- Add best-fitting model selection workflow for CryoFold
- CryFold tutorial script adapted for command-line use