[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M74M98AFX3QAASPB4BFTCBQG":3,"skill-same:01M74M98AFX3QAASPB4BFTCBQG":70},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":16,"categories":17,"keywords":18,"publishedAt":19,"linkOnly":20,"markdownContent":21,"markdownBody":22,"frontMatter":23,"refName":37,"commitSha":38,"skillDirectory":39,"forksCount":40,"repositoryHtmlUrl":41,"repositoryHomepageUrl":42,"ownerHtmlUrl":43,"ownerBlogUrl":6,"ownerDescription":44,"updatedAt":45,"install":46,"resources":59},"01M74M98AFX3QAASPB4BFTCBQG","esmfold2",null,"Biohub ESMFold2 \u002F ESMFold2-Fast all-atom co-folding (Candido et al. 2026, github.com\u002FBiohub\u002Fesm). Single-sequence and MSA modes; protein, DNA, RNA, ligand (CCD\u002FSMILES), modified residues. FoldBench Ab-Ag 50-55%, PPI 70-77% DockQ-pass. Also covers the ESMC-{300M,600M,6B} protein language models from the same release: masked-LM logits, hidden states, mutation scoring, contact prediction, and the SAE interpretability head. MIT-licensed weights on HuggingFace org `biohub`. Use this skill when: (1) Predicting complex structures with single-sequence input, (2) Validating designed binders with ESMFold2-Fast, (3) Running ESMFold2 with MSA input, (4) Getting ESMC embeddings or per-residue mutation scores, (5) Choosing kernel backend and sampling-step settings for paper-faithful throughput.","pengqianhan\u002FAI-Human-Research-OS","pengqianhan","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F4286531?v=4","01M74M3DT9HNZ6HPAK8TX69EX9","PengqianHan","01M74M3DTVSN2GXW0JH2W396FK",3,"Apache-2.0","https:\u002F\u002Fgithub.com\u002Fpengqianhan\u002FAI-Human-Research-OS\u002Fblob\u002Fmain\u002Fresearch-skills-hub\u002Fclaude-science-skills\u002Fskills\u002Fesmfold2\u002FSKILL.md",[],[],"2026-11-11 15:11:14",false,"---\nname: esmfold2\ndescription: >\n  Biohub ESMFold2 \u002F ESMFold2-Fast all-atom co-folding (Candido et al. 2026,\n  github.com\u002FBiohub\u002Fesm). Single-sequence and MSA modes; protein, DNA, RNA,\n  ligand (CCD\u002FSMILES), modified residues. FoldBench Ab-Ag 50-55%, PPI 70-77%\n  DockQ-pass. Also covers the ESMC-{300M,600M,6B} protein language models from\n  the same release: masked-LM logits, hidden states, mutation scoring, contact\n  prediction, and the SAE interpretability head. MIT-licensed weights on\n  HuggingFace org `biohub`. Use this skill when: (1) Predicting complex\n  structures with single-sequence input, (2) Validating designed binders with\n  ESMFold2-Fast, (3) Running ESMFold2 with MSA input, (4) Getting ESMC\n  embeddings or per-residue mutation scores, (5) Choosing kernel backend and\n  sampling-step settings for paper-faithful throughput.\n\nlicense: Apache-2.0\ncategory: biomodels\nrequirements: [gpu]\nmetadata:\n  display-name: ESMFold2\n  # SKILL.md body: \"**License:** MIT (code github.com\u002FBiohub\u002Fesm + weights HF\n  # `biohub\u002F*`)\"\n  # github.com\u002FBiohub\u002Fesm\u002Fblob\u002Fmain\u002FLICENSE.md: MIT (© 2026 Chan Zuckerberg\n  # Biohub, Inc.). verified 2026-06-30\n  third_party:\n    - kind: weights\n      name: ESMFold2 \u002F ESMC\n      provider: Biohub\n      license: MIT\n      terms_url: https:\u002F\u002Fgithub.com\u002FBiohub\u002Fesm\u002Fblob\u002Fmain\u002FLICENSE.md\n---\n\n# ESMFold2 (Biohub)\n\nAll-atom diffusion co-folding from the Biohub ESM release (2026). ESMFold2 =\n48 pair layers with MSA support; ESMFold2-Fast = 24 layers, single-sequence\nonly, ~1.7x faster.\n\n**License:** MIT (code github.com\u002FBiohub\u002Fesm + weights HF `biohub\u002F*`).\n**Paper:** \"Language Modeling Materializes a World Model of Protein Biology\" (2026).