{
  "id": 664239,
  "title": "Protein Function Prediction Software vs Methods",
  "url": "/competitions/cafa-6-protein-function-prediction/discussion/664239",
  "author_name": "",
  "post_date": "2025-12-23T11:23:41.417548200Z",
  "votes": 7,
  "comment_count": 7,
  "views": 0,
  "content": "<table>\n<thead>\n<tr>\n<th><strong>Software (year)</strong></th>\n<th><strong>Sequence similarity / transfer</strong></th>\n<th><strong>Domain signatures</strong></th>\n<th><strong>Classic DL</strong></th>\n<th><strong>Transformers / pLM</strong></th>\n<th><strong>Use of GO ontology</strong></th>\n<th><strong>Zero-shot</strong></th>\n<th><strong>3D structure source</strong></th>\n<th><strong>GNN / Graph Transformer</strong></th>\n<th><strong>Text mining / text retrieval</strong></th>\n<th><strong>PPI networks</strong></th>\n<th><strong>Multimodal / hybrid</strong></th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td><strong>DeepGOZero (2022)</strong></td>\n<td></td>\n<td>✔️ (InterPro-derived features used as inputs)</td>\n<td>✔️ (feedforward/NN modules for scoring)</td>\n<td>✔️ (ontology &amp; embedding pipeline; not pLM-centered)</td>\n<td>✔️ (ontology axioms → ontology embeddings; exploits GO axioms).</td>\n<td>✔️ (ontology-axiom zero-shot prediction).</td>\n<td></td>\n<td></td>\n<td></td>\n<td></td>\n<td>✔️ (sequence + ontology embeddings hybrid).</td>\n</tr>\n<tr>\n<td><strong>NetGO 3.0 (2023)</strong></td>\n<td>✔️ (homology features / BLAST-like signals included in ensemble)</td>\n<td>✔️ (InterProScan used as a component)</td>\n<td></td>\n<td>✔️ (ESM-1b embeddings via LR-ESM component).</td>\n<td>✔️ (GO term supervision &amp; hierarchy integrated in scoring).</td>\n<td></td>\n<td></td>\n<td></td>\n<td>✔️ (functional text embeddings for some components).</td>\n<td>✔️ (STRING integrated for network features).</td>\n<td>✔️ (ensemble: sequence + pLM + network + domains).</td>\n</tr>\n<tr>\n<td><strong>TALE (2021)</strong></td>\n<td></td>\n<td></td>\n<td></td>\n<td>✔️ (custom Transformer; joint sequence–label embedding)</td>\n<td>✔️ (label embeddings aligned to GO terms)</td>\n<td>✔️ (improves tail-label generalisation / few-shot)</td>\n<td></td>\n<td></td>\n<td></td>\n<td></td>\n<td>✔️ (sequence-only transformer with label embedding).</td>\n</tr>\n<tr>\n<td><strong>ProTranslator / ProTransfer (2022)</strong></td>\n<td>✔️ (can transfer via text→sequence similarity in practice)</td>\n<td></td>\n<td></td>\n<td>✔️ (text⇄sequence “translation” approach; NLP models)</td>\n<td>✔️ (relies on textual GO descriptions to align functions)</td>\n<td>✔️ (translation-based zero-shot to novel textual descriptors)</td>\n<td></td>\n<td></td>\n<td>✔️ (uses GO textual definitions / retrieval)</td>\n<td></td>\n<td>✔️ (text-driven + sequence matching).</td>\n</tr>\n<tr>\n<td><strong>DeepFRI (2020)</strong></td>\n<td></td>\n<td></td>\n<td>✔️ (sequence LSTM/CNN features originally used)</td>\n<td>✔️ (uses sequence embeddings; later work uses pLM features)</td>\n<td></td>\n<td></td>\n<td>✔️ (PDB structures / modeled contact maps)</td>\n<td>✔️ (GCN on residue contact graph)</td>\n<td></td>\n<td></td>\n<td>✔️ (structure + sequence + GCN).</td>\n</tr>\n<tr>\n<td><strong>InterProScan (2005)</strong></td>\n<td>✔️ (signature-based mapping → annotation transfer)</td>\n<td>✔️ (Pfam, PROSITE, SMART, CATH-Gene3D, TIGRFAMs etc.)</td>\n<td></td>\n<td></td>\n<td></td>\n<td></td>\n<td></td>\n<td></td>\n<td></td>\n<td></td>\n<td>✔️ (classic signature-based pipeline used as component).</td>\n</tr>\n<tr>\n<td><strong>TransFun (2022)</strong></td>\n<td></td>\n<td>✔️ (optional InterPro features)</td>\n<td></td>\n<td>✔️ (ESM embeddings used)</td>\n<td>✔️ (predicts GO terms; uses GO supervision)</td>\n<td></td>\n<td>✔️ (AlphaFold2 predicted structures / PDB when available)</td>\n<td>✔️ (EGNN — 3D-equivariant GNN)</td>\n<td></td>\n<td></td>\n<td>✔️ (sequence pLM + AlphaFold + EGNN ensemble).</td>\n</tr>\n<tr>\n<td><strong>ProtNLM (UniProt, 2024)</strong></td>\n<td></td>\n<td>✔️ (UniProt pipeline includes InterPro features)</td>\n<td></td>\n<td>✔️ (protein LLM to generate descriptive text / labels)</td>\n<td>✔️ (mapped to GO via UniProt rules/pipelines)</td>\n<td>✔️ (generative annotation can assign novel descriptions)</td>\n<td></td>\n<td></td>\n<td>✔️ (UniProt text-mining pipelines; integration with literature)</td>\n<td>✔️ (cross-DB links; UniProt integrates PPI sources)</td>\n<td>✔️ (LLM-generated descriptions + rule-based pipelines).</td>\n</tr>\n<tr>\n<td><strong>PFresGO (2023)</strong></td>\n<td></td>\n<td>✔️ (InterPro integration mentioned)</td>\n<td>✔️ (attention-based DL for residue localization)</td>\n<td>✔️ (ProtTrans / ProtT5 embeddings often used)</td>\n<td>✔️ (explicit modelling of GO hierarchical relationships)</td>\n<td></td>\n<td></td>\n<td>✔️ (GO graph attention / cross-attention)</td>\n<td></td>\n<td></td>\n<td>✔️ (sequence embeddings + GO graph attention).</td>\n</tr>\n<tr>\n<td><strong>GOAnnotator (various, 2019– )</strong></td>\n<td>✔️ (can use BLAST/DIAMOND for evidence)</td>\n<td>✔️ (optionally uses InterProScan outputs)</td>\n<td></td>\n<td>✔️ (sentence-transformers / SBERT for retrieval in modern versions)</td>\n<td>✔️ (maps extracted evidence to GO terms)</td>\n<td>✔️ (retrieval-based assignment to new terms)</td>\n<td></td>\n<td>✔️ (optional graph fusion with PPI)</td>\n<td>✔️ (PubMed retrieval; PubTator / sentence-transformers / BM25 + rerank)</td>\n<td>✔️ (STRING / BioGRID integration optional)</td>\n<td>✔️ (text retrieval + sequence/domain evidence).</td>\n</tr>\n<tr>\n<td><strong>STAR-GO (2025, preprint)</strong></td>\n<td></td>\n<td>✔️ (domain features optional)</td>\n<td></td>\n<td>✔️ (Transformer + GO sentence embeddings)</td>\n<td>✔️ (joint semantic + structural GO embeddings)</td>\n<td>✔️ (designed for semantic zero-shot on GO terms)</td>\n<td>✔️ (evaluated with PDB / AlphaFold-based datasets)</td>\n<td>✔️ (GO graph integration; hierarchical propagation)</td>\n<td>✔️ (GO definitions via sentence encoders)</td>\n<td></td>\n<td>✔️ (semantic GO embeddings + sequence alignment).