Pan-Cancer Synthetic-Lethality Explorer
A tool for finding drug + mutation pairs where a drug kills cancer cells that carry a specific mutation more than otherwise-similar cells that lack it. That preferential killing is called synthetic lethality, and it is one of the cleanest routes to a cancer therapy with a built-in target. This app scans large public drug-screening datasets to surface those leads and lets you inspect the evidence behind each one.
We believe synthetic lethality is understudied in cancers and deserves a deeper focus. See this framing from contemporary related work by another lab:
"Synthetic lethality (SL) is defined as a genetic interaction wherein concurrent functional loss of two genes results in cellular or organismal death, while inactivation of either gene alone remains compatible with survival. This unique biological phenomenon has emerged as a transformative paradigm in oncology, primarily owing to its capacity to address three critical limitations of conventional cancer therapies: reversing acquired or intrinsic drug resistance, minimising off-target toxicity and enhancing the precision of therapeutic targeting."
What is synthetic lethality?
Two events are synthetically lethal when neither one alone kills a cell, but the two together do. In cancer, one event is a mutation the tumour already carries (you can't undo it) and the other is a drug you can give. If a drug is lethal only in cells that carry the mutation, then the tumour's own mutation becomes its weakness — the drug harms cancer cells and largely leaves healthy cells alone.
The one metric to understand: S′, and how it becomes a lead
S′ is a single-value metric — similar to AUC — that summarizes a whole dose-response curve into one potency/efficacy number. Instead of comparing entire curves, every drug-on-a-cell-line experiment collapses to one S′ score (higher S′ = more killing).
Wild-type (WT) = cells without the mutation. Mutant = cells with it. So the app's headline question is simply: for which drugs is ΔpS′ strongly negative?
Method & validation: read the manuscript (doi.org/10.21203/rs.3.rs-9559070/v1).
Where to go — pick a starting point
Every jargon term underlined like this ⓘUnderlined terms have a plain-language definition — hover, tap, or keyboard-focus them. has a plain-language definition you can open. If a chart says the warehouse is unavailable, the data connection is down — the page stays legible and you can retry later.
What am I looking at?
Screen one gene for synthetic-lethal drugs. Choose a drug-screening dataset and a mutated gene. The app splits every cell line into mutant (carries the gene mutation) vs. wild-type (doesn't), then ranks every drug by ΔpS′ ⓘΔpS′ = pS′(wild-type) − pS′(mutant). Strongly negative = the drug kills mutant cells much more than wild-type = a synthetic-lethal lead. — how much more it kills the mutant cells.
The volcano plot: each dot is a drug. Left (negative ΔpS′) = mutant-selective killing. Higher up = more statistically significant. Red dots pass the significance cutoff (q ≤ 0.1) — those are your candidate leads. A permutation test guards against chance, so bigger datasets take a moment.
How the ranking works, in plain terms: for each drug the app compares how hard it hit the mutant cell lines versus the wild-type ones (that gap is ΔpS′). To be sure a gap is real and not chance, it runs a permutation test — it randomly re-labels the cell lines many times (the "Permutations" count above) and checks how often pure luck produces a gap that big. Drugs whose real gap is rarely beaten by luck get a low q-value and show up as red dots — those are your candidate synthetic-lethal leads.
What am I looking at?
Two ways of judging whether two drugs are "alike," side by side. One is structure (do they look chemically similar?); the other is response (do they kill the same cells the same way?). Usually these agree — similar chemistry, similar effect.
The scatter plot puts every close-structure drug pair at (structural similarity, response similarity). The interesting cases are the red dots: near-identical chemistry but opposite cell-kill effect. That is an activity cliff ⓘTwo nearly-identical molecules with sharply different biological effects — a warning that you can't assume they share a mechanism just because they look alike. — a "do NOT assume these share a mechanism" flag.
Below, search any compound to see it through every lens at once, with its neighbours and an imputed mechanism (a best-guess MoA borrowed from its trustworthy look-alikes), plus the trust gates that back — or veto — that guess.
The hero view. Every compound sits in two independent spaces at once — structure (Lens 1, ECFP4/Tanimoto) and response (Lens 2, its 60-line NCI-60 viability vector). Where they agree (structurally close AND response-correlated) we can transitively impute a drug's mechanism from its labeled neighbours. Where they disagree — near-identical chemistry, opposite cellular outcome — that is an activity cliff: the "do NOT annotate across this" signal. Thresholds are derived from this dataset at load time (no magic numbers).
Disagreement map — structural similarity vs response concordance (red = activity cliff: close structure, discordant response)
Drug-detail card — one compound through every ready lens
What am I looking at?
This is a similarity matrix between drug-mechanism classes — not genes. "MoA" = mechanism of action, the way a drug works (e.g. "topoisomerase inhibitor", "tubulin binder"). Every drug in a class is pooled into one cell-kill fingerprint: how it kills across a standard panel of 60 cell lines (NCI-60).
