Audit-grade pre-clinical due diligence & translational intelligence
Mixed-format ChEMBL comparisons and fluent AI summaries hide what was never measured. Mosaic audits the biological, chemical and clinical axes of any human target — mouse-versus-human potency compared like for like, tool-compound liabilities, resistance and synthetic-lethal leads — and states, for every count, the query behind it and whether it is complete, a floor, or missing. Every dossier passes a rule-based verifier before delivery.
| axis | count | state | query |
|---|---|---|---|
| papers | — | not addressable | "TTN"[Title/Abstract] — symbol not attributable |
| compounds | 4 | fetched on demand | ChEMBL activities · pChEMBL ≥ 5 |
| trials | ≥ 298 | truncated · floor | ClinicalTrials.gov |
| ppi | 620 | fetched on demand | STRING v12.0 |
| essentiality | 1,208 | fetched on demand | DepMap CRISPR gene effect |
| coessentiality | 61 | fetched on demand | DepMap · corr ≥ 0.20 |
| structure | — | fetch failed | AlphaFold |
| mouse_compounds | 0 | measured · zero | ChEMBL · mouse Ttn · pChEMBL ≥ 5 |
| admet | ≥ 0 | truncated · floor | ChEMBL ADME · 2 assessed compounds |
Panel A · The ChEMBL format trap
Ask ChEMBL for every compound measured on both the human and the mouse protein and the mouse often looks far weaker. Most of that gap is the assay, not the species: mouse values tend to come from cell-based assays, human values from purified protein. Mosaic compares only compounds measured cell-free with the same standard type on both species, and gives no verdict below 8 of them.
All formats mixed
up to 148×
apparent mouse-vs-human potency shift, 11 of 60 targets past 5×
Like for like
within ±0.43 pChEMBL
all 6 targets with ≥ 8 like-for-like compounds (BCL2 125, CSF1R 95, MAPK1 22, KDR 12, MAPK14 12, PARP1 12): no species gap
| Target | All formats mixed | Like for like | Mosaic verdict |
|---|---|---|---|
| BCL2L1 | 148× weaker on mouse · 79 compounds | Δ +2.75 · 1 compound | no verdict · fewer than 8 |
| ADAM17 | 90× weaker on mouse · 48 compounds | 0 compounds · no cell-free mouse data | no like-for-like data |
| FLT1 | 25× weaker on mouse · 3 compounds | Δ −1.75 · 1 compound | no verdict · fewer than 8 |
| SRC | 11× weaker on mouse · 4 compounds | Δ +0.00 · 2 compounds | no verdict · fewer than 8 |
| EGFR | 9.2× weaker on mouse · 23 compounds | Δ −1.01 · 2 compounds | no verdict · fewer than 8 |
| CSF1R | 8.5× weaker on mouse · 241 compounds | Δ +0.00 · 95 compounds | no gap |
| FGFR4 | 8.4× weaker on mouse · 72 compounds | 0 compounds · no cell-free mouse data | no like-for-like data |
| AXL | 7.2× weaker on mouse · 32 compounds | 0 compounds · no cell-free mouse data | no like-for-like data |
| ABL1 | 6.6× weaker on mouse · 8 compounds | Δ +0.47 · 2 compounds | no verdict · fewer than 8 |
| PDGFRB | 6.4× weaker on mouse · 8 compounds | Δ −1.01 · 5 compounds | no verdict · fewer than 8 |
| MAPK14 | 6.2× weaker on mouse · 44 compounds | Δ +0.01 · 12 compounds | no gap |
Δ = mouse pChEMBL − human pChEMBL, median over the compounds compared. Rows: targets where comparing every shared compound across all assay formats gives |delta pChEMBL| >= 0.7 over >= 3 compounds. Like for like = measured cell-free (BAO_0000357, no cell or tissue fields, no mutant protein) with the same standard type on both species; a verdict needs >= 8 compounds. Source: Mosaic Stage 2 tables (target_species_gap, target_tool_compound_liabilities, compound_liabilities), production; ChEMBL activities under the declared query (IC50/Ki/Kd/EC50, relation '=', pChEMBL >= 5). kg_version 327.
