Disease2Target — Pancreatic Cancer KRAS-Combination Case Study
Systems Pharmacology AI Research Center (SPARC), University of Alabama at Birmingham Prepared 26 September 2026.
Retrospective validation. These pages were transcribed from the manuscript in the Box folder
Disease2Target_Pancreatic_Case_Study.
Looking for more detail than this? The seven full target reports are the deepest layer — every raw metric behind every score, with its source. They are listed with direct links in supporting-data. Figures stay in Box under lab access control; the analysis is here.
Abstract
Background. The SPARC pancreatic cancer case study (“Drug Discovery at the Speed of AI”, slides 25–40 and 67) proposes that a KRAS inhibitor could be combined with a FAK inhibitor or an ERBB inhibitor, because pancreatic tumours rely on FAK at baseline and switch on ERBB2/3 when KRAS is blocked. We asked whether Disease2Target, SPARC’s target-prioritisation web application, independently holds evidence for the claims the case study makes.
Methods. We loaded the pancreatic adenocarcinoma dataset in Disease2Target (snapshot #143 v9, 6,000 genes). We pulled the full evidence report for the seven genes the case study relies on (KRAS, PTK2/FAK, SRC, ERBB2, ERBB3, EGFR, FN1) and put the slide 67 clinical question to the app’s AI co-pilot twice. Each of 14 case study claims was classed as supported, partly supported, not testable in the app, or missed. All app scores were then recomputed by hand from their raw inputs, using the app’s published source code.
Results. The app supports 8 of the 14 claims and partly supports 3. Two need data the app does not hold (single-cell co-expression, survival modelling), and one publication was missed by the co-pilot. The app’s DepMap dependency for FAK (Chronos −0.64, 58% of pancreatic lines dependent) matches the case study’s value (−0.616, 57%) within 0.03. KRAS is the strongest dependency (Chronos −2.14, 96% of lines) and ranks #2 of 6,000 targets. SRC sits at the 99.8th percentile of network centrality but is a weak dependency (Chronos −0.22), as the case study argues. ERBB2 and ERBB3 both score 100/100 on tractability. Every displayed score was reproduced from its raw inputs (49 of 49 criterion scores and all 7 overall scores, within one point of rounding).
Conclusion. Disease2Target confirms the case study’s target-level evidence quickly and traceably. The central new finding — ERBB2/3 induction after KRAS blockade — comes from the lab’s own perturbation data and cannot be shown in the app. The co-pilot’s literature recall still needs human checking.
Pages
- methods — dataset, gene selection, capture procedure, co-pilot protocol, claim classification, score reproduction
- results — per-gene findings, claim tally, co-pilot runs, score reproduction
- discussion — what the app adds, limitations, points to reconcile, conclusion
- supporting-data — inventory of the 32 supporting figures and where they live
- references — full reference list
Claim tally at a glance
| Classification | Count |
|---|---|
| Supported | 8 |
| Partly supported | 3 |
| Not testable in app | 2 |
| Missed by co-pilot | 1 |
| Total claims assessed | 14 |
The seven genes
| Gene | Rank (Ranking Board) | Overall score | DepMap Chronos | % lines dependent | Note |
|---|---|---|---|---|---|
| KRAS | #2 / 6,000 | 100 | −2.14 | 96% | Top-tier candidate; backbone of the combination |
| PTK2 (FAK) | #293 | — | −0.64 | 58% | Standing dependency; low genetics/expression scores |
| SRC | — | — | −0.22 | 15% | Network hub (99.8th pct), weak dependency |
| ERBB2 | #72 | — | — | — | Tractability 100/100 |
| ERBB3 | #108 | — | — | — | Tractability 100/100 |
| EGFR | — | — | — | — | SRC network neighbour |
| FN1 | — | — | — | 0 (dependency score) | SRC network neighbour |
Introduction
Choosing a drug target means answering several questions at once. Does the tumour need the gene? Can a drug reach the protein? Has anyone tried it in patients? Is it safe to block? Normally each question is answered in a different database (DepMap, Open Targets, ChEMBL, ClinicalTrials.gov, gnomAD, PubMed, and so on), and the reasoning is rarely written down in one place.
Disease2Target puts these sources together into one scored card per gene, for every gene in a disease. Its ranking is a transparent weighted sum over eight criteria with no machine-learning model, and every score opens to the raw numbers behind it.
The SPARC case study makes a specific, testable set of claims about seven genes. That makes it a good test of whether the app’s evidence agrees with a completed piece of research. This is a retrospective validation: the case study was carried out with the lab’s own analysis pipeline, and here we ask whether the app, on its own, holds the same evidence.