\n\n## Install\n\nCUDA 12.x GPU (H100\u002FA100-class); Python **3.12 only**. Fresh venv; needs\negress to HF Hub, GitHub, PyPI:\n\n```bash\npip install --no-cache-dir uv\nuv venv --python 3.12 \u002Fwork\u002Fvenv && source \u002Fwork\u002Fvenv\u002Fbin\u002Factivate\nuv pip install \\\n  \"torch>=2.5,\u003C2.8\" einops \"biotite>=1.0\" rdkit msgpack-numpy biopython \\\n  scikit-learn brotli attrs pandas cloudpathlib httpx tenacity zstd pydssp \\\n  pygtrie accelerate huggingface_hub safetensors \"numpy\u003C3\" networkx \\\n  sentencepiece tokenizers regex packaging filelock pyyaml typing_extensions \\\n  \"transformers @ git+https:\u002F\u002Fgithub.com\u002FBiohub\u002Ftransformers.git@3a8956fb4d4ea16b0ec8e71deef2c2909b6a5cbf\"\nuv pip install --no-deps \"esm @ git+https:\u002F\u002Fgithub.com\u002FBiohub\u002Fesm.git@f652b471\"\n# OPTIONAL — only affects ESMC attention; trunk speedup comes from set_kernel_backend(\"fused\")\nuv pip install ninja packaging wheel setuptools\nMAX_JOBS=8 uv pip install --no-deps --no-build-isolation \"flash-attn\u003C3\"\n# Do NOT install transformer-engine — RuntimeError (not ImportError) on import\n# slips ESMC's guard and kills ESMFold2Model import.\n```\n\nThe bundled `esmfold2_gpu` Modal env (remote-compute-modal skill) is the\ncanonical, version-pinned recipe.\n\n**Gotchas:**\n- **Default kernel backend is `None`** (reference PyTorch, ~12x slower than paper). Call `model.set_kernel_backend('fused')` after `from_pretrained()`. See section below.\n- Match torch CUDA build to your driver; the pin `\u003C2.8` targets CUDA 12.2.\n- Weights via Xet bridge ~300 MB\u002Fs: ESMFold2 1.36 GB, ESMFold2-Fast 0.76 GB. Set `HF_HOME=\u002Fwork\u002Fhf_cache`.\n\n## Usage — local model\n\n```python\nfrom esm.models.esmfold2 import (\n    ESMFold2InputBuilder, StructurePredictionInput,\n    ProteinInput, DNAInput, RNAInput, LigandInput, Modification,\n)\nfrom transformers.models.esmfold2.modeling_esmfold2 import ESMFold2Model\n\nmodel = ESMFold2Model.from_pretrained(\"biohub\u002FESMFold2\").cuda().eval()\n# or \"biohub\u002FESMFold2-Fast\" (24 layers, no MSA, ~1.7x faster)\n# or \"biohub\u002FESMFold2-Experimental{,-Fast}{,-Cutoff2025}\" (4 design-critic models)\n\nspi = StructurePredictionInput(sequences=[\n    ProteinInput(id=\"A\", sequence=target_seq),\n    ProteinInput(id=\"B\", sequence=binder_seq),\n    # DNAInput(id=\"C\", sequence=\"ACGT\", modifications=[Modification(position=5, ccd=\"C36\")]),\n    # RNAInput(id=\"D\", sequence=\"ACGU\"),\n    # LigandInput(id=\"L\", ccd=[\"SAH\"]),  # or smiles=\"...\"\n])\n# Homodimer: ProteinInput(id=[\"A\",\"B\"], sequence=seq)\n\nresults = ESMFold2InputBuilder().fold(\n    model, spi,\n    num_loops=10,             # paper FoldBench eval: 10; 20-loop variant: 20\n    num_sampling_steps=68,    # paper eval: 68 (truncated EDM)\n    num_diffusion_samples=5,  # paper eval: 5\u002Fseed\n    seed=0,\n)\n# fold() returns list[Prediction], one per diffusion sample. Each carries\n# .plddt [L], .ptm, .iptm, .pae [L,L], .pair_chains_iptm, .complex.to_mmcif().\n# Rank by ipTM for complexes \u002F mean pLDDT for monomers:\nbest = max(results, key=lambda r: float(r.iptm if r.iptm is not None\n                                        else r.plddt.mean()))\nopen(\"pred.cif\", \"w\").write(best.complex.to_mmcif())\n```\n\n**Paper-faithful FoldBench settings:** 10 loops, 68 sampling steps, 25 seeds\nx 5 diffusion samples; rank by ipTM (complexes) or pLDDT (monomers); MSA mode\nadds `msa_depth=1024` with 10% column masking and ESMC dropout 0.3.