{
  "paper_doi": "10.48550/arXiv.2509.18480",
  "paper_title": "SimpleFold: Folding Proteins is Simpler than You Think",
  "paper_year": 2025,
  "headline_metric": {
    "name": "CASP14 SimpleFold-3B median LDDT",
    "value": 0.709,
    "tolerance_pct": 10,
    "units": "LDDT"
  },
  "data_sources": [
    "PDB experimental data",
    "AFDB SwissProt",
    "AFESM",
    "AFESM-E",
    "CAMEO22",
    "CASP14"
  ],
  "models": [
    {
      "name": "SimpleFold-3B",
      "weights_url": "https://ml-site.cdn-apple.com/models/simplefold/simplefold_3B.ckpt",
      "code_url": "https://github.com/apple/ml-simplefold",
      "apple_silicon_path": "native_mlx"
    }
  ],
  "evaluation_set": [
    "T1024",
    "T1025",
    "T1026",
    "T1027",
    "T1028",
    "T1029",
    "T1030",
    "T1031",
    "T1032",
    "T1033",
    "T1034",
    "T1035",
    "T1036s1",
    "T1037",
    "T1038",
    "T1039",
    "T1040",
    "T1041",
    "T1042",
    "T1043",
    "T1045s1",
    "T1045s2",
    "T1046s1",
    "T1046s2",
    "T1047s1",
    "T1047s2",
    "T1048",
    "T1049",
    "T1050",
    "T1052",
    "T1053",
    "T1054",
    "T1055",
    "T1056",
    "T1057",
    "T1058",
    "T1060s2",
    "T1060s3",
    "T1061",
    "T1062",
    "T1064",
    "T1065s1",
    "T1065s2",
    "T1067",
    "T1068",
    "T1070",
    "T1072s1",
    "T1073",
    "T1074",
    "T1076",
    "T1078",
    "T1079",
    "T1080",
    "T1082",
    "T1083",
    "T1084",
    "T1085",
    "T1087",
    "T1088",
    "T1089",
    "T1090",
    "T1091",
    "T1092",
    "T1093",
    "T1094",
    "T1095",
    "T1096",
    "T1099",
    "T1100",
    "T1101"
  ],
  "evaluation_metric": "LDDT",
  "quotes": [
    {
      "field": "paper_title",
      "text": "SimpleFold: Folding Proteins is Simpler than You Think",
      "page": 1
    },
    {
      "field": "paper_doi",
      "text": "https://doi.org/10.48550/arXiv.2509.18480",
      "page": 1
    },
    {
      "field": "models.name",
      "text": "We scale SimpleFold to 3B parameters",
      "page": 1
    },
    {
      "field": "models.code_url",
      "text": "Code: https://github.com/apple/ml-simplefold",
      "page": 1
    },
    {
      "field": "models.apple_silicon_path",
      "text": "Due to its general-purpose architecture, SimpleFold shows efficiency in deployment and inference on consumer-level hardware.",
      "page": 1
    },
    {
      "field": "models.apple_silicon_path",
      "text": "We provide support for both PyTorch and MLX (recommended for Apple hardware) backends in inference.",
      "page": 0
    },
    {
      "field": "data_sources",
      "text": "We train SimpleFold with a data mix of 3 different sources.",
      "page": 6
    },
    {
      "field": "data_sources",
      "text": "we train our largest SimpleFold-3B on the distilled AFESM-E data together with PDB and SwissProt.",
      "page": 6
    },
    {
      "field": "evaluation_set",
      "text": "We evaluate SimpleFold on two widely adopted protein structure prediction benchmarks: CAMEO22 and CASP14",
      "page": 8
    },
    {
      "field": "evaluation_set",
      "text": "List of 70 targets in CASP14",
      "page": 24
    },
    {
      "field": "evaluation_metric",
      "text": "We report standard structure prediction metrics: TM-score and GDT-TS assess global structural similarity; LDDT and LDDT-Ca measure local atomic accuracy",
      "page": 8
    },
    {
      "field": "headline_metric",
      "text": "SimpleFold-3B                      0.720 / 0.792   0.639 / 0.703   0.666 / 0.709",
      "page": 9
    },
    {
      "field": "manual_curation_status",
      "text": "MANUAL_CURATED_READER_OUTPUT: produced by pilot/local-runner/scripts/manual-simplefold-reader.mjs from primary arXiv PDF text plus official Apple repository README; use until command-agent credentials are available.",
      "page": 0
    }
  ]
}