Releases: jwohlwend/boltz
Releases · jwohlwend/boltz
Boltz v2.2.1: Minor bug fixes
What's Changed
- fix: redundant PoseBustersPotential bounds by @simonlevine in #478
- Fix help message in --use_potentials flag by @Alberto024 in #487
- Minor simplification of Diffusion Loss by @shenoynikhil in #486
- Fix: update filter_inputs_affinity to only filter records if override=False by @tobiasploetz in #483
- Allow case-insensitive input suffixes and correct docstring mismatch by @Jnelen in #481
- Minor guidance_update renaming cleanup by @JoshuaMeyers in #489
- bugfix: NoneType not iterable when inference_pocket_constraints is None by @JoshuaMeyers in #493
- Bug fix in bias repeat when multiplicity > 1 by @shenoynikhil in #524
- Pad cyclic_period to ensure tensors passed to forward have a static size by @Adam27X in #546
- Fix bf16 dropout by @tlitfin-unsw in #554
- Minor tweaks to the documentation by @timholy in #561
- Minor bugfix to potentials by @JoshuaMeyers in #548
New Contributors
- @Alberto024 made their first contribution in #487
- @Jnelen made their first contribution in #481
- @JoshuaMeyers made their first contribution in #489
- @Adam27X made their first contribution in #546
- @tlitfin-unsw made their first contribution in #554
- @timholy made their first contribution in #561
Full Changelog: v2.2.0...v2.2.1
Contributors
timholy, Adam27X, and 7 other contributors
Assets 2
Boltz v2.2.0: Contact conditioning fixes, new potentials
Highlight
- Fixes for contact and pocket conditioning which was not fully functional for Boltz2
- New potentials
- Kernels are now optional to avoid installation issues on non-linux platforms
- New security features for use of the MSA server (thanks @papagala!)
And many others!
What's Changed
- Generalize dropping leaving atoms for covalent modifications and noncanonical modifications. by @benf549 in #349
- Respect existing CUEQ envvars for tuning by @pechersky in #348
- Typos by @CurtisColwell in #351
- Fixes for training and Torch compatibility by @simonlevine in #364
- Fix Typo: Correct "neighboorhood" to "neighborhood" in Comments by @zeevick10 in #369
- Fix typo: "explicitly" instead of "explicitely" in prediction.md documentation by @maximevtush in #371
- Minor Documentation and Code Comments Improvements by @leopardracer in #386
- Fix for use_kernels for later torch versions for PairformerNoSeqModule by @simonlevine in #380
- Update mols.tar download logic by @brianloyal in #379
- fixed typo in prediction.md documentation by @fisand08 in #375
- Fix flag by @JohnKirwan in #440
- Docs: Update README and Prediction Documentation by @kilavvy in #410
- Optimize images by @paulpham157 in #376
- Fix --output_format pdb with --model boltz1 by @benf549 in #365
- Fix template matched by @Dawincy134 in #445
- Add support for dumping embeddings during prediction by @ribesstefano in #457
- Update prediction.md by @jansteinkuehler in #423
- Contacts and Template Potentials by @Noahb930 in #462
- Fix over-eager skip logic in filter_inputs_structure by @mohiuddin-khan-shiam in #426
- Fixes issue when training Boltz-1x with steering by @darianyang in #255
- Improve
TrainingDataset __getitem__performance by @gcroci2 in #441 - fix: pass full BoltzSteeringParams to affinity prediction by @tobiasploetz in #468
- Adding MSA server security by @papagala in #466
- Fix typo in schema for contact constraint by @glukhove in #472
- fix issue with non-linux installation by @gcorso in #479
New Contributors
- @pechersky made their first contribution in #348
- @CurtisColwell made their first contribution in #351
- @zeevick10 made their first contribution in #369
- @maximevtush made their first contribution in #371
- @leopardracer made their first contribution in #386
- @brianloyal made their first contribution in #379
- @fisand08 made their first contribution in #375
- @JohnKirwan made their first contribution in #440
- @kilavvy made their first contribution in #410
- @paulpham157 made their first contribution in #376
- @Dawincy134 made their first contribution in #445
- @ribesstefano made their first contribution in #457
- @jansteinkuehler made their first contribution in #423
- @mohiuddin-khan-shiam made their first contribution in #426
- @gcroci2 made their first contribution in #441
- @tobiasploetz made their first contribution in #468
- @papagala made their first contribution in #466
- @glukhove made their first contribution in #472
- @gcorso made their first contribution in #479
Full Changelog: v2.1.1...v2.2.0
Contributors
JohnKirwan, pechersky, and 21 other contributors
Assets 2
Boltz v2.1.1: New CUDA kernels from NVIDIA
New Features
This release introduces two novel kernels from NVIDIA for the triangular attention and triangular multiplication layers, achieve 2x speedup and large memory savings!