- Each row and each column is one MoA class. Each square shows how correlated two classes' cell-kill patterns are — high correlation means "these two mechanisms kill the same kinds of cells."
- Colour = correlation. Red ≈ +1 (very alike), white ≈ 0 (unrelated), blue ≈ −1 (opposite patterns).
- The diagonal is always red (1.0) — every class is identical to itself.
- Red blocks off the diagonal = families of related mechanisms that behave alike, discovered purely from cell-kill data — no chemical structure and no gene/target information used.
This tab is not about driver genes or drug targets. It answers one question: "do drugs grouped by mechanism actually behave in coherent, distinguishable ways?" (They do — that is the point.)
Dataset & detail (below). The NCI-60 substrate is the cached, pre-pooled set (~40 classes, fixed at ≥5 drugs/class). Switch to PRISM (live warehouse) to see many more mechanism classes, and lower the min drugs / class threshold to surface even more (≥5 → 77 classes, ≥3 → 133). A lower threshold trades per-class robustness for breadth of mechanisms.
What am I looking at?
How chemically self-similar the drug library is. Each drug's chemical structure is turned into a fingerprint, and we measure similarity between drugs with the Tanimoto score ⓘA 0-to-1 chemical-similarity score. 1.0 = identical structure; ~0 = nothing in common. 0.5+ is usually considered a close analog. (0 = unrelated, 1 = identical).
The histogram shows, for every drug, how similar its closest neighbour is. A bar on the right (high Tanimoto) means many drugs have a near-twin in the library — those analog pairs are the raw material for reasoning "if drug A works this way, its close analog B probably does too."
Scope — read this first: this lens is computed on ONLY the NCI-60 compound library (25,245 compounds), the app's fixed structural reference set. It is not a dataset you pick and not a single compound — the chart summarises the whole library, where every compound contributes one bar at its single closest-neighbour similarity. To make it concrete, use the search below to drop any compound onto this distribution and see its actual nearest structural neighbours.
Could this run on other datasets — and does it matter? Structural similarity depends only on the molecules, not on cell lines, so a compound's fingerprint is identical in every dataset — one structural reference already covers any overlapping compound. Running it over other libraries (e.g. PRISM/CoderData compounds, which also ship SMILES) is future work that would let mechanism-transfer operate directly on those drugs; it wouldn't change the chemistry shown here. NCI-60 is used because it is the largest annotated compound set and the substrate for the MoA-imputation story.
ECFP4 nearest-neighbour density over the NCI-60 compound library. Analogs seed structural transitive annotation.
Why this page matters
Structural similarity is Lens 1 — the foundation of the app's mechanism-transfer idea: an unlabelled drug can borrow a likely mechanism from a close chemical analog whose mechanism is already known. This page measures how much of the library even has such analogs. It becomes actionable on the 🔀 Cross-Lens Disagreement tab, where Lens 1 (structure) is checked against Lens 2 (cell-kill response): where the two agree, a mechanism can be transferred; where near-identical chemistry produces opposite effects, that pair is an activity cliff and the transfer is blocked.
Method & validation: read the manuscript (doi.org/10.21203/rs.3.rs-9559070/v1).
What am I looking at?
The main deliverable: mutant-vs-wild-type drug selectivity for a whole panel of driver genes at once. Pick a preset panel (or build your own), and every chart below regenerates live from the warehouse. Three terms run through all of them:
- Response distribution — the spread of drug effect (pS′) and of the genotype gap (ΔpS′). Mass to the left of the red 0-line = mutant-selective.
- Selectivity map — each dot is a drug: wild-type pS′ (x) vs mutant pS′ (y). Dots above the diagonal kill mutant cells more; red dots clear the synthetic-lethal cutoff.
- MoA signature — which drug-mechanism classes are, on average, the most mutant-selective (most negative ΔpS′).
- Metric concordance — a sanity check that S′ tracks the standard AUC metric.
- Replication validation — our ΔpS′ vs the manuscript's published values, to show the pipeline reproduces known results.
The synthetic-lethal cutoff is set per dataset (a validated preset pins the manuscript standard; a new panel derives it from its own ΔpS′ spread) and is adjustable in the controls below — no single magic number.
Genotype-selectivity views over any preset gene panel — response distribution, WT–mutant selectivity map, MoA signature, and summary stats, all regenerated live from the warehouse. ΔpS′ = WT − mutant pooled pS′; the synthetic-lethal cutoff is set per dataset (a validated preset pins an absolute standard; a new dataset derives it from its own ΔpS′ spread) and is overridable below.
Build your own panel — any dataset × any driver genes
- preset default — use the cutoff this dataset ships with (a validated panel pins the published standard; a newly built panel derives one from its own data). Start here.
- absolute ΔpS′ — a fixed number you type, e.g. −2 (the manuscript standard). Simplest, and directly comparable across datasets.