Panel B · True pre-clinical blind spots
40 of 60
targets have no cell-free mouse protein chemistry in ChEMBL: 21 have no mouse activity at all, 15 only cell-based mouse assays, and 4 a mouse target ChEMBL holds only inside a family or complex. A mouse efficacy study on these rests on compounds with no cell-free potency on the mouse protein recorded in ChEMBL (pChEMBL ≥ 5, IC50/Ki/Kd/EC50).
1,347 of 2,661 (51%)
of the most potent tool compounds assessed have recorded activity at pChEMBL ≥ 6 on no human protein except their own target. That is unprofiled, not selective — a floor, never a clean bill of health.
| # | Target | Red flags |
|---|---|---|
| 1 | DYRK1A |
|
| 2 | FLT3 |
|
| 3 | CA9 |
|
| 4 | FES |
|
| 5 | MAP2K2 |
|
| 6 | FLT4 |
|
| 7 | KIT |
|
| 8 | NTRK3 |
|
| 9 | CDK1 |
|
| 10 | CHEK1 |
|
One flag each: a like-for-like potency gap (|delta pChEMBL| >= 0.7 over >= 8 compounds measured cell-free with the same standard type); no mouse chemistry while the human target has >= 100 activities; mouse data cellular-only; >= 30% of the assessed compounds promiscuous (more than 5 other human targets at pChEMBL >= 6); any PAINS-flagged compound; a binding site < 95% identical to mouse. Ranked by flag count, then like-for-like shift, then promiscuous share. Off-targets include paralogues: a carbonic anhydrase IX inhibitor that also hits CA I and CA II counts as promiscuous. Across the 60 oncology targets in the curated graph; kg_version 327.
Built to de-risk Phase II
Live in dossiers + MCP
Human and mouse potency compared only like for like — compounds measured cell-free with the same standard type on both species — beside binding-site identity from aligned UniProt binding residues. Below 8 compounds the verdict is “no verdict”. Across 60 curated targets, 40 have no cell-free mouse protein chemistry at all: the blind spot is reported, not zeroed.
Live in dossiers + MCP
Each target's 50 most potent compounds: PAINS and Brenk structural alerts, promiscuity (more than 5 other human targets at pChEMBL ≥ 6), hERG testing and ADME-assay coverage — every count a floor. 1,347 of 2,661 show no recorded activity at pChEMBL ≥ 6 on any other human protein: unprofiled, not selective.
Live via MCP · expanding
mosaic_resistance_bypass_map surfaces targets co-mentioned with yours in resistance-context abstracts, ranked by strong resistance evidence and low development activity, each with its source snippet — leads for combination strategy, not evidence of resistance.
Live via MCP · expanding
DepMap CRISPR essentiality across 1,208 cell lines, co-essentiality partners at r ≥ 0.20, and mosaic_synthetic_lethal_whitespace: partners coupled to a drugged target by co-essentiality, STRING or Reactome that are themselves undeveloped — with partners outside the curated graph marked unknown, never zero.
Preview · not built yet
Baseline RNA and protein range across lineages, to estimate the expression threshold a response needs. Not built yet — ask for it in a custom dossier.
Preview · not built yet
Feature matrices for virtual-cell and digital-twin training where “measured · zero”, “not in universe” and “never measured” stay separate values instead of collapsing to 0.0. Not built yet as an export.
A count of zero and a count nobody took look identical in every other tool. Here they are different values: an earned zero names the query and the date the source answered it. Nothing unmeasured ever renders as zero.
Where ingestion is capped, the count says ≥ and the state says truncated — it supports “at least N” and never an absence claim. The trials axis is a floor everywhere, and the dossiers say so on every page.
Every dossier passes a rule-based verifier: no claim may rest on a failed axis, precision figures must disclose sampling, and any figure that can’t be traced to fetched evidence is listed as unsupported — in the dossier itself.