--- license: other license_name: embl-ebi-terms-of-use license_link: https://www.ebi.ac.uk/about/terms-of-use/ task_categories: - feature-extraction tags: - proteomics - mass-spectrometry - peptide-sequencing pretty_name: InstaNovo-FM training corpus configs: # Fully explicit names: no bare tier name exists, so a config can never be # ambiguous about which representation it refers to. # # splits/ uses data_dir because the filenames carry train/validation/test and # datasets detects them. by_project/ must NOT: with data_dir, split detection # runs inside the directory, finds no split keyword in the run names, and # collapses everything into a split called "train" -- advertising a training # split that does not exist. data_files with `split: full` names it honestly. - config_name: hcfm_splits default: true data_dir: splits/hcfm - config_name: mcfm_splits data_dir: splits/mcfm - config_name: lcfm_splits data_dir: splits/lcfm - config_name: hcfm_by_project data_files: - split: full path: by_project/hcfm/**/*.parquet - config_name: mcfm_by_project data_files: - split: full path: by_project/mcfm/**/*.parquet - config_name: lcfm_by_project data_files: - split: full path: by_project/lcfm/**/*.parquet - config_name: peptide_registry data_files: - split: full path: peptide_registry.parquet --- # InstaNovo-FM training corpus Tandem mass spectra with peptide-spectrum-match labels, uniformly reprocessed from public PRIDE submissions, used to pretrain and evaluate InstaNovo-FM. ## Layout Spectra are grouped into three **confidence tiers**. Each is a Foundational Model dataset named for how stringently its peptide-spectrum matches (PSMs) were filtered: **LCFM** (Low), **MCFM** (Medium) and **HCFM** (High). The names are relative, not absolute — every labelled tier consists of high-confidence PSMs, and "low confidence" marks LCFM as the broadest, least stringently filtered one, from which the stricter subsets are derived. They are nested: HCFM ⊂ MCFM ⊂ LCFM. A fourth tier, **ACFM** (All Confidence), is the unlabelled superset and is not published here. Each tier is published twice, once under `splits/` and once under `by_project/`. Which you want depends on whether you are consuming the corpus or rebuilding it: - **`splits/`** — the train/validation/test partitions the model actually consumed: quality-filtered, shuffled, and peptide-disjoint. Take this to reproduce or extend the published results. - **`by_project/`** — the same tier *before* filtering and splitting, one directory per PRIDE accession. Take this to apply your own quality criteria or derive your own partitions, which `splits/` cannot support because the filtering is lossy. The tier sits *inside* the folder rather than above it — `splits/lcfm/`, not `lcfm/splits/`. That nesting is deliberate: it means a download pattern like `--include "by_project/*"` cannot stray outside the folder you named. Nested the other way, `lcfm/*` would have matched both folders at once and quietly handed you ~900 GB containing two overlapping copies of the same spectra, one filtered and one not. ``` InstaDeepAI/InstaNovo │ ├── splits/ FILTERED · SHUFFLED · PEPTIDE-DISJOINT │ ├── hcfm/ 11 files · 11.4 GB · 3,670,113 rows │ │ ├── hcfm-train-00000-of-00007.parquet … 00006-of-00007 │ │ ├── hcfm-validation-00000-of-00001.parquet │ │ └── hcfm-test-00000-of-00003.parquet … 00002-of-00003 │ ├── mcfm/ 44 files · 52.5 GB · 18,255,265 rows │ │ ├── mcfm-train-00000-of-00029.parquet … 00028-of-00029 │ │ ├── mcfm-validation-00000-of-00001.parquet │ │ └── mcfm-test-00000-of-00014.parquet … 00013-of-00014 │ └── lcfm/ 454 files · 467.1 GB · 181,777,591 rows │ ├── lcfm-train-00000-of-00293.parquet … 00292-of-00293 │ ├── lcfm-validation-00000-of-00019.parquet … 00018-of-00019 │ └── lcfm-test-00000-of-00142.parquet … 00141-of-00142 │ ├── by_project/ COMPLETE TIER · NOT FILTERED · NOT SPLIT │ ├── hcfm/ 15,166 files · 11.4 GB · 3,684,448 rows │ │ ├── PXD000561/ 82 accessions in every tier │ │ │ ├── Adult_Adrenalgland_Gel_Elite_49_f01.parquet │ │ │ └── … one file per instrument run │ │ └── PXD000865/ … │ ├── mcfm/ 15,244 files · 51.7 GB · 18,422,236 rows │ │ └── PXD000561/ … same runs, fewer rows each │ └── lcfm/ 15,286 files · ~450 GB · 184,607,213 rows │ └── PXD000561/ … │ ├── peptide_registry.parquet 5,613,657 peptides and their split │ └── manifests/ └── empty_runs.csv 162 runs with no PSMs at their threshold ``` **Pick one tier, from one folder.