</td>\n</tr>\n<tr>\n<td><strong>ProtFun (2025)</strong></td>\n<td></td>\n<td>✔️ (InterPro signatures used to build protein family network)</td>\n<td></td>\n<td>✔️ (uses ANKH protein LLM embeddings)</td>\n<td>✔️ (predicts GO labels supervised)</td>\n<td></td>\n<td></td>\n<td>✔️ (GAT on protein family network / PFN)</td>\n<td></td>\n<td>✔️ (custom PFN; may include family-based links)</td>\n<td>✔️ (LLM embeddings + InterPro-based PFN + GAT).</td>\n</tr>\n<tr>\n<td><strong>DIAMOND2GO / Diamond2GO (2024)</strong></td>\n<td>✔️ (DIAMOND sequence similarity → transfer)</td>\n<td></td>\n<td></td>\n<td></td>\n<td></td>\n<td></td>\n<td></td>\n<td></td>\n<td></td>\n<td></td>\n<td>✔️ (fast sequence-based GO transfer via DIAMOND; optional InterPro enrichment).</td>\n</tr>\n<tr>\n<td><strong>DPFunc (2024)</strong></td>\n<td></td>\n<td>✔️ (InterProScan-guided domain features)</td>\n<td>✔️ (deep learning architecture focused on interpretability)</td>\n<td>✔️ (ESM-1b embeddings included)</td>\n<td></td>\n<td></td>\n<td>✔️ (PDB / AlphaFold2 structures)</td>\n<td>✔️ (GCN layers on residue/structure graph)</td>\n<td></td>\n<td>✔️ (optional PPI datasets for evaluation)</td>\n<td>✔️ (domain-guided structural DL + GCN).</td>\n</tr>\n<tr>\n<td><strong>FANTASIA (2024)</strong></td>\n<td>✔️ (ProtT5 / ProtTrans embeddings used for similarity)</td>\n<td>✔️ (domain-aware clustering in pipeline)</td>\n<td></td>\n<td>✔️ (ProtT5 / GOPredSim pipeline; embedding similarity)</td>\n<td></td>\n<td>✔️ (embedding-similarity generalization for dark proteome)</td>\n<td></td>\n<td></td>\n<td>✔️ (embedding retrieval/search)</td>\n<td></td>\n<td>✔️ (embedding-similarity GO term transfer at proteome scale).</td>\n</tr>\n<tr>\n<td><strong>ProtBoost (2024) – CAFA-5 top</strong></td>\n<td>✔️ (pLM embeddings used in ensemble)</td>\n<td>✔️ (InterPro features included in ensembles)</td>\n<td></td>\n<td>✔️ (ESM / ProtTrans embeddings used)</td>\n<td></td>\n<td></td>\n<td></td>\n<td>✔️ (GCN used in stacking/ensemble)</td>\n<td>✔️ (text/features optional in ensemble)</td>\n<td>✔️ (STRING / BioGRID optional)</td>\n<td>✔️ (Py-Boost + GCN ensemble; multi-source stacking).</td>\n</tr>\n<tr>\n<td><strong>AnnoPRO (2024)</strong></td>\n<td></td>\n<td>✔️ (multiscale domain features included)</td>\n<td>✔️ (multi-scale DL modules for long-tail GO terms)</td>\n<td>✔️ (pretrained encoders used)</td>\n<td>✔️ (explicitly targets long-tail GO terms)</td>\n<td></td>\n<td></td>\n<td>✔️ (hybrid deep networks)</td>\n<td>✔️ (optional text features)</td>\n<td>✔️ (optional PPI features)</td>\n<td>✔️ (multiscale features + hybrid DL).</td>\n</tr>\n<tr>\n<td><strong>TEMPROT (2023)</strong></td>\n<td>✔️ (BLAST / homology used in ensemble versions)</td>\n<td></td>\n<td></td>\n<td>✔️ (ProtBERT-BFD embeddings)</td>\n<td></td>\n<td></td>\n<td></td>\n<td></td>\n<td></td>\n<td></td>\n<td>✔️ (transformer embeddings + homology ensemble).</td>\n</tr>\n<tr>\n<td><strong>Prot2Text-V2 (2025, arXiv)</strong></td>\n<td></td>\n<td>✔️ (domains &amp; structure optional)</td>\n<td></td>\n<td>✔️ (multimodal protein LLM + sequence encoders)</td>\n<td>✔️ (maps generated text to GO semantic terms)</td>\n<td>✔️ (text-generation based zero-shot explanations)</td>\n<td>✔️ (AlphaFold / ESMFold / PDB optional)</td>\n<td>✔️ (GNN + Transformer multimodal encoder)</td>\n<td>✔️ (generative functional text output)</td>\n<td>✔️ (optional PPI ingestion)</td>\n<td>✔️ (multimodal generation: text + graph + structure).</td>\n</tr>\n<tr>\n<td><strong>GoCurator (2023)</strong></td>\n<td></td>\n<td></td>\n<td></td>\n<td>✔️ (BioBERT / sentence-transformers for curation assistance)</td>\n<td>✔️ (curation maps literature evidence to GO terms)</td>\n<td>✔️ (literature-driven zero-shot candidate suggestions)</td>\n<td></td>\n<td></td>\n<td>✔️ (PubMed / GOA sentence extraction; PubTator-like tools)</td>\n<td></td>\n<td>✔️ (human-in-the-loop curation + automated retrieval).</td>\n</tr>\n</tbody>\n</table>",
  "messages": [
    {
      "id": "3380923",
      "postDate": "12/23/2025 11:23:41",
      "content": "<table>\n<thead>\n<tr>\n<th><strong>Software (year)</strong></th>\n<th><strong>Sequence similarity / transfer</strong></th>\n<th><strong>Domain signatures</strong></th>\n<th><strong>Classic DL</strong></th>\n<th><strong>Transformers / pLM</strong></th>\n<th><strong>Use of GO ontology</strong></th>\n<th><strong>Zero-shot</strong></th>\n<th><strong>3D structure source</strong></th>\n<th><strong>GNN / Graph Transformer</strong></th>\n<th><strong>Text mining / text retrieval</strong></th>\n<th><strong>PPI networks</strong></th>\n<th><strong>Multimodal / hybrid</strong></th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td><strong>DeepGOZero (2022)</strong></td>\n<td></td>\n<td>✔️ (InterPro-derived features used as inputs)</td>\n<td>✔️ (feedforward/NN modules for scoring)</td>\n<td>✔️ (ontology &amp; embedding pipeline; not pLM-centered)</td>\n<td>✔️ (ontology axioms → ontology embeddings; exploits GO axioms).</td>\n<td>✔️ (ontology-axiom zero-shot prediction).</td>\n<td></td>\n<td></td>\n<td></td>\n<td></td>\n<td>✔️ (sequence + ontology embeddings hybrid).</td>\n</tr>\n<tr>\n<td><strong>NetGO 3.0 (2023)</strong></td>\n<td>✔️ (homology features / BLAST-like signals included in ensemble)</td>\n<td>✔️ (InterProScan used as a component)</td>\n<td></td>\n<td>✔️ (ESM-1b embeddings via LR-ESM component).</td>\n<td>✔️ (GO term supervision &amp; hierarchy integrated in scoring).</td>\n<td></td>\n<td></td>\n<td></td>\n<td>✔️ (functional text embeddings for some components).</td>\n<td>✔️ (STRING integrated for network features).</td>\n<td>✔️ (ensemble: sequence + pLM + network + domains).