\n\n## Model variants on HF `biohub\u002F`\n\n| repo | size | pair layers | MSA | use |\n|---|---|---|---|---|\n| `ESMFold2` | 0.94 GB + ccd.pkl 0.42 GB | 48 | yes | full eval |\n| `ESMFold2-Fast` | 0.76 GB | 24 | no | fast single-seq |\n| `ESMFold2-Experimental{,-Fast}` | 0.90 \u002F 0.72 GB | 48 \u002F 24 | — | design search (Alg 11) |\n| `ESMFold2-Experimental{,-Fast}-Cutoff2025` | 0.90 \u002F 0.72 GB | — | — | design search + critic |\n| `ESMFold2-Experimental-Fast-base{300M,600M,6B}-step{250k..1500k}` | — | — | — | 15 critic ensemble |\n\n## Throughput: `set_kernel_backend(\"fused\")` is REQUIRED\n\n**Default is the slow path.** `ESMFold2Model.from_pretrained(...)` loads with\n`_kernel_backend=None` (reference PyTorch) and `chunk_size=64`. You MUST call:\n\n```python\nmodel = ESMFold2Model.from_pretrained(\"biohub\u002FESMFold2\").cuda().eval()\nmodel.set_kernel_backend(\"fused\")   # vendored Triton TriMul\u002FLN+SwiGLU\u002Fpair-bias kernels\nmodel.set_chunk_size(None)           # optimal & OOM-safe L\u003C=1024; use 256 above\n```\n\n`\"fused\"` gives ~1.5–6× trunk speedup over the reference backend, growing with\nL; end-to-end `fold()` is diffusion-bound at short L so fused breaks even\naround L≈300–400. Fused vs reference outputs are numerically consistent (pLDDT\nwithin noise). `\"fused\"` (Triton, bundled with the GPU torch wheel) is\n**inference-only** — auto-disables under backprop. Above ~L=1400\n(`chunk_size=128`) it hits illegal memory access — fall back to\n`set_kernel_backend(None)` + `set_chunk_size(64)`; validated through L=1024.\n\n**Do NOT use** `set_kernel_backend(\"cuequivariance\")`: the\n`cuequivariance-torch==0.10.0` wheel lacks the compiled ops and **silently\nfalls back** to the reference path. **`apply_torch_compile()`** is an\nalternative (NOT additive — call `set_kernel_backend(None)` first).\n\n## ESMFold2-Experimental* — design hook\n\nExperimental variants expose `res_type_soft` for gradient-guided design — see\n`references\u002Fdesign-hook.md`. Do NOT use the fused backend with them (fp32\u002Fbf16\ndtype crash; the reference path is correct).\n\n## Gotcha: cusolver SVD poison + structseq constructor\n\nThe Kabsch alignment in `modeling_esmfold2_common.py` calls\n`torch.linalg.svd(H32, driver=\"gesvd\")` on batched 3x3 matrices. NaN\u002FInf inputs\n(degenerate diffusion samples) corrupt the cusolver workspace — **all subsequent\nCUDA calls fail with \"illegal memory access\"**. Monkeypatch: redirect small\nbatched SVDs to CPU:\n\n```python\n_orig_svd = torch.linalg.svd\ndef _safe_svd(A, full_matrices=True, driver=None):\n    if A.is_cuda and A.shape[-1] \u003C= 4 and A.shape[-2] \u003C= 4:\n        Acpu = A.detach().float().cpu()\n        if not torch.isfinite(Acpu).all():\n            Acpu = torch.nan_to_num(Acpu, nan=0.0, posinf=1e6, neginf=-1e6)\n        out = _orig_svd(Acpu, full_matrices=full_matrices)\n        # torch.return_types.linalg_svd is a C structseq -> ctor takes ONE tuple.\n        return type(out)(tuple(t.to(A.device, A.dtype) for t in out))\n    return _orig_svd(A, full_matrices=full_matrices, driver=driver)\ntorch.linalg.svd = _safe_svd\n```\n\nNote `type(out)(tuple(...))`, **not** `type(out)(*(...))` — `torch.return_types.*` are\nC structseqs whose constructor takes a single tuple argument.\n\n## With-MSA mode\n\nESMFold2 supports per-chain MSA input via `ProteinInput(id, sequence, msa=MSA)`.