** Because the tiers are nested, a second tier re-downloads the same spectra at a stricter threshold — and `splits/*` or `by_project/*` fetches all three, roughly 531 GB and 510 GB. ### `splits/` — start here `train`, `validation` and `test` parquet, exactly as the model consumed them. The partitions are **peptide-disjoint** 80/10/10: a peptide sequence appears in only one of the three, so evaluation does not reward memorisation. Rows are shuffled and have passed the quality filters below. Rows per split, measured: | | train | validation | test | |---|---:|---:|---:| | `hcfm_splits` | 2,459,391 (67.0%) | 178,997 (4.9%) | 1,031,725 (28.1%) | | `mcfm_splits` | 11,703,040 (64.1%) | 790,417 (4.3%) | 5,761,808 (31.6%) | | `lcfm_splits` | 117,230,014 (64.5%) | 7,716,481 (4.3%) | 56,831,096 (31.3%) | The 80/10/10 ratio is over **peptides**, not spectra: the registry assigns each peptide to one split and every spectrum of that peptide follows it. Peptides differ in how many spectra they have, and test peptides carry disproportionately many — 11.5% of peptides but ~31% of rows — so the test partition is about 2.7x its nominal share and validation about half of its own. Use this to reproduce or extend the published results. ### `by_project/` — the input the splits came from Each tier before filtering and splitting. Published because the filtering is lossy: rows dropped by the quality gates are not recoverable from `splits/`. Use this to apply different quality criteria, or to re-derive the partitions with `peptide_registry.parquet` (see below). ### Quality filters applied to `splits/` but not to `by_project/` | field | kept | |---|---| | retention time | <= 10800 s | | lower isolation offset | <= 300 Da | | precursor charge | 0-7 **inclusive** (0 is kept) | | precursor *m/z* | <= 2000 | | modification annotation | only if resolvable to a UNIMOD identifier | A null in any numeric field **passes** its condition, so a reimplementation that discards nulls instead would produce a different dataset. These tiers contain no nulls in any filtered field, so the two behave identically here — but the distinction matters if you apply the same criteria to new data. Measured effect on LCFM (184,607,213 -> 181,777,591 rows, 98.5% kept): retention time removes ~2.54 M, the unresolved-modification condition ~385 k, lower offset 8,147, and the charge and *m/z* bounds remove **nothing** — no row in the corpus exceeds either. The unresolved annotations are modifications the pipeline could not resolve to a UNIMOD identifier (`[IN:]`; 175 tokens over ids 3000–3174, predominantly N-glycans on asparagine but **not** exclusively — `K[IN:3174]` is on lysine). Glycopeptides whose glycan *does* have a UNIMOD identifier are retained: 100 distinct UNIMOD modifications occur across 30.7% of the corpus. ### `peptide_registry.parquet` The split assignment of every peptide in the corpus. It is what makes the peptide-disjoint partitions reproducible, and it lets new data be partitioned consistently with these splits rather than at random — apply it to `by_project/`, or to your own spectra, to keep a held-out set genuinely held out. ### Joining `by_project/` to the registry Join on **`registry_key`** — not on `sequence`, and not on `unmodified_peptide`: ```python from datasets import load_dataset # hcfm is the smallest tier (11.4 GB); swap in mcfm or lcfm when you need more. runs = load_dataset("InstaDeepAI/InstaNovo", "hcfm_by_project", split="full") registry = load_dataset("InstaDeepAI/InstaNovo", "peptide_registry", split="full") labelled = runs.to_polars().join( registry.to_polars(), left_on="registry_key", right_on="peptide", how="left" ) ``` Both calls download and cache from the Hub, so nothing needs fetching by hand. `to_polars()` materialises the tier in memory, and `mz_array` and `intensity_array` dominate that, so