</td>\n</tr>\n<tr>\n<td><strong>TALE (2021)</strong></td>\n<td></td>\n<td></td>\n<td></td>\n<td>✔️ (custom Transformer; joint sequence–label embedding)</td>\n<td>✔️ (label embeddings aligned to GO terms)</td>\n<td>✔️ (improves tail-label generalisation / few-shot)</td>\n<td></td>\n<td></td>\n<td></td>\n<td></td>\n<td>✔️ (sequence-only transformer with label embedding).</td>\n</tr>\n<tr>\n<td><strong>ProTranslator / ProTransfer (2022)</strong></td>\n<td>✔️ (can transfer via text→sequence similarity in practice)</td>\n<td></td>\n<td></td>\n<td>✔️ (text⇄sequence “translation” approach; NLP models)</td>\n<td>✔️ (relies on textual GO descriptions to align functions)</td>\n<td>✔️ (translation-based zero-shot to novel textual descriptors)</td>\n<td></td>\n<td></td>\n<td>✔️ (uses GO textual definitions / retrieval)</td>\n<td></td>\n<td>✔️ (text-driven + sequence matching).</td>\n</tr>\n<tr>\n<td><strong>DeepFRI (2020)</strong></td>\n<td></td>\n<td></td>\n<td>✔️ (sequence LSTM/CNN features originally used)</td>\n<td>✔️ (uses sequence embeddings; later work uses pLM features)</td>\n<td></td>\n<td></td>\n<td>✔️ (PDB structures / modeled contact maps)</td>\n<td>✔️ (GCN on residue contact graph)</td>\n<td></td>\n<td></td>\n<td>✔️ (structure + sequence + GCN).</td>\n</tr>\n<tr>\n<td><strong>InterProScan (2005)</strong></td>\n<td>✔️ (signature-based mapping → annotation transfer)</td>\n<td>✔️ (Pfam, PROSITE, SMART, CATH-Gene3D, TIGRFAMs etc.)</td>\n<td></td>\n<td></td>\n<td></td>\n<td></td>\n<td></td>\n<td></td>\n<td></td>\n<td></td>\n<td>✔️ (classic signature-based pipeline used as component).</td>\n</tr>\n<tr>\n<td><strong>TransFun (2022)</strong></td>\n<td></td>\n<td>✔️ (optional InterPro features)</td>\n<td></td>\n<td>✔️ (ESM embeddings used)</td>\n<td>✔️ (predicts GO terms; uses GO supervision)</td>\n<td></td>\n<td>✔️ (AlphaFold2 predicted structures / PDB when available)</td>\n<td>✔️ (EGNN — 3D-equivariant GNN)</td>\n<td></td>\n<td></td>\n<td>✔️ (sequence pLM + AlphaFold + EGNN ensemble).</td>\n</tr>\n<tr>\n<td><strong>ProtNLM (UniProt, 2024)</strong></td>\n<td></td>\n<td>✔️ (UniProt pipeline includes InterPro features)</td>\n<td></td>\n<td>✔️ (protein LLM to generate descriptive text / labels)</td>\n<td>✔️ (mapped to GO via UniProt rules/pipelines)</td>\n<td>✔️ (generative annotation can assign novel descriptions)</td>\n<td></td>\n<td></td>\n<td>✔️ (UniProt text-mining pipelines; integration with literature)</td>\n<td>✔️ (cross-DB links; UniProt integrates PPI sources)</td>\n<td>✔️ (LLM-generated descriptions + rule-based pipelines).</td>\n</tr>\n<tr>\n<td><strong>PFresGO (2023)</strong></td>\n<td></td>\n<td>✔️ (InterPro integration mentioned)</td>\n<td>✔️ (attention-based DL for residue localization)</td>\n<td>✔️ (ProtTrans / ProtT5 embeddings often used)</td>\n<td>✔️ (explicit modelling of GO hierarchical relationships)</td>\n<td></td>\n<td></td>\n<td>✔️ (GO graph attention / cross-attention)</td>\n<td></td>\n<td></td>\n<td>✔️ (sequence embeddings + GO graph attention).</td>\n</tr>\n<tr>\n<td><strong>GOAnnotator (various, 2019– )</strong></td>\n<td>✔️ (can use BLAST/DIAMOND for evidence)</td>\n<td>✔️ (optionally uses InterProScan outputs)</td>\n<td></td>\n<td>✔️ (sentence-transformers / SBERT for retrieval in modern versions)</td>\n<td>✔️ (maps extracted evidence to GO terms)</td>\n<td>✔️ (retrieval-based assignment to new terms)</td>\n<td></td>\n<td>✔️ (optional graph fusion with PPI)</td>\n<td>✔️ (PubMed retrieval; PubTator / sentence-transformers / BM25 + rerank)</td>\n<td>✔️ (STRING / BioGRID integration optional)</td>\n<td>✔️ (text retrieval + sequence/domain evidence).</td>\n</tr>\n<tr>\n<td><strong>STAR-GO (2025, preprint)</strong></td>\n<td></td>\n<td>✔️ (domain features optional)</td>\n<td></td>\n<td>✔️ (Transformer + GO sentence embeddings)</td>\n<td>✔️ (joint semantic + structural GO embeddings)</td>\n<td>✔️ (designed for semantic zero-shot on GO terms)</td>\n<td>✔️ (evaluated with PDB / AlphaFold-based datasets)</td>\n<td>✔️ (GO graph integration; hierarchical propagation)</td>\n<td>✔️ (GO definitions via sentence encoders)</td>\n<td></td>\n<td>✔️ (semantic GO embeddings + sequence alignment).</td>\n</tr>\n<tr>\n<td><strong>ProtFun (2025)</strong></td>\n<td></td>\n<td>✔️ (InterPro signatures used to build protein family network)</td>\n<td></td>\n<td>✔️ (uses ANKH protein LLM embeddings)</td>\n<td>✔️ (predicts GO labels supervised)</td>\n<td></td>\n<td></td>\n<td>✔️ (GAT on protein family network / PFN)</td>\n<td></td>\n<td>✔️ (custom PFN; may include family-based links)</td>\n<td>✔️ (LLM embeddings + InterPro-based PFN + GAT).</td>\n</tr>\n<tr>\n<td><strong>DIAMOND2GO / Diamond2GO (2024)</strong></td>\n<td>✔️ (DIAMOND sequence similarity → transfer)</td>\n<td></td>\n<td></td>\n<td></td>\n<td></td>\n<td></td>\n<td></td>\n<td></td>\n<td></td>\n<td></td>\n<td>✔️ (fast sequence-based GO transfer via DIAMOND; optional InterPro enrichment).</td>\n</tr>\n<tr>\n<td><strong>DPFunc (2024)</strong></td>\n<td></td>\n<td>✔️ (InterProScan-guided domain features)</td>\n<td>✔️ (deep learning architecture focused on interpretability)</td>\n<td>✔️ (ESM-1b embeddings included)</td>\n<td></td>\n<td></td>\n<td>✔️ (PDB / AlphaFold2 structures)</td>\n<td>✔️ (GCN layers on residue/structure graph)</td>\n<td></td>\n<td>✔️ (optional PPI datasets for evaluation)</td>\n<td>✔️ (domain-guided structural DL + GCN).</td>\n</tr>\n<tr>\n<td><strong>FANTASIA (2024)</strong></td>\n<td>✔️ (ProtT5 / ProtTrans embeddings used for similarity)</td>\n<td>✔️ (domain-aware clustering in pipeline)</td>\n<td></td>\n<td>✔️ (ProtT5 / GOPredSim pipeline; embedding similarity)</td>\n<td></td>\n<td>✔️ (embedding-similarity generalization for dark proteome)</td>\n<td></td>\n<td></td>\n<td>✔️ (embedding retrieval/search)</td>\n<td></td>\n<td>✔️ (embedding-similarity GO term transfer at proteome scale).