Evidence atlas of Hippo/RAM × actin candidate associations inside a 359-protein STRING v12 neighborhood. Each pair is joined to BioGRID and IntAct on UniProt accession, labeled with STRING channel scores, and overlaid with a GraphSAGE rank. STRING functional edges are context, not a physical-binding claim.
Left: exclusive catalog class by species. Right: GraphSAGE score versus STRING degree product.
The Hippo pathway (MST1/2–LATS1/2–YAP/TAZ/TEAD in humans) controls organ size and contact inhibition. YAP/TAZ activity is tightly coupled to F-actin and cytoskeletal tension, but catalogs disagree about which pairs have physical support.
This project stays small enough for a 16 GB Apple M2:
- Human (NCBI 9606): Hippo seeds YAP1, WWTR1 (TAZ), LATS1/2, STK3/4, SAV1, MOB1A, NF2, TEAD1 and actin seeds ACTB, CFL1, PFN1, GSN, DIAPH1, ACTN1, VCL, RHOA, CDC42, RAC1.
- Yeast (4932): there is no YAP/TAZ. The RAM/MOR network (CBK1, KIC1, MOB2, TAO3, SOG2, HYM1) is the Hippo-like module, paired with ACT1 and polarity/actin regulators.
STRING partners are expanded at combined score ≥ 700 (max 20 per seed). The result is 359 proteins and 2,484 undirected edges — not a whole-proteome dump. The atlas itself is the 160 same-species Hippo × actin pairs among the 36 seeds.
BioGRID organism TAB3 (physical vs genetic) and IntAct PSI-MI TAB are joined on UniProt accession. STRING channel scores are fetched for atlas proteins. Each pair is physical_curated, string_functional_only, unreported, or artifact_risk.
Each sequence is encoded with Meta’s facebook/esm2_t12_35M_UR50D (hidden size 480). Residue hidden states are mean-pooled, excluding CLS/EOS/pad. Sequences longer than 1022 amino acids are truncated to the ESM-2 context limit.
Observed STRING edges are split per species 80 / 10 / 10. Message passing uses train edges only. Negatives are random same-species non-edges. That AUROC is a methods check on functional STRING edges. Physical BioGRID/IntAct labels are a separate benchmark (make benchmark).
Atlas (160 Hippo × actin pairs; 156 absent from STRING ≥ 700):
| STRING-absent class | n | fraction |
|---|---|---|
| unreported | 132 | 84.6% |
| physical_curated (BioGRID/IntAct) | 13 | 8.3% |
| artifact_risk | 11 | 7.1% |
GraphSAGE probability vs STRING degree product, Spearman 0.58. Yeast Act1–Kic1 is unreported; human YAP1–ACTB is already Affinity Capture–MS but STRING combined score 0.515. These are candidate associations, not bindings.
STRING-edge test AUROC 0.910 vs Adamic–Adar 0.904 (random negatives, one seed) is neighborhood recovery, not physical discovery. On BioGRID/IntAct physical labels (3,124 edges, 20 seeds), logistic regression on degree and common neighbors reaches AUROC 0.741 ± 0.019; GraphSAGE+ESM-2 is 0.720 ± 0.026 and does not beat Adamic–Adar (0.732 ± 0.018). Sequence features help on degree-matched negatives.
The runtime is the hippoact:latest image (linux/arm64, CPU PyTorch). There is no local virtualenv. Docker Desktop on Mac cannot use Metal (mps); in-container training is CPU and is fast on this 359-node graph.
docker compose build
make test # unit tests (no live BioGRID/IntAct download)
make fetch # STRING + UniProt download
make embed # ESM-2 node features
make arch # GNN shape check
make train # train + hit list
make viz # NetworkX figure
make atlas # BioGRID/IntAct atlas; runs benchmark if the Phase 1 gate opens
make benchmark # physical-label benchmark only (needs atlas outputs)Set HIPPO_SKIP_BENCHMARK=1 on the atlas container to stop after the CSV/QC JSON.
docker compose run --rm -e HIPPO_SKIP_BENCHMARK=1 --entrypoint python pipeline -m src.evidence_atlasReclaim disk (image + Hugging Face cache volume):
make clean| Path | Contents |
|---|---|
data/raw/proteins.csv, interactions.csv, proteins.fasta |
STRING/UniProt graph |
data/raw/evidence/ |
BioGRID TAB3, IntAct MITAB, STRING channels, UniProt locations |
data/processed/hippo_actin_atlas.csv |
160-row evidence atlas |
data/processed/atlas_qc.json, atlas_stats.json |
QC and missingness stats |
data/processed/physical_edges.csv |
BioGRID/IntAct physical edges in the 359-protein set |
data/processed/node_embeddings.pt |
[359, 480] float32 features |
data/processed/gnn_best.pt |
Best GraphSAGE weights (STRING-edge training) |
data/processed/top_predicted_interactions.csv |
Ranked STRING-absent Hippo–actin pairs (top 50) |
data/processed/benchmark_metrics.json |
20-seed physical-label AUROC/AP/CIs |
data/processed/benchmark_stability.csv |
Median ranks of STRING-absent Hippo×actin pairs |
figures/figure4_evidence_classes.png |
Evidence-class stacked bars |
figures/figure5_physical_benchmark.png |
Physical-label AUROC bars |
- Target: Apple M2, 16 GB unified memory.
- ESM-2 35M peak RSS during embedding was ~650 MB.
- PyPI linux/aarch64
torchwheels can pull CUDA libraries; the Dockerfile installs CPU torch fromhttps://download.pytorch.org/whl/cpu.
- A high GraphSAGE probability is not experimental evidence and is not a binary contact.
- BioGRID “physical” is not the same as a reconstituted pair.
- STRING ≥ 700 is incomplete; some “novel” pairs exist in BioGRID/IntAct or at lower STRING scores.
- 89 long proteins were truncated at 1022 residues.
- Human and yeast graphs are trained together but never share edges.
- Docker uses CPU, not MPS.
STRING (https://string-db.org), UniProt, BioGRID, and IntAct are used via their public files/APIs for research. ESM-2 weights are loaded from Hugging Face (facebook/esm2_t12_35M_UR50D).