GWAS association in the relevant disease context starts from a fundamentally different evidential position than one without.
Tier 2: Functional Evidence
Experimental knockout and knockdown phenotype data, protein-protein interaction network evidence, and pathway enrichment analyses inform the mechanistic plausibility of a target. A target with Tier 1 genetic support and concordant functional data is significantly more compelling than one with genetic support alone.
Tier 3: Literature Evidence
Publication frequency, citation recency weighting, and preprint signals collectively reflect the research community's confidence in a target. This tier is weighted lower than Tiers 1 and 2 because it is subject to study bias: well-studied targets accrue more literature, which can create a self-reinforcing signal regardless of true causal relevance.
Omics Integration
Handling real-world data: incomplete, batch-affected, multi-modal
Most target ID projects do not arrive with clean, complete multi-omics datasets. The system is built for the reality of what research teams actually have.
Normalization and batch correction
Expression data from different sample preparation batches are batch-corrected using ComBat-seq-equivalent normalization before cross-modal integration. RNA-seq data can be provided as raw counts or pre-normalized TPM/FPKM; proteomics abundance as LFQ or iBAQ values.
Sparse input handling
If a dataset covers two of four supported modalities (e.g. bulk RNA-seq plus GWAS, but no proteomics and no scRNA-seq), the model scores available evidence and reports the modalities that are absent. Missing modalities do not inflate confidence scores.
Cross-modal harmonization
Gene expression, protein abundance, and GWAS signals are mapped to a common gene-level target space using Ensembl identifiers with alias resolution. Conflicting signals between modalities (e.g. high transcriptomic but low proteomic expression) are flagged and reported as a confidence modifier.
Single-cell integration
scRNA-seq datasets are processed at the cell-type level. Cell-type-specific differential expression is computed and summarized into a target-level signal reflecting disease-relevant cell-type enrichment. Useful for identifying targets specific to a tissue compartment rather than bulk expression artifacts.
Validation Approach
What we check our work against
We benchmark retrospectively. We are honest about what early-stage validation means for a company founded in 2024.
Retrospective Phase 1/2 benchmarking
We apply our scoring model to targets that subsequently entered Phase 1 or 2 trials (publicly available from ClinicalTrials.gov outcomes data) and examine whether our pre-trial composite score was concordant with the trial outcome direction. This is our primary internal benchmark.
Cross-program early-access validation
Through our early-access program (Q3-Q4 2025), research teams from 8 early-stage programs used Assaygrove rankings alongside their internal analyses. In 6 of 8 cases, teams adjusted their lead target prioritization based on our output. We track these reprioritization decisions as prospective validation data.
Known limitations we report explicitly
Study bias in protein interaction networks means well-studied targets accrue more functional evidence independent of true causal relevance. Rare disease indications with limited GWAS data have sparser Tier 1 evidence. The scoring model does not assess druggability, patent landscape, or chemistry tractability.
Versioned scoring model
Each analysis reports the scoring model version used. If you re-run an analysis after a model update, both outputs are preserved, and the change log describes what changed in the evidence weights or input data sources. Your historical rankings remain auditable.