</td>\n</tr>\n<tr>\n<td><strong>ProtBoost (2024) – CAFA-5 top</strong></td>\n<td>✔️ (pLM embeddings used in ensemble)</td>\n<td>✔️ (InterPro features included in ensembles)</td>\n<td></td>\n<td>✔️ (ESM / ProtTrans embeddings used)</td>\n<td></td>\n<td></td>\n<td></td>\n<td>✔️ (GCN used in stacking/ensemble)</td>\n<td>✔️ (text/features optional in ensemble)</td>\n<td>✔️ (STRING / BioGRID optional)</td>\n<td>✔️ (Py-Boost + GCN ensemble; multi-source stacking).</td>\n</tr>\n<tr>\n<td><strong>AnnoPRO (2024)</strong></td>\n<td></td>\n<td>✔️ (multiscale domain features included)</td>\n<td>✔️ (multi-scale DL modules for long-tail GO terms)</td>\n<td>✔️ (pretrained encoders used)</td>\n<td>✔️ (explicitly targets long-tail GO terms)</td>\n<td></td>\n<td></td>\n<td>✔️ (hybrid deep networks)</td>\n<td>✔️ (optional text features)</td>\n<td>✔️ (optional PPI features)</td>\n<td>✔️ (multiscale features + hybrid DL).</td>\n</tr>\n<tr>\n<td><strong>TEMPROT (2023)</strong></td>\n<td>✔️ (BLAST / homology used in ensemble versions)</td>\n<td></td>\n<td></td>\n<td>✔️ (ProtBERT-BFD embeddings)</td>\n<td></td>\n<td></td>\n<td></td>\n<td></td>\n<td></td>\n<td></td>\n<td>✔️ (transformer embeddings + homology ensemble).</td>\n</tr>\n<tr>\n<td><strong>Prot2Text-V2 (2025, arXiv)</strong></td>\n<td></td>\n<td>✔️ (domains &amp; structure optional)</td>\n<td></td>\n<td>✔️ (multimodal protein LLM + sequence encoders)</td>\n<td>✔️ (maps generated text to GO semantic terms)</td>\n<td>✔️ (text-generation based zero-shot explanations)</td>\n<td>✔️ (AlphaFold / ESMFold / PDB optional)</td>\n<td>✔️ (GNN + Transformer multimodal encoder)</td>\n<td>✔️ (generative functional text output)</td>\n<td>✔️ (optional PPI ingestion)</td>\n<td>✔️ (multimodal generation: text + graph + structure).</td>\n</tr>\n<tr>\n<td><strong>GoCurator (2023)</strong></td>\n<td></td>\n<td></td>\n<td></td>\n<td>✔️ (BioBERT / sentence-transformers for curation assistance)</td>\n<td>✔️ (curation maps literature evidence to GO terms)</td>\n<td>✔️ (literature-driven zero-shot candidate suggestions)</td>\n<td></td>\n<td></td>\n<td>✔️ (PubMed / GOA sentence extraction; PubTator-like tools)</td>\n<td></td>\n<td>✔️ (human-in-the-loop curation + automated retrieval).</td>\n</tr>\n</tbody>\n</table>",
      "rawMarkdown": "| **Software (year)** | **Sequence similarity / transfer** | **Domain signatures** | **Classic DL** | **Transformers / pLM** | **Use of GO ontology** | **Zero-shot** | **3D structure source** | **GNN / Graph Transformer** | **Text mining / text retrieval** | **PPI networks** | **Multimodal / hybrid** |\n| --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |\n| **DeepGOZero (2022)** |  | ✔️ (InterPro-derived features used as inputs) | ✔️ (feedforward/NN modules for scoring) | ✔️ (ontology & embedding pipeline; not pLM-centered) | ✔️ (ontology axioms → ontology embeddings; exploits GO axioms). | ✔️ (ontology-axiom zero-shot prediction). |  |  |  |  | ✔️ (sequence + ontology embeddings hybrid). |\n| **NetGO 3.0 (2023)** | ✔️ (homology features / BLAST-like signals included in ensemble) | ✔️ (InterProScan used as a component) |  | ✔️ (ESM-1b embeddings via LR-ESM component). | ✔️ (GO term supervision & hierarchy integrated in scoring). |  |  |  | ✔️ (functional text embeddings for some components). | ✔️ (STRING integrated for network features). | ✔️ (ensemble: sequence + pLM + network + domains). |\n| **TALE (2021)** |  |  |  | ✔️ (custom Transformer; joint sequence–label embedding) | ✔️ (label embeddings aligned to GO terms) | ✔️ (improves tail-label generalisation / few-shot) |  |  |  |  | ✔️ (sequence-only transformer with label embedding). |\n| **ProTranslator / ProTransfer (2022)** | ✔️ (can transfer via text→sequence similarity in practice) |  |  | ✔️ (text⇄sequence “translation” approach; NLP models) | ✔️ (relies on textual GO descriptions to align functions) | ✔️ (translation-based zero-shot to novel textual descriptors) |  |  | ✔️ (uses GO textual definitions / retrieval) |  | ✔️ (text-driven + sequence matching). |\n| **DeepFRI (2020)** |  |  | ✔️ (sequence LSTM/CNN features originally used) | ✔️ (uses sequence embeddings; later work uses pLM features) |  |  | ✔️ (PDB structures / modeled contact maps) | ✔️ (GCN on residue contact graph) |  |  | ✔️ (structure + sequence + GCN). |\n| **InterProScan (2005)** | ✔️ (signature-based mapping → annotation transfer) | ✔️ (Pfam, PROSITE, SMART, CATH-Gene3D, TIGRFAMs etc.) |  |  |  |  |  |  |  |  | ✔️ (classic signature-based pipeline used as component). |\n| **TransFun (2022)** |  | ✔️ (optional InterPro features) |  | ✔️ (ESM embeddings used) | ✔️ (predicts GO terms; uses GO supervision) |  | ✔️ (AlphaFold2 predicted structures / PDB when available) | ✔️ (EGNN — 3D-equivariant GNN) |  |  | ✔️ (sequence pLM + AlphaFold + EGNN ensemble). |\n| **ProtNLM (UniProt, 2024)** |  | ✔️ (UniProt pipeline includes InterPro features) |  | ✔️ (protein LLM to generate descriptive text / labels) | ✔️ (mapped to GO via UniProt rules/pipelines) | ✔️ (generative annotation can assign novel descriptions) |  |  | ✔️ (UniProt text-mining pipelines; integration with literature) | ✔️ (cross-DB links; UniProt integrates PPI sources) | ✔️ (LLM-generated descriptions + rule-based pipelines). |\n| **PFresGO (2023)** |  | ✔️ (InterPro integration mentioned) | ✔️ (attention-based DL for residue localization) | ✔️ (ProtTrans / ProtT5 embeddings often used) | ✔️ (explicit modelling of GO hierarchical relationships) |  |  | ✔️ (GO graph attention / cross-attention) |  |  | ✔️ (sequence embeddings + GO graph attention). |\n| **GOAnnotator (various, 2019– )** | ✔️ (can use BLAST/DIAMOND for evidence) | ✔️ (optionally uses InterProScan outputs) |  | ✔️ (sentence-transformers / SBERT for retrieval in modern versions) | ✔️ (maps extracted evidence to GO terms) | ✔️ (retrieval-based assignment to new terms) |  | ✔️ (optional graph fusion with PPI) | ✔️ (PubMed retrieval; PubTator / sentence-transformers / BM25 + rerank) | ✔️ (STRING / BioGRID integration optional) | ✔️ (text retrieval + sequence/domain evidence). |\n| **STAR-GO (2025, preprint)** |  | ✔️ (domain features optional) |  | ✔️ (Transformer + GO sentence embeddings) | ✔️ (joint semantic + structural GO embeddings) | ✔️ (designed for semantic zero-shot on GO terms) | ✔️ (evaluated with PDB / AlphaFold-based datasets) | ✔️ (GO graph integration; hierarchical propagation) | ✔️ (GO definitions via sentence encoders) |  | ✔️ (semantic GO embeddings + sequence alignment). |\n| **ProtFun (2025)** |  | ✔️ (InterPro signatures used to build protein family network) |  | ✔️ (uses ANKH protein LLM embeddings) | ✔️ (predicts GO labels supervised) |  |  | ✔️ (GAT on protein family network / PFN) |  | ✔️ (custom PFN; may include family-based links) | ✔️ (LLM embeddings + InterPro-based PFN + GAT). |\n| **DIAMOND2GO / Diamond2GO (2024)** | ✔️ (DIAMOND sequence similarity → transfer) |  |  |  |  |  |  |  |  |  | ✔️ (fast sequence-based GO transfer via DIAMOND; optional InterPro enrichment). |\n| **DPFunc (2024)** |  | ✔️ (InterProScan-guided domain features) | ✔️ (deep learning architecture focused on interpretability) | ✔️ (ESM-1b embeddings included) |  |  | ✔️ (PDB / AlphaFold2 structures) | ✔️ (GCN layers on residue/structure graph) |  | ✔️ (optional PPI datasets for evaluation) | ✔️ (domain-guided structural DL + GCN). |\n| **FANTASIA (2024)** | ✔️ (ProtT5 / ProtTrans embeddings used for similarity) | ✔️ (domain-aware clustering in pipeline) |  | ✔️ (ProtT5 / GOPredSim pipeline; embedding similarity) |  | ✔️ (embedding-similarity generalization for dark proteome) |  |  | ✔️ (embedding retrieval/search) |  | ✔️ (embedding-similarity GO term transfer at proteome scale). |\n| **ProtBoost (2024) – CAFA-5 top** | ✔️ (pLM embeddings used in ensemble) | ✔️ (InterPro features included in ensembles) |  | ✔️ (ESM / ProtTrans embeddings used) |  |  |  | ✔️ (GCN used in stacking/ensemble) | ✔️ (text/features optional in ensemble) | ✔️ (STRING / BioGRID optional) | ✔️ (Py-Boost + GCN ensemble; multi-source stacking). |\n| **AnnoPRO (2024)** |  | ✔️ (multiscale domain features included) | ✔️ (multi-scale DL modules for long-tail GO terms) | ✔️ (pretrained encoders used) | ✔️ (explicitly targets long-tail GO terms) |  |  | ✔️ (hybrid deep networks) | ✔️ (optional text features) | ✔️ (optional PPI features) | ✔️ (multiscale features + hybrid DL). |\n| **TEMPROT (2023)** | ✔️ (BLAST / homology used in ensemble versions) |  |  | ✔️ (ProtBERT-BFD embeddings) |  |  |  |  |  |  | ✔️ (transformer embeddings + homology ensemble). |\n| **Prot2Text-V2 (2025, arXiv)** |  | ✔️ (domains & structure optional) |  | ✔️ (multimodal protein LLM + sequence encoders) | ✔️ (maps generated text to GO semantic terms) | ✔️ (text-generation based zero-shot explanations) | ✔️ (AlphaFold / ESMFold / PDB optional) | ✔️ (GNN + Transformer multimodal encoder) | ✔️ (generative functional text output) | ✔️ (optional PPI ingestion) | ✔️ (multimodal generation: text + graph + structure). |\n| **GoCurator (2023)** |  |  |  | ✔️ (BioBERT / sentence-transformers for curation assistance) | ✔️ (curation maps literature evidence to GO terms) | ✔️ (literature-driven zero-shot candidate suggestions) |  |  | ✔️ (PubMed / GOA sentence extraction; PubTator-like tools) |  | ✔️ (human-in-the-loop curation + automated retrieval). |",
      "votes": null
    },
    {
      "id": "3381069",
      "postDate": "12/23/2025 17:18:33",
      "content": "<p>InterProScan performance on my VM is a major concern; it’s taking nearly 33 days to complete. This has become the biggest hurdle in my pipeline ;(</p>",
      "rawMarkdown": "InterProScan performance on my VM is a major concern; it’s taking nearly 33 days to complete. This has become the biggest hurdle in my pipeline ;(",
      "votes": null
    },
    {
      "id": "3392298",
      "postDate": "01/16/2026 16:54:48",
      "content": "<p>Why did your InterProScan run take so long? I was able to annotate using interpro and panther for the testsuperset in about half an hour on a standard workstation (64 GB RAM, Intel i5), so the runtime difference is quite surprising.</p>",
      "rawMarkdown": "Why did your InterProScan run take so long? I was able to annotate using interpro and panther for the testsuperset in about half an hour on a standard workstation (64 GB RAM, Intel i5), so the runtime difference is quite surprising.",
      "votes": null
    },
    {
      "id": "3392299",
      "postDate": "01/16/2026 16:56:42",
      "content": "<p>The newest version of FANTASIA supports multiple models (ProtT5, ProstT5, ESM2, ESM3c, and Ankh) and multiple embedding layers, not only ProtT5.</p>",
      "rawMarkdown": "The newest version of FANTASIA supports multiple models (ProtT5, ProstT5, ESM2, ESM3c, and Ankh) and multiple embedding layers, not only ProtT5.",
      "votes": null
    },
    {
      "id": "3392301",
      "postDate": "01/16/2026 17:03:25",
      "content": "<p><a href=\"https://www.kaggle.com/xaxipirulazo\" target=\"_blank\">@xaxipirulazo</a> I finished it in 4 days. I realized that I had used the --dp flag, which forces InterProScan to calculate everything from scratch, and that led to it taking 33 days.</p>\n<p>My workstation is an Intel Xeon Bronze with 32 GB RAM. Anyway, finishing InterProScan in 1 hour is very impressive. If possible, could you share the command you used to run InterProScan? I gave up on this competition, but I still want to try InterProScan on the last submission</p>",
      "rawMarkdown": "xaxipirulazo I finished it in 4 days. I realized that I had used the --dp flag, which forces InterProScan to calculate everything from scratch, and that led to it taking 33 days.\n\nMy workstation is an Intel Xeon Bronze with 32 GB RAM. Anyway, finishing InterProScan in 1 hour is very impressive. If possible, could you share the command you used to run InterProScan? I gave up on this competition, but I still want to try InterProScan on the last submission",