\nThe `MSA` object lives at `esm.utils.msa.msa.MSA`:\n\n```python\nfrom esm.utils.msa.msa import MSA\n# ProteinInput, StructurePredictionInput as imported above\n\nmsa_A = MSA.from_a3m(\"\u002Fpath\u002Fchain_A.a3m\", max_sequences=2048)\nmsa_B = MSA.from_a3m(\"\u002Fpath\u002Fchain_B.a3m\", max_sequences=2048)\ninp = StructurePredictionInput(sequences=[\n    ProteinInput(id=\"A\", sequence=seq_A, msa=msa_A),\n    ProteinInput(id=\"B\", sequence=seq_B, msa=msa_B),\n])\n```\n\n**Gotchas:**\n- `MSA.from_a3m(remove_insertions=True)` asserts equal row lengths after\n  insertion removal. ColabFold a3m files often carry trailing **null bytes** and\n  off-by-one rows vs the query — `tr -d '\\000'` and force row 0 to the exact\n  query sequence (or `MSA.from_sequences` on manually cleaned, query-length rows).\n- **ESMFold2-Fast does NOT support MSA** (single-seq only).\n- With-MSA mode improves AbAg interface pass-rate per the paper's evaluation.\n\n## Paper-matched inference configuration\n\nThe paper's FoldBench protocol (section A.2.11):\n\n| Parameter | Paper default | Paper \"20lp\" | Notes |\n|---|---|---|---|\n| `num_loops` (folding-trunk recycles) | **10** | 20 | +2pp on AbAg |\n| `num_sampling_steps` (diffusion) | **68** | 68 | EDM-tuned; do **NOT** use 200 |\n| seeds x diffusion samples | 25 x 5 | 25 x 5 | Fig S6\u002FS7 oracle = best-of-125 |\n\n## Training data cutoff\n\nESMFold2 and ESMFold2-Fast both use a **Sept 2021** PDB training cutoff (HF\n`biohub\u002FESMFold2` README).\n\n## ESMC language model\n\nESMC is the Biohub successor to ESM-2; three sizes: 300M (30L), 600M (36L),\n6B (80L, d=2560). HF path: `AutoModelForMaskedLM.from_pretrained(\"biohub\u002FESMC-6B\")`.\n\n**Mask token is `\u003Cmask>`** (id 32) — use `tok.mask_token`. The native-SDK `_`\nconvention does NOT apply to the HF tokenizer: `_` is not in the vocab and\nencodes to `\u003Cunk>`, silently corrupting mutation scores.\n\nFull API, mutation scoring, SAE features, contact prediction: see\n`references\u002Fesmc.md`.\n","# ESMFold2 (Biohub)\n\nAll-atom diffusion co-folding from the Biohub ESM release (2026). ESMFold2 =\n48 pair layers with MSA support; ESMFold2-Fast = 24 layers, single-sequence\nonly, ~1.7x faster.\n\n**License:** MIT (code github.com\u002FBiohub\u002Fesm + weights HF `biohub\u002F*`).\n**Paper:** \"Language Modeling Materializes a World Model of Protein Biology\" (2026).\n\n## Install\n\nCUDA 12.x GPU (H100\u002FA100-class); Python **3.12 only**. Fresh venv; needs\negress to HF Hub, GitHub, PyPI:\n\n```bash\npip install --no-cache-dir uv\nuv venv --python 3.12 \u002Fwork\u002Fvenv && source \u002Fwork\u002Fvenv\u002Fbin\u002Factivate\nuv pip install \\\n  \"torch>=2.5,\u003C2.8\" einops \"biotite>=1.0\" rdkit msgpack-numpy biopython \\\n  scikit-learn brotli attrs pandas cloudpathlib httpx tenacity zstd pydssp \\\n  pygtrie accelerate huggingface_hub safetensors \"numpy\u003C3\" networkx \\\n  sentencepiece tokenizers regex packaging filelock pyyaml typing_extensions \\\n  \"transformers @ git+https:\u002F\u002Fgithub.com\u002FBiohub\u002Ftransformers.git@3a8956fb4d4ea16b0ec8e71deef2c2909b6a5cbf\"\nuv pip install --no-deps \"esm @ git+https:\u002F\u002Fgithub.com\u002FBiohub\u002Fesm.git@f652b471\"\n# OPTIONAL — only affects ESMC attention; trunk speedup comes from set_kernel_backend(\"fused\")\nuv pip install ninja packaging wheel setuptools\nMAX_JOBS=8 uv pip install --no-deps --no-build-isolation \"flash-attn\u003C3\"\n# Do NOT install transformer-engine — RuntimeError (not ImportError) on import\n# slips ESMC's guard and kills ESMFold2Model import.\n```\n\nThe bundled `esmfold2_gpu` Modal env (remote-compute-modal skill) is the\ncanonical, version-pinned recipe.\n\n**Gotchas:**\n- **Default kernel backend is `None`** (reference PyTorch, ~12x slower than paper). Call `model.set_kernel_backend('fused')` after `from_pretrained()`. See section below.\n- Match torch CUDA build to your driver; the pin `\u003C2.8` targets CUDA 12.2.\n- Weights via Xet bridge ~300 MB\u002Fs: ESMFold2 1.36 GB, ESMFold2-Fast 0.76 GB. Set `HF_HOME=\u002Fwork\u002Fhf_cache`.