CHANGELOG.md

To Do

• Expand test suite to include function unittests (maximize coverage)

v0

v0.3

v0.3.1: Replace BioPython with biotite

• Replace requirement for local foldseek installation in extract-3di
• Replace all fasta parsing with biotite.sequence.io.fasta

v0.3.0: D-SCRIPT with BMPI

• Introduce predict_block to support multi-GPU prediction of protein interactions with semi-on-the-fly embedding loading, resolving #6 and #11
• Add sparse embedding loading to predict_block, in which only the required subset of each block of embeddings is loaded at each time when predicing only some pairs
• Add predict_bipartite for predicting interactions between all proteins from two disjoint sets, e.g., proteins from two species. This implementation is a bit more efficient than using predict_block with cross-set pairs specified.
• Improve method for loading embeddings
• Modernized D-SCRIPT repository
• Significantly updated unittest coverage
• Migrate to loguru under the hood
• Update pyproject, GitHub actions, and other continuous integration/installation
• Linting and formatting with Ruff

v0.2

v0.2.2

• Resolve #35 to use require_dataset -- can now add multiple .fasta files to the same h5 file
• Update pretrained API and docs to include Topsy-Turvy
• Add retry decorator to get_pretrained if download fails
• Add ability to set a random seed for training
• Update evaluate code to also store metrics in a file

v0.2.1: 2022-06-28 -- Bug fixes

• Add biopython to setup.py

v0.2.0: 2022-06-24 -- Integration of Topsy Turvy

• Integrate Topsy-Turvy to allow for top-down supervision
• Use utils.log function across all commands
• Speed up loading embeddings into memory using parallel processing
• Update fasta parse and write to use BioPython SeqIO (better error checking)
• More comprehensive test suite for main commands

v0.1

v0.1.9: 2022-03-07 -- Minor bug fixes

• Updated model loading on new version to handle re-named parameters
• Updated cpu-only loading during prediction with map_location

v0.1.8: 2022-02-08 -- Finalized training bug fix

• Resolve #24 by fixing training
• Can now run dscript train --train data/pairs/human_train.tsv --test data/pairs/human_test.tsv --embedding /afs/csail/u/s/samsl/Work/databases/STRING/homo.sapiens/human_nonRed.h5 --output [output] --save-prefix [prefix] --device 0 to replicate paper results
• Updated code formatting with black and pre-commit

v0.1.7-dev9: 2022-02-07 -- Update training code to be clean again

• Following previous update, addresses #24 by fixing model training while maintaining preferred API and command line usage

v0.1.7-dev1: 2021-12-16 -- Major overhaul in training

• Fixed significant bug in how training was run by reverting to older code
• Should address issue #24: unable to replicate paper results
• To do: code cleaning to bring up to formatting standards while maintaining performance

v0.1.6: 2021-09-06 -- Bug Fix - Augmentation and proper defaults

• Augmentation fix in v0.1.5 was bugged still and would throw an error, now resets index
• Change --use-w and --augment to --no-w and --no-augment with store false

v0.1.5: 2021-06-23 -- Bug Fix - Augment and Documentation

• Updated package level imports
• Updated documentation
• Fixed issue #13: improper augmentation of data
• Fixed issue #12: overwrites cmap data sets if they already exist

v0.1.4: 2021-03-05 -- Bug Fix - Typo in ContactModule.forward()

• Fixed issue #7: bug which would crash contact module if called directly

v0.1.3: 2021-02-17 -- Bug Fix - Pairs too large for GPU

• Fixed issues #3, #4
• Basic logging system implemented to report skipped pairs
• Fixed wrong variable name in loading from sequence file
• Updated documentation

v0.1.2: 2020-11-30 -- Bug Fix - Eval Mode

• Model should be put into eval() mode before prediction or evaluation, and when new models are downloaded - this makes the output deterministic by disabling dropout layers

v0.1.1: 2020-11-18 -- First Beta Release