      "votes": null
    },
    {
      "id": "3392317",
      "postDate": "01/16/2026 17:34:59",
      "content": "<p>Hello,</p>\n<p>I no longer have the exact command in my history, so I can’t verify it precisely, but it was something along the lines of:</p>\n<pre><code>nextflow run ebi-pf-team/interproscan6 \\\n  -r 6.0.0 \\\n  -profile docker,test \\\n  --datadir data \\\n  --goterms \\\n  --input &lt;input_fasta&gt;\n</code></pre>\n<p>The first time I ran it with the sample FASTA, it took around 3–4 hours, mainly due to downloading the required data and performing the initial processing. After that initial run, however, the behavior surprised me quite a lot.</p>\n<p>Anyway, the run ended up using more than 32 GB of RAM. Running it with swap enabled would likely make it take much longer, if it doesn’t crash altogether.</p>\n<p>You made me wonder whether I might actually be running it in test mode.</p>",
      "rawMarkdown": "Hello,\n\nI no longer have the exact command in my history, so I can’t verify it precisely, but it was something along the lines of:\n\n```bash\nnextflow run ebi-pf-team/interproscan6 \\\n  -r 6.0.0 \\\n  -profile docker,test \\\n  --datadir data \\\n  --goterms \\\n  --input <input_fasta>\n```\n\nThe first time I ran it with the sample FASTA, it took around 3–4 hours, mainly due to downloading the required data and performing the initial processing. After that initial run, however, the behavior surprised me quite a lot.\n\nAnyway, the run ended up using more than 32 GB of RAM. Running it with swap enabled would likely make it take much longer, if it doesn’t crash altogether.\n\nYou made me wonder whether I might actually be running it in test mode.",
      "votes": null
    },
    {
      "id": "3392323",
      "postDate": "01/16/2026 17:45:04",
      "content": "<p><a href=\"https://www.kaggle.com/xaxipirulazo\" target=\"_blank\">@xaxipirulazo</a> thank for the command. I split the large FASTA file into smaller FASTA files (100 proteins each). After that, I ran InterProScan 5 on each FASTA file. I don’t remember the exact details, but it only used ~5 GB of RAM (maybe that’s why it took 4 days 😭). I’ll probably try running InterProScan again.</p>",
      "rawMarkdown": "xaxipirulazo thank for the command. I split the large FASTA file into smaller FASTA files (100 proteins each). After that, I ran InterProScan 5 on each FASTA file. I don’t remember the exact details, but it only used ~5 GB of RAM (maybe that’s why it took 4 days 😭). I’ll probably try running InterProScan again.",
      "votes": null
    },
    {
      "id": "3392857",
      "postDate": "01/17/2026 16:39:04",
      "content": "<p><a href=\"https://www.kaggle.com/nguyncdngs\" target=\"_blank\">@nguyncdngs</a> hey!, I did just relaunch it</p>\n<p><code>nextflow run ebi-pf-team/interproscan6   -r 6.0.0   -profile docker   --datadir data --input testsuperset.fasta --goterms</code></p>\n<p>Completed at: 17-Jan-2026 17:09:19\nDuration    : 32m 3s\nCPU hours   : 2.9\nSucceeded   : 183</p>",
      "rawMarkdown": "nguyncdngs hey!, I did just relaunch it\n\n`nextflow run ebi-pf-team/interproscan6   -r 6.0.0   -profile docker   --datadir data --input testsuperset.fasta --goterms`\n\nCompleted at: 17-Jan-2026 17:09:19\nDuration    : 32m 3s\nCPU hours   : 2.9\nSucceeded   : 183",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 3381069,
      "author_name": "nguyncdngs",
      "author_url": "",
      "post_date": "12/23/2025 17:18:33",
      "content": "<p>InterProScan performance on my VM is a major concern; it’s taking nearly 33 days to complete. This has become the biggest hurdle in my pipeline ;(</p>",
      "votes": null,
      "replies": [
        {
          "id": 3392298,
          "author_name": "xaxipirulazo",
          "author_url": "",
          "post_date": "01/16/2026 16:54:48",
          "content": "<p>Why did your InterProScan run take so long? I was able to annotate using interpro and panther for the testsuperset in about half an hour on a standard workstation (64 GB RAM, Intel i5), so the runtime difference is quite surprising.</p>",
          "votes": null,
          "replies": [
            {
              "id": 3392301,
              "author_name": "nguyncdngs",
              "author_url": "",
              "post_date": "01/16/2026 17:03:25",
              "content": "<p><a href=\"https://www.kaggle.com/xaxipirulazo\" target=\"_blank\">@xaxipirulazo</a> I finished it in 4 days. I realized that I had used the --dp flag, which forces InterProScan to calculate everything from scratch, and that led to it taking 33 days.</p>\n<p>My workstation is an Intel Xeon Bronze with 32 GB RAM. Anyway, finishing InterProScan in 1 hour is very impressive. If possible, could you share the command you used to run InterProScan? I gave up on this competition, but I still want to try InterProScan on the last submission</p>",
              "votes": null,
              "replies": [
                {
                  "id": 3392317,
                  "author_name": "xaxipirulazo",
                  "author_url": "",
                  "post_date": "01/16/2026 17:34:59",
                  "content": "<p>Hello,</p>\n<p>I no longer have the exact command in my history, so I can’t verify it precisely, but it was something along the lines of:</p>\n<pre><code>nextflow run ebi-pf-team/interproscan6 \\\n  -r 6.0.0 \\\n  -profile docker,test \\\n  --datadir data \\\n  --goterms \\\n  --input &lt;input_fasta&gt;\n</code></pre>\n<p>The first time I ran it with the sample FASTA, it took around 3–4 hours, mainly due to downloading the required data and performing the initial processing. After that initial run, however, the behavior surprised me quite a lot.</p>\n<p>Anyway, the run ended up using more than 32 GB of RAM. Running it with swap enabled would likely make it take much longer, if it doesn’t crash altogether.</p>\n<p>You made me wonder whether I might actually be running it in test mode.</p>",