call `runs.select_columns(["registry_key", "experiment_name", "scan"])` first if you only need the split assignment. To work through one accession at a time rather than a whole tier, pass `streaming=True` and filter on `experiment_name`, or fetch single files directly: ```python from huggingface_hub import hf_hub_download path = hf_hub_download( "InstaDeepAI/InstaNovo", "by_project/hcfm/PXD000561/Adult_Adrenalgland_Gel_Elite_49_f01.parquet", repo_type="dataset", ) ``` `registry_key` is `unmodified_peptide` with every `I` rewritten to `L`, which is the convention the registry itself uses: its `peptide` column contains no `I` at all. Isoleucine and leucine are isobaric, so a peptide pair differing only by I/L is indistinguishable to MS/MS, and collapsing them into one key is what stops one variant training while the other tests. Mind the direction. Collapsing `L` to `I` instead yields keys that match nothing, and a failed join returns *no rows* rather than visibly wrong ones — which reads naturally as "these peptides are new". The collapse applies to the key only. `sequence` and `unmodified_peptide` keep their original I and L residues, and the model was trained on those unnormalised sequences. There is no `normalised_peptide` column. Earlier internal copies carried one that was never populated; it is removed here so nothing invites a join that would silently match nothing. ## Confidence tiers The tiers are nested subsets at increasing PSM-confidence thresholds: LCFM is the full labelled corpus, MCFM and HCFM are progressively stricter. MCFM and HCFM inherit LCFM's split assignments, so a peptide has the same split in every tier. The unlabelled **ACFM** tier is not published here. It is approximately 5.7 TB and comprises every MS/MS scan from the same raw files, so it is reconstructible from the PRIDE accessions with the conversion pipeline deposited at Figshare, and is otherwise available from the authors on request. ## Loading a specific tier The configs above name every tier and flavour, so nothing is inferred from the directory layout: ```python from datasets import load_dataset load_dataset("InstaDeepAI/InstaNovo", "hcfm_splits") # train/validation/test load_dataset("InstaDeepAI/InstaNovo", "lcfm_splits", split="test") load_dataset("InstaDeepAI/InstaNovo", "hcfm_by_project") # one "full" split ``` To fetch files without loading them, select by path — tier and flavour are directory prefixes: ```bash hf download InstaDeepAI/InstaNovo --repo-type dataset --include "hcfm/*" ``` ## File naming Split files follow the Hub's sharding convention, `{split}-{index:05d}-of-{total:05d}.parquet`: ``` splits/hcfm/hcfm-train-00000-of-00007.parquet splits/hcfm/hcfm-validation-00000-of-00001.parquet ``` The `-of-{total}` suffix makes an incomplete download self-evident, and the zero-padding sorts correctly in a plain lexicographic listing. `by_project/` deliberately does **not** use that convention. It has no train, validation or test split, and any of those keywords in a filename would make the Hub advertise a split that does not exist. Its files are named `data-{index:05d}-of-{total:05d}.parquet` within each project accession directory, so per-project selection stays possible: ``` by_project/lcfm/PXD012345/.parquet ``` ## Versioning `v0.1` is the first release, and the one the paper's results were produced from. Pin it rather than tracking `main`, so a later addition to the corpus cannot change what you fetch: ```python from datasets import load_dataset ds = load_dataset("InstaDeepAI/InstaNovo", "hcfm_splits", revision="v0.1") ``` `main` moves as tiers are extended or corrected; a tag does not. ## Licence The spectra derive from public submissions to [PRIDE](https://www.ebi.ac.uk/pride/), so use of this dataset is governed by the [EMBL-EBI terms of use](https://www.ebi.ac.uk/about/terms-of-use/). The code that produced the corpus is available on [Figshare](https://doi.org/10.6084/m9.figshare.33368752) under a CC BY 4.0 licence, and the model checkpoints are CC BY-NC-SA 4.0. ## Citation See the InstaNovo-FM manuscript. Code, including the pipeline that produced these files, is at .