\n\n## Usage — local model\n\n```python\nfrom esm.models.esmfold2 import (\n    ESMFold2InputBuilder, StructurePredictionInput,\n    ProteinInput, DNAInput, RNAInput, LigandInput, Modification,\n)\nfrom transformers.models.esmfold2.modeling_esmfold2 import ESMFold2Model\n\nmodel = ESMFold2Model.from_pretrained(\"biohub\u002FESMFold2\").cuda().eval()\n# or \"biohub\u002FESMFold2-Fast\" (24 layers, no MSA, ~1.7x faster)\n# or \"biohub\u002FESMFold2-Experimental{,-Fast}{,-Cutoff2025}\" (4 design-critic models)\n\nspi = StructurePredictionInput(sequences=[\n    ProteinInput(id=\"A\", sequence=target_seq),\n    ProteinInput(id=\"B\", sequence=binder_seq),\n    # DNAInput(id=\"C\", sequence=\"ACGT\", modifications=[Modification(position=5, ccd=\"C36\")]),\n    # RNAInput(id=\"D\", sequence=\"ACGU\"),\n    # LigandInput(id=\"L\", ccd=[\"SAH\"]),  # or smiles=\"...\"\n])\n# Homodimer: ProteinInput(id=[\"A\",\"B\"], sequence=seq)\n\nresults = ESMFold2InputBuilder().fold(\n    model, spi,\n    num_loops=10,             # paper FoldBench eval: 10; 20-loop variant: 20\n    num_sampling_steps=68,    # paper eval: 68 (truncated EDM)\n    num_diffusion_samples=5,  # paper eval: 5\u002Fseed\n    seed=0,\n)\n# fold() returns list[Prediction], one per diffusion sample. Each carries\n# .plddt [L], .ptm, .iptm, .pae [L,L], .pair_chains_iptm, .complex.to_mmcif().\n# Rank by ipTM for complexes \u002F mean pLDDT for monomers:\nbest = max(results, key=lambda r: float(r.iptm if r.iptm is not None\n                                        else r.plddt.mean()))\nopen(\"pred.cif\", \"w\").write(best.complex.to_mmcif())\n```\n\n**Paper-faithful FoldBench settings:** 10 loops, 68 sampling steps, 25 seeds\nx 5 diffusion samples; rank by ipTM (complexes) or pLDDT (monomers); MSA mode\nadds `msa_depth=1024` with 10% column masking and ESMC dropout 0.3.\n\n## Model variants on HF `biohub\u002F`\n\n| repo | size | pair layers | MSA | use |\n|---|---|---|---|---|\n| `ESMFold2` | 0.94 GB + ccd.pkl 0.42 GB | 48 | yes | full eval |\n| `ESMFold2-Fast` | 0.76 GB | 24 | no | fast single-seq |\n| `ESMFold2-Experimental{,-Fast}` | 0.90 \u002F 0.72 GB | 48 \u002F 24 | — | design search (Alg 11) |\n| `ESMFold2-Experimental{,-Fast}-Cutoff2025` | 0.90 \u002F 0.72 GB | — | — | design search + critic |\n| `ESMFold2-Experimental-Fast-base{300M,600M,6B}-step{250k..1500k}` | — | — | — | 15 critic ensemble |\n\n## Throughput: `set_kernel_backend(\"fused\")` is REQUIRED\n\n**Default is the slow path.** `ESMFold2Model.from_pretrained(...)` loads with\n`_kernel_backend=None` (reference PyTorch) and `chunk_size=64`. You MUST call:\n\n```python\nmodel = ESMFold2Model.from_pretrained(\"biohub\u002FESMFold2\").cuda().eval()\nmodel.set_kernel_backend(\"fused\")   # vendored Triton TriMul\u002FLN+SwiGLU\u002Fpair-bias kernels\nmodel.set_chunk_size(None)           # optimal & OOM-safe L\u003C=1024; use 256 above\n```\n\n`\"fused\"` gives ~1.5–6× trunk speedup over the reference backend, growing with\nL; end-to-end `fold()` is diffusion-bound at short L so fused breaks even\naround L≈300–400. Fused vs reference outputs are numerically consistent (pLDDT\nwithin noise). `\"fused\"` (Triton, bundled with the GPU torch wheel) is\n**inference-only** — auto-disables under backprop. Above ~L=1400\n(`chunk_size=128`) it hits illegal memory access — fall back to\n`set_kernel_backend(None)` + `set_chunk_size(64)`; validated through L=1024.\n\n**Do NOT use** `set_kernel_backend(\"cuequivariance\")`: the\n`cuequivariance-torch==0.10.0` wheel lacks the compiled ops and **silently\nfalls back** to the reference path. **`apply_torch_compile()`** is an\nalternative (NOT additive — call `set_kernel_backend(None)` first).