Skip to content

backend_notes.md

Latest commit

 

History

History
87 lines (77 loc) · 5.07 KB

backend_notes.md

File metadata and controls

87 lines (77 loc) · 5.07 KB
Raw

PA backend notes

Integration and reproducibility notes for using PA as a de-novo backend (e.g. the denovo_benchmarks harness, or the borgonovo reference-free assembler). See the top-level README.md for the container and the high-level I/O contract; this file collects the lower-level gotchas. Every claim is cited to file:line.

Environment / build gotchas

  • Submodule URL is SSH. .gitmodules points depthcharge at git@github.com:Alfred-N/depthcharge.git, so a recursive clone fails without an SSH key. The container works around this by rewriting the URL to HTTPS before git submodule update (misc/container.def:19-21); do the same (or clone the fork over HTTPS) outside the container.
  • pkg_resources / setuptools. environment.yml:11 pins pytorch-lightning==2.1.1, which still imports the deprecated pkg_resources. Recent setuptools (≥ 81) no longer ship it, and environment.yml does not pin setuptools — if you hit ModuleNotFoundError: pkg_resources, add setuptools<81 to the env.
  • mzML support is untested in the bundled env. depthcharge parses mzML/mzXML through pyteomics.mzml/pyteomics.mzxml (depthcharge depthcharge/data/parsers.py:14-15), which typically also needs psims for PSI-MS CV terms. environment.yml only exercises the MGF path, so expect to install extra pyteomics dependencies before using --suffix .mzML.

Running src/main.py directly

  • Run from the repo root. Config paths are relative — configs/master_bm.yaml references configs/downstream/casanovo_bm.yaml (configs/master_bm.yaml:38) — and the container cds into the repo before invoking it (misc/denovo_benchmarks_pairwise/make_predictions.sh:15-20).
  • --predict_only takes a value. It is type=int, default=0 (src/parse_args.py:221-226) cast to bool (parse_args.py:306), i.e. --predict_only=1, not a bare flag. (In the container it comes from predict_only: 1 in configs/master_bm.yaml:30.) Prediction then runs through trainer.predict (src/main.py:207-209).
  • Output-dir creation aborts if outs/ exists. create_output_dirs calls os.makedirs(..., exist_ok=False) for both output_dir and log_dir (src/parse_args.py:391-392). With fixed_output_dir: 1 (configs/master_bm.yaml:31) the paths are not uniquified, so a leftover outs/ from a previous run will crash the next one — delete it first.

Driving the model without the Lightning Trainer

  • setup() needs a Trainer. BenchmarkDataModule.setup (and LanceDataModule.setup) read self.trainer.global_rank / self.trainer.world_size to build the ShardedBatchSampler (src/data/lance_data_module.py:119-125). To build a dataloader without a real pl.Trainer, attach a dummy, e.g. dm.trainer = SimpleNamespace(global_rank=0, world_size=1) before calling dm.setup("predict").

Calling the decoder directly (per-step distributions)

Neither the container nor make_predictions.sh emits per-step probability distributions over the token vocabulary (the mzTab carries only scalar per-residue confidences, and outputs.csv simulates even those — output_mapper.py:191). A borgonovo-style backend that needs the distributions must call the model directly:

  • The decoder behind decoder_model: casanovo_decoder is PeptideDecoder.forward(tokens, precursors, memory, memory_key_padding_mask) (src/models/casanovo/decoder_interface.py:109,166).
  • It returns (scores, tokens), with scores of shape (B, L, V) (raw logits; softmax for probabilities) — return self.final(preds), tokens (decoder_interface.py:163).
  • The first step uses decoder(None, precursors, memory, mem_masks); None tokens means only the precursor embedding is fed in (decoder_interface.py:137-141; called at src/wrappers/beam_search.py:84, then iteratively at beam_search.py:121-126).
  • precursors is [mass, charge, mz], where mass = mz·charge (protonated cluster mass — see the precursor-mass note in the README) (src/wrappers/casanovo_trainer_wrapper.py:150-154).
  • decoder.reverse controls decoding direction; the released checkpoint runs with reverse: True (configs/downstream/casanovo_bm.yaml), set onto the decoder at src/wrappers/casanovo_trainer_wrapper.py:99, which means C→N order (stop token $ appears first; beam_search.py:232).
  • Greedy decoding is just beam search with n_beams: 1 (configs/downstream/casanovo_bm.yaml; wrapper default at src/wrappers/casanovo_trainer_wrapper.py:76).
  • Token labels are C+57.021, M+15.995, … (src/utils.py:33,49), not Casanovo's C[Carbamidomethyl] / ProForma C[UNIMOD:4].