Changes
The kernels are on by default for CUDA users with Ampere cards or newer, otherwise they are turned off.
Boltz v2.1.0
fix
Boltz v2.0.3: Patch CCD affinity
Merge branch 'main' of github.com:jwohlwend/boltz
Boltz v2.0.2
Another patch, small issue with multi ligand copies and affinity.
Boltz v2.0.1: Patch
Fixes an important issue with checkpoint download. Please delete your cache to trigger re-download of the checkpoints if you encounter an issue.
Boltz v2.0.0: Boltz-2 release
Boltz-2 is out!
Boltz-2 is our latest model and a step forward in molecular modeling with the introduction of binding affinity prediction. Boltz-2 also comes with exciting new features such as contact, method and template conditioning for extra user control. We’re excited to see the cool things you’ll build with it.
Highlights
- Binding affinity prediction for protein-small molecules, closing the gap with free energy perturbation methods at 1000x speedup!
- More user control via structural templates (including mutlimeric), contact and method conditioning.
- Improved performance on antibodies and other modalities
- Improved modeling of ensembles when using MD conditioning
Boltz v1.0.0: Boltz-1x release
We're excited to release Boltz-1x!
Boltz-1x introduces inference-time steering for much higher physical quality, CUDA kernels for faster, more memory-efficient inference and training, and more!
Highlights
- Inference time steering to ensure quality physical poses. Amazing work by our own @Noahb930 !
- Trifast kernel integration for the triangular attention module. Thank you @latkins !
- A new
cyclicflag for proteins in our input YAML, which tweaks the relative positional encoding for significant improvements to cyclic peptides. Thank you @ioannisa92
All changes
- Check EMA parameter compatibility after initializing EMA by @xluo233 in #154
- Update training.md by @Oracle16 in #156
- add name to wandb logger by @holgerroth in #158
- mmcif: scale qa_metric.PLDDT to [0,100] by @phue in #170
- preprocessing bug: chain clash filter gets skipped by @pedersor in #177
- Fix get_resolution in data preprocessing by @kkleidal in #144
- Latkins/faster construct paired msa by @latkins in #175
- raise error when atom name encoding is violating the expected dimension by @wckdouglas in #143
- Small fixes to process cif files and use them in training by @lbugnon in #141
- fix random_subset logit for MSA subsampling by @simonlevine in #186
- bugfix in determining existing manifest records by @timodonnell in #222
- fix: Minimum Python version update + results folder .gitignore by @dsmith111 in #224
- Fix Boltz writing waters in mmcif format, issue #227. by @tomgoddard in #228
- Fix the expected mask shape for AtomAttentionPairBias by @wiwnopgm in #226
- feat: allow cache directory to be set via BOLTZ_CACHE environment variable by @ihayhurst in #215
- fix: correctly handle chirality information by @boraml in #219
- raise exception instead of dict in OOM handler by @TanushGoel in #237
- Feat/cyclic offset by @ioannisa92 in #213
- Fixes ligand pLDDTs in PDB/mmCIF b-factor column bugs. by @benf549 in #240
- add a missing comma in defining structure_module by @LtECoD in #152
- Add trifast. by @latkins in #147
New Contributors
- @LtECoD made their first contribution in #152
- @xluo233 made their first contribution in #154
- @Oracle16 made their first contribution in #156
- @holgerroth made their first contribution in #158
- @phue made their first contribution in #170
- @pedersor made their first contribution in #177
- @kkleidal made their first contribution in #144
- @latkins made their first contribution in #175
- @wckdouglas made their first contribution in #143
- @simonlevine made their first contribution in #186
- @timodonnell made their first contribution in #222
- @dsmith111 made their first contribution in #224
- @tomgoddard made their first contribution in #228
- @wiwnopgm made their first contribution in #226
- @ihayhurst made their first contribution in #215
- @boraml made their first contribution in #219
- @TanushGoel made their first contribution in #237
- @ioannisa92 made their first contribution in #213
- @benf549 made their first contribution in #240
Full Changelog: v0.4.1...v1.0.0
Contributors
timodonnell, latkins, and 19 other contributors
Assets 2
Boltz v0.4.1: Small patch release
Small patch due to an error in chucking mode