\n\n## ESMFold2-Experimental* — design hook\n\nExperimental variants expose `res_type_soft` for gradient-guided design — see\n`references\u002Fdesign-hook.md`. Do NOT use the fused backend with them (fp32\u002Fbf16\ndtype crash; the reference path is correct).\n\n## Gotcha: cusolver SVD poison + structseq constructor\n\nThe Kabsch alignment in `modeling_esmfold2_common.py` calls\n`torch.linalg.svd(H32, driver=\"gesvd\")` on batched 3x3 matrices. NaN\u002FInf inputs\n(degenerate diffusion samples) corrupt the cusolver workspace — **all subsequent\nCUDA calls fail with \"illegal memory access\"**. Monkeypatch: redirect small\nbatched SVDs to CPU:\n\n```python\n_orig_svd = torch.linalg.svd\ndef _safe_svd(A, full_matrices=True, driver=None):\n    if A.is_cuda and A.shape[-1] \u003C= 4 and A.shape[-2] \u003C= 4:\n        Acpu = A.detach().float().cpu()\n        if not torch.isfinite(Acpu).all():\n            Acpu = torch.nan_to_num(Acpu, nan=0.0, posinf=1e6, neginf=-1e6)\n        out = _orig_svd(Acpu, full_matrices=full_matrices)\n        # torch.return_types.linalg_svd is a C structseq -> ctor takes ONE tuple.\n        return type(out)(tuple(t.to(A.device, A.dtype) for t in out))\n    return _orig_svd(A, full_matrices=full_matrices, driver=driver)\ntorch.linalg.svd = _safe_svd\n```\n\nNote `type(out)(tuple(...))`, **not** `type(out)(*(...))` — `torch.return_types.*` are\nC structseqs whose constructor takes a single tuple argument.\n\n## With-MSA mode\n\nESMFold2 supports per-chain MSA input via `ProteinInput(id, sequence, msa=MSA)`.\nThe `MSA` object lives at `esm.utils.msa.msa.MSA`:\n\n```python\nfrom esm.utils.msa.msa import MSA\n# ProteinInput, StructurePredictionInput as imported above\n\nmsa_A = MSA.from_a3m(\"\u002Fpath\u002Fchain_A.a3m\", max_sequences=2048)\nmsa_B = MSA.from_a3m(\"\u002Fpath\u002Fchain_B.a3m\", max_sequences=2048)\ninp = StructurePredictionInput(sequences=[\n    ProteinInput(id=\"A\", sequence=seq_A, msa=msa_A),\n    ProteinInput(id=\"B\", sequence=seq_B, msa=msa_B),\n])\n```\n\n**Gotchas:**\n- `MSA.from_a3m(remove_insertions=True)` asserts equal row lengths after\n  insertion removal. ColabFold a3m files often carry trailing **null bytes** and\n  off-by-one rows vs the query — `tr -d '\\000'` and force row 0 to the exact\n  query sequence (or `MSA.from_sequences` on manually cleaned, query-length rows).\n- **ESMFold2-Fast does NOT support MSA** (single-seq only).\n- With-MSA mode improves AbAg interface pass-rate per the paper's evaluation.\n\n## Paper-matched inference configuration\n\nThe paper's FoldBench protocol (section A.2.11):\n\n| Parameter | Paper default | Paper \"20lp\" | Notes |\n|---|---|---|---|\n| `num_loops` (folding-trunk recycles) | **10** | 20 | +2pp on AbAg |\n| `num_sampling_steps` (diffusion) | **68** | 68 | EDM-tuned; do **NOT** use 200 |\n| seeds x diffusion samples | 25 x 5 | 25 x 5 | Fig S6\u002FS7 oracle = best-of-125 |\n\n## Training data cutoff\n\nESMFold2 and ESMFold2-Fast both use a **Sept 2021** PDB training cutoff (HF\n`biohub\u002FESMFold2` README).\n\n## ESMC language model\n\nESMC is the Biohub successor to ESM-2; three sizes: 300M (30L), 600M (36L),\n6B (80L, d=2560). HF path: `AutoModelForMaskedLM.from_pretrained(\"biohub\u002FESMC-6B\")`.\n\n**Mask token is `\u003Cmask>`** (id 32) — use `tok.mask_token`. The native-SDK `_`\nconvention does NOT apply to the HF tokenizer: `_` is not in the vocab and\nencodes to `\u003Cunk>`, silently corrupting mutation scores.\n\nFull API, mutation scoring, SAE features, contact prediction: see\n`references\u002Fesmc.md`.