                  "votes": null,
                  "replies": [
                    {
                      "id": 3392323,
                      "author_name": "nguyncdngs",
                      "author_url": "",
                      "post_date": "01/16/2026 17:45:04",
                      "content": "<p><a href=\"https://www.kaggle.com/xaxipirulazo\" target=\"_blank\">@xaxipirulazo</a> thank for the command. I split the large FASTA file into smaller FASTA files (100 proteins each). After that, I ran InterProScan 5 on each FASTA file. I don’t remember the exact details, but it only used ~5 GB of RAM (maybe that’s why it took 4 days 😭). I’ll probably try running InterProScan again.</p>",
                      "votes": null,
                      "replies": [
                        {
                          "id": 3392857,
                          "author_name": "xaxipirulazo",
                          "author_url": "",
                          "post_date": "01/17/2026 16:39:04",
                          "content": "<p><a href=\"https://www.kaggle.com/nguyncdngs\" target=\"_blank\">@nguyncdngs</a> hey!, I did just relaunch it</p>\n<p><code>nextflow run ebi-pf-team/interproscan6   -r 6.0.0   -profile docker   --datadir data --input testsuperset.fasta --goterms</code></p>\n<p>Completed at: 17-Jan-2026 17:09:19\nDuration    : 32m 3s\nCPU hours   : 2.9\nSucceeded   : 183</p>",
                          "votes": null,
                          "replies": []
                        }
                      ]
                    }
                  ]
                }
              ]
            }
          ]
        }
      ]
    },
    {
      "id": 3392299,
      "author_name": "xaxipirulazo",
      "author_url": "",
      "post_date": "01/16/2026 16:56:42",
      "content": "<p>The newest version of FANTASIA supports multiple models (ProtT5, ProstT5, ESM2, ESM3c, and Ankh) and multiple embedding layers, not only ProtT5.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "3380923": "| **Software (year)** | **Sequence similarity / transfer** | **Domain signatures** | **Classic DL** | **Transformers / pLM** | **Use of GO ontology** | **Zero-shot** | **3D structure source** | **GNN / Graph Transformer** | **Text mining / text retrieval** | **PPI networks** | **Multimodal / hybrid** |\n| --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |\n| **DeepGOZero (2022)** |  | ✔️ (InterPro-derived features used as inputs) | ✔️ (feedforward/NN modules for scoring) | ✔️ (ontology & embedding pipeline; not pLM-centered) | ✔️ (ontology axioms → ontology embeddings; exploits GO axioms). | ✔️ (ontology-axiom zero-shot prediction). |  |  |  |  | ✔️ (sequence + ontology embeddings hybrid). |\n| **NetGO 3.0 (2023)** | ✔️ (homology features / BLAST-like signals included in ensemble) | ✔️ (InterProScan used as a component) |  | ✔️ (ESM-1b embeddings via LR-ESM component). | ✔️ (GO term supervision & hierarchy integrated in scoring). |  |  |  | ✔️ (functional text embeddings for some components). | ✔️ (STRING integrated for network features). | ✔️ (ensemble: sequence + pLM + network + domains). |\n| **TALE (2021)** |  |  |  | ✔️ (custom Transformer; joint sequence–label embedding) | ✔️ (label embeddings aligned to GO terms) | ✔️ (improves tail-label generalisation / few-shot) |  |  |  |  | ✔️ (sequence-only transformer with label embedding). |\n| **ProTranslator / ProTransfer (2022)** | ✔️ (can transfer via text→sequence similarity in practice) |  |  | ✔️ (text⇄sequence “translation” approach; NLP models) | ✔️ (relies on textual GO descriptions to align functions) | ✔️ (translation-based zero-shot to novel textual descriptors) |  |  | ✔️ (uses GO textual definitions / retrieval) |  | ✔️ (text-driven + sequence matching). |\n| **DeepFRI (2020)** |  |  | ✔️ (sequence LSTM/CNN features originally used) | ✔️ (uses sequence embeddings; later work uses pLM features) |  |  | ✔️ (PDB structures / modeled contact maps) | ✔️ (GCN on residue contact graph) |  |  | ✔️ (structure + sequence + GCN). |\n| **InterProScan (2005)** | ✔️ (signature-based mapping → annotation transfer) | ✔️ (Pfam, PROSITE, SMART, CATH-Gene3D, TIGRFAMs etc.) |  |  |  |  |  |  |  |  | ✔️ (classic signature-based pipeline used as component). |\n| **TransFun (2022)** |  | ✔️ (optional InterPro features) |  | ✔️ (ESM embeddings used) | ✔️ (predicts GO terms; uses GO supervision) |  | ✔️ (AlphaFold2 predicted structures / PDB when available) | ✔️ (EGNN — 3D-equivariant GNN) |  |  | ✔️ (sequence pLM + AlphaFold + EGNN ensemble). |\n| **ProtNLM (UniProt, 2024)** |  | ✔️ (UniProt pipeline includes InterPro features) |  | ✔️ (protein LLM to generate descriptive text / labels) | ✔️ (mapped to GO via UniProt rules/pipelines) | ✔️ (generative annotation can assign novel descriptions) |  |  | ✔️ (UniProt text-mining pipelines; integration with literature) | ✔️ (cross-DB links; UniProt integrates PPI sources) | ✔️ (LLM-generated descriptions + rule-based pipelines). |\n| **PFresGO (2023)** |  | ✔️ (InterPro integration mentioned) | ✔️ (attention-based DL for residue localization) | ✔️ (ProtTrans / ProtT5 embeddings often used) | ✔️ (explicit modelling of GO hierarchical relationships) |  |  | ✔️ (GO graph attention / cross-attention) |  |  | ✔️ (sequence embeddings + GO graph attention). |\n| **GOAnnotator (various, 2019– )** | ✔️ (can use BLAST/DIAMOND for evidence) | ✔️ (optionally uses InterProScan outputs) |  | ✔️ (sentence-transformers / SBERT for retrieval in modern versions) | ✔️ (maps extracted evidence to GO terms) | ✔️ (retrieval-based assignment to new terms) |  | ✔️ (optional graph fusion with PPI) | ✔️ (PubMed