- mean − k·σ or median − k·MAD — data-derived: place the line k spread-widths below the middle of this dataset's own ΔpS′ values (σ = standard deviation; MAD = a robust, outlier-resistant spread). No magic number — the cutoff adapts to the data.
- k-th percentile — keep only the most extreme k% of gaps (e.g. the bottom 5%).
Response distribution — operational drug effect (pS′) and genotype shift (ΔpS′)
Selectivity map — WT vs mutant pS′ (points above diagonal = mutant-selective)
MoA signature — median ΔpS′ per mechanism class
Response classes — representative compound per S′-defined class
Summary statistics per genotype
Metric concordance — S′ vs standard AUC (dataset-wide; confirms S′ tracks the established metric)
Replication validation — our ΔpS′ vs published values (pipeline accountability: does it reproduce known synthetic-lethal effects?)
⚠️ Experimental — not peer-reviewed. These reports have not been peer-reviewed and should be considered experimental. This is an early hackathon build for research exploration, not clinical use.
What this app does
A working pan-cancer synthetic-lethality discovery tool that leverages numerous open cancer datasets:
- Genotype-selectivity engine (S′ / ΔpS′) — every drug–cell-line dose-response curve collapses to one S′ score; pooled per genotype, the mutant-vs-wild-type gap ΔpS′ ranks synthetic-lethal leads.
- The transform reproduces the manuscript exactly (SCALE = 100).
- SL Scan — one gene, every drug — pick a screen and a driver gene; a permutation test ranks every compound by mutant-selective killing with FDR q-values.
- Runs across all viable datasets: nine CoderData screens plus DepMap PRISM (the manuscript's primary substrate, r = 0.968).
- Presence-gated, so no option in the UI ever errors on use.
- Selectivity Explorer — a whole driver-gene panel at once — response distribution, WT-vs-mutant selectivity map, MoA signature, and summary stats, all regenerated on demand.
- A validated lung preset reproduces the published synthetic-lethal counts exactly.
- A build-your-own-panel tool runs the same method on any dataset (PRISM included) × any driver genes × any tissue.
- Cross-lens mechanism transfer (the hero) — chemical structure (Lens 1) and cell-kill response (Lens 2) checked against each other.
- Where they agree, an understudied drug's mechanism is imputed from trustworthy chemical look-alikes.
- Where near-identical chemistry gives opposite effects, that activity cliff blocks the transfer — evidence trail and trust gates shown.
- Two similarity lenses, live
- 🧪 Structural — ECFP4 / Tanimoto over 25,245 compounds, searchable per compound.
- 🔬 MoA — mechanism-class similarity by cell-kill pattern.
On the roadmap
- Response as a standalone lens tab (already powering the Cross-Lens view)
- Physicochemical and Pathway similarity lenses
- General cross-screen replication report (beyond the validated lung case)
- Reliability envelope — S′ spread across labs sharing cell lines
- Genetic dependency — RNAi / CRISPR essentiality vs chemical synthetic lethality
- In-vivo / clinical — PDX tumour-growth and patient survival
Datasets & citations
All results are computed from public datasets — we do not redistribute source data, and each source retains its own license and terms of use. Full MLA citations and a reproduction guide are in the repository's CITATIONS.md (source code on GitHub). Datasets used:
- DepMap PRISM Repurposing (19Q4) — Corsello et al., Nature Cancer, 2020.
- CCLE / DepMap mutations (24Q2) — Ghandi et al., Nature, 2019; DepMap Portal.
- CTRPv2 — Seashore-Ludlow et al., Cancer Discovery, 2015; Rees et al., Nature Chemical Biology, 2016.
- gCSI — Haverty et al., Nature, 2016.
- BeatAML — Tyner et al., Nature, 2018.
- Novartis PDX — Gao et al., Nature Medicine, 2015.
- GDSC — Iorio et al., Cell, 2016; Yang et al., Nucleic Acids Research, 2013.
- NCI-60 — Shoemaker, Nature Reviews Cancer, 2006; NCI Developmental Therapeutics Program.
- FIMM & CoderData tissue cohorts (bladder, colorectal, liver, pancreatic, sarcoma, MPNST) — harmonized via CoderData / IMPROVE (PNNL, U.S. DOE & NCI).
- Gene identifiers — NCBI Gene / Entrez (Sayers et al., Nucleic Acids Research, 2022).
Method & references
The S′ / ΔpS′ metric, the permutation-based synthetic-lethality calling, and the cross-lens mechanism-transfer framework are described in the project manuscript; the validated lung preset reproduces its published counts exactly.
- Project manuscript (preprint): doi.org/10.21203/rs.3.rs-9559070/v1
- Synthetic-lethality background (quoted on Start Here): Li et al., Clinical and Translational Medicine, 2026, doi:10.1002/ctm2.70586
API docs: /docs