\n",{"name":5,"license":15,"category":24,"metadata":25,"description":34,"requirements":35},"biomodels",{"third_party":26,"display-name":33},[27],{"kind":28,"name":29,"license":30,"provider":31,"terms_url":32},"weights","ESMFold2 \u002F ESMC","MIT","Biohub","https:\u002F\u002Fgithub.com\u002FBiohub\u002Fesm\u002Fblob\u002Fmain\u002FLICENSE.md","ESMFold2","Biohub ESMFold2 \u002F ESMFold2-Fast all-atom co-folding (Candido et al. 2026, github.com\u002FBiohub\u002Fesm). Single-sequence and MSA modes; protein, DNA, RNA, ligand (CCD\u002FSMILES), modified residues. FoldBench Ab-Ag 50-55%, PPI 70-77% DockQ-pass. Also covers the ESMC-{300M,600M,6B} protein language models from the same release: masked-LM logits, hidden states, mutation scoring, contact prediction, and the SAE interpretability head. MIT-licensed weights on HuggingFace org `biohub`. Use this skill when: (1) Predicting complex structures with single-sequence input, (2) Validating designed binders with ESMFold2-Fast, (3) Running ESMFold2 with MSA input, (4) Getting ESMC embeddings or per-residue mutation scores, (5) Choosing kernel backend and sampling-step settings for paper-faithful throughput.\n",[36],"gpu","main","e71245365df78e06cd166602744bf2491f79910e","research-skills-hub\u002Fclaude-science-skills\u002Fskills\u002Fesmfold2",0,"https:\u002F\u002Fgithub.com\u002Fpengqianhan\u002FAI-Human-Research-OS","https:\u002F\u002Fpengqianhan.github.io\u002FAI-Human-Research-OS\u002F","https:\u002F\u002Fgithub.com\u002Fpengqianhan","Ph.D. candidate of The University of Auckland, New Zealand","2026-11-14 18:08:16",{"source":8,"skillName":5,"commands":47,"prompt":57,"zipAvailable":58},[48,51,54],{"tool":49,"command":50},"npx","npx skills add pengqianhan\u002FAI-Human-Research-OS --skill esmfold2",{"tool":52,"command":53},"bunx","bunx skills add pengqianhan\u002FAI-Human-Research-OS --skill esmfold2",{"tool":55,"command":56},"pnpm","pnpm dlx skills add pengqianhan\u002FAI-Human-Research-OS --skill esmfold2","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002Fpengqianhan\u002FAI-Human-Research-OS\u002Fblob\u002Fmain\u002Fresearch-skills-hub\u002Fclaude-science-skills\u002Fskills\u002Fesmfold2\u002FSKILL.md",true,[60,66],{"path":61,"resourceType":62,"resourceTypeDesc":63,"sizeBytes":64,"sourceHtmlUrl":65},"references\u002Fdesign-hook.md","REFERENCE","参考文档",1452,"https:\u002F\u002Fgithub.com\u002Fpengqianhan\u002FAI-Human-Research-OS\u002Fblob\u002Fmain\u002Fresearch-skills-hub\u002Fclaude-science-skills\u002Fskills\u002Fesmfold2\u002Freferences\u002Fdesign-hook.md",{"path":67,"resourceType":62,"resourceTypeDesc":63,"sizeBytes":68,"sourceHtmlUrl":69},"references\u002Fesmc.md",3052,"https:\u002F\u002Fgithub.com\u002Fpengqianhan\u002FAI-Human-Research-OS\u002Fblob\u002Fmain\u002Fresearch-skills-hub\u002Fclaude-science-skills\u002Fskills\u002Fesmfold2\u002Freferences\u002Fesmc.md",[71,78,85,92,99,106],{"skillNo":72,"name":73,"nameZh":6,"description":74,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":30,"licenseFileUrl":6,"sourceHtmlUrl":75,"categories":76,"keywords":77,"publishedAt":19},"01M74MWYTHT498SVMN823896T7","workflow-skill-creator","Distills a completed user workflow or interaction into a reusable agent skill. Use when the user asks to turn their workflow, interaction, or multi-step process into a skill, or when they say \"make this a skill\", \"create a skill from what we just did\", \"package this workflow\" or similar. Do not use for creating skills from scratch without an existing workflow (use a generic skill-creator for