retrieval; PubTator / sentence-transformers / BM25 + rerank) | ✔️ (STRING / BioGRID integration optional) | ✔️ (text retrieval + sequence/domain evidence). |\n| **STAR-GO (2025, preprint)** |  | ✔️ (domain features optional) |  | ✔️ (Transformer + GO sentence embeddings) | ✔️ (joint semantic + structural GO embeddings) | ✔️ (designed for semantic zero-shot on GO terms) | ✔️ (evaluated with PDB / AlphaFold-based datasets) | ✔️ (GO graph integration; hierarchical propagation) | ✔️ (GO definitions via sentence encoders) |  | ✔️ (semantic GO embeddings + sequence alignment). |\n| **ProtFun (2025)** |  | ✔️ (InterPro signatures used to build protein family network) |  | ✔️ (uses ANKH protein LLM embeddings) | ✔️ (predicts GO labels supervised) |  |  | ✔️ (GAT on protein family network / PFN) |  | ✔️ (custom PFN; may include family-based links) | ✔️ (LLM embeddings + InterPro-based PFN + GAT). |\n| **DIAMOND2GO / Diamond2GO (2024)** | ✔️ (DIAMOND sequence similarity → transfer) |  |  |  |  |  |  |  |  |  | ✔️ (fast sequence-based GO transfer via DIAMOND; optional InterPro enrichment). |\n| **DPFunc (2024)** |  | ✔️ (InterProScan-guided domain features) | ✔️ (deep learning architecture focused on interpretability) | ✔️ (ESM-1b embeddings included) |  |  | ✔️ (PDB / AlphaFold2 structures) | ✔️ (GCN layers on residue/structure graph) |  | ✔️ (optional PPI datasets for evaluation) | ✔️ (domain-guided structural DL + GCN). |\n| **FANTASIA (2024)** | ✔️ (ProtT5 / ProtTrans embeddings used for similarity) | ✔️ (domain-aware clustering in pipeline) |  | ✔️ (ProtT5 / GOPredSim pipeline; embedding similarity) |  | ✔️ (embedding-similarity generalization for dark proteome) |  |  | ✔️ (embedding retrieval/search) |  | ✔️ (embedding-similarity GO term transfer at proteome scale). |\n| **ProtBoost (2024) – CAFA-5 top** | ✔️ (pLM embeddings used in ensemble) | ✔️ (InterPro features included in ensembles) |  | ✔️ (ESM / ProtTrans embeddings used) |  |  |  | ✔️ (GCN used in stacking/ensemble) | ✔️ (text/features optional in ensemble) | ✔️ (STRING / BioGRID optional) | ✔️ (Py-Boost + GCN ensemble; multi-source stacking). |\n| **AnnoPRO (2024)** |  | ✔️ (multiscale domain features included) | ✔️ (multi-scale DL modules for long-tail GO terms) | ✔️ (pretrained encoders used) | ✔️ (explicitly targets long-tail GO terms) |  |  | ✔️ (hybrid deep networks) | ✔️ (optional text features) | ✔️ (optional PPI features) | ✔️ (multiscale features + hybrid DL). |\n| **TEMPROT (2023)** | ✔️ (BLAST / homology used in ensemble versions) |  |  | ✔️ (ProtBERT-BFD embeddings) |  |  |  |  |  |  | ✔️ (transformer embeddings + homology ensemble). |\n| **Prot2Text-V2 (2025, arXiv)** |  | ✔️ (domains & structure optional) |  | ✔️ (multimodal protein LLM + sequence encoders) | ✔️ (maps generated text to GO semantic terms) | ✔️ (text-generation based zero-shot explanations) | ✔️ (AlphaFold / ESMFold / PDB optional) | ✔️ (GNN + Transformer multimodal encoder) | ✔️ (generative functional text output) | ✔️ (optional PPI ingestion) | ✔️ (multimodal generation: text + graph + structure). |\n| **GoCurator (2023)** |  |  |  | ✔️ (BioBERT / sentence-transformers for curation assistance) | ✔️ (curation maps literature evidence to GO terms) | ✔️ (literature-driven zero-shot candidate suggestions) |  |  | ✔️ (PubMed / GOA sentence extraction; PubTator-like tools) |  | ✔️ (human-in-the-loop curation + automated retrieval). |",
    "3381069": "InterProScan performance on my VM is a major concern; it’s taking nearly 33 days to complete. This has become the biggest hurdle in my pipeline ;(",
    "3392298": "Why did your InterProScan run take so long? I was able to annotate using interpro and panther for the testsuperset in about half an hour on a standard workstation (64 GB RAM, Intel i5), so the runtime difference is quite surprising.",
    "3392299": "The newest version of FANTASIA supports multiple models (ProtT5, ProstT5, ESM2, ESM3c, and Ankh) and multiple embedding layers, not only ProtT5.",
    "3392301": "xaxipirulazo I finished it in 4 days. I realized that I had used the --dp flag, which forces InterProScan to calculate everything from scratch, and that led to it taking 33 days.\n\nMy workstation is an Intel Xeon Bronze with 32 GB RAM. Anyway, finishing InterProScan in 1 hour is very impressive. If possible, could you share the command you used to run InterProScan? I gave up on this competition, but I still want to try InterProScan on the last submission",
    "3392317": "Hello,\n\nI no longer have the exact command in my history, so I can’t verify it precisely, but it was something along the lines of:\n\n```bash\nnextflow run ebi-pf-team/interproscan6 \\\n  -r 6.0.0 \\\n  -profile docker,test \\\n  --datadir data \\\n  --goterms \\\n  --input <input_fasta>\n```\n\nThe first time I ran it with the sample FASTA, it took around 3–4 hours, mainly due to downloading the required data and performing the initial processing. After that initial run, however, the behavior surprised me quite a lot.\n\nAnyway, the run ended up using more than 32 GB of RAM. Running it with swap enabled would likely make it take much longer, if it doesn’t crash altogether.\n\nYou made me wonder whether I might actually be running it in test mode.",
    "3392323": "xaxipirulazo thank for the command. I split the large FASTA file into smaller FASTA files (100 proteins each). After that, I ran InterProScan 5 on each FASTA file. I don’t remember the exact details, but it only used ~5 GB of RAM (maybe that’s why it took 4 days 😭). I’ll probably try running InterProScan again.",
    "3392857": "nguyncdngs hey!, I did just relaunch it\n\n`nextflow run ebi-pf-team/interproscan6   -r 6.0.0   -profile docker   --datadir data --input testsuperset.fasta --goterms`\n\nCompleted at: 17-Jan-2026 17:09:19\nDuration    : 32m 3s\nCPU hours   : 2.9\nSucceeded   : 183"
  },
  "source": "meta"
}