that).","https:\u002F\u002Fgithub.com\u002Fpengqianhan\u002FAI-Human-Research-OS\u002Fblob\u002Fmain\u002Fresearch-skills-hub\u002Fscience-skills\u002Fworkflow_skill_creator\u002FSKILL.md",[],[],{"skillNo":79,"name":80,"nameZh":6,"description":81,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":30,"licenseFileUrl":6,"sourceHtmlUrl":82,"categories":83,"keywords":84,"publishedAt":19},"01M74MWWVCFZ8HQS8840F1V1FT","uv","Checks whether the uv Python package manager is installed and installs it if missing. Ensures uv is on PATH. Use when another skill requires uv as a prerequisite.","https:\u002F\u002Fgithub.com\u002Fpengqianhan\u002FAI-Human-Research-OS\u002Fblob\u002Fmain\u002Fresearch-skills-hub\u002Fscience-skills\u002Fuv\u002FSKILL.md",[],[],{"skillNo":86,"name":87,"nameZh":6,"description":88,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":30,"licenseFileUrl":6,"sourceHtmlUrl":89,"categories":90,"keywords":91,"publishedAt":19},"01M74MWRH5BTSWZ2C5PCTD5FV9","uniprot-database","Access protein metadata, function, taxonomy, and sequences across UniProtKB, UniParc, and UniRef. Use when searching for proteins, mapping identifiers, or retrieving functional annotations and publications. Don't use for sequence alignment, protein folding, or sequence similarity search (use specialized skills for those tasks).","https:\u002F\u002Fgithub.com\u002Fpengqianhan\u002FAI-Human-Research-OS\u002Fblob\u002Fmain\u002Fresearch-skills-hub\u002Fscience-skills\u002Funiprot_database\u002FSKILL.md",[],[],{"skillNo":93,"name":94,"nameZh":6,"description":95,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":30,"licenseFileUrl":6,"sourceHtmlUrl":96,"categories":97,"keywords":98,"publishedAt":19},"01M74MWN1D3ER9F1FZPWH8HQC1","unibind-database","Queries the UniBind database for experimentally validated transcription factor (TF) binding sites. Use when retrieving direct TF-DNA interaction datasets, downloading binding site coordinates (BED\u002FFASTA) for local analysis, or listing available datasets by species, cell line, or TF name. Don't use to query specific intervals, locations, genes, motif models or expression data.","https:\u002F\u002Fgithub.com\u002Fpengqianhan\u002FAI-Human-Research-OS\u002Fblob\u002Fmain\u002Fresearch-skills-hub\u002Fscience-skills\u002Funibind_database\u002FSKILL.md",[],[],{"skillNo":100,"name":101,"nameZh":6,"description":102,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":30,"licenseFileUrl":6,"sourceHtmlUrl":103,"categories":104,"keywords":105,"publishedAt":19},"01M74MWHHB5QZY3NR2FYA1EK4N","ucsc-conservation-and-tfbs","Fetch Evolutionary Conservation scores (phyloP, phastCons) and Transcription Factor Binding Sites (TFBS) from the UCSC Genome Browser. Use when analyzing whether genomic variants or regions are evolutionarily conserved, functionally important, or bounded by TF regulators across major projects (ENCODE, JASPAR, ReMap).","https:\u002F\u002Fgithub.com\u002Fpengqianhan\u002FAI-Human-Research-OS\u002Fblob\u002Fmain\u002Fresearch-skills-hub\u002Fscience-skills\u002Fucsc_conservation_and_tfbs\u002FSKILL.md",[],[],{"skillNo":107,"name":108,"nameZh":6,"description":109,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":30,"licenseFileUrl":6,"sourceHtmlUrl":110,"categories":111,"keywords":112,"publishedAt":19},"01M74MWERYHSY87P21FDMVE6BA","string-database","Query the STRING database for protein-protein interactions (PPIs), functional enrichment, and homology. Use when the user asks about interactions between specific proteins, interaction evidence, confidence scores, protein interaction partners, or pathway enrichments.","https:\u002F\u002Fgithub.com\u002Fpengqianhan\u002FAI-Human-Research-OS\u002Fblob\u002Fmain\u002Fresearch-skills-hub\u002Fscience-skills\u002Fstring_database\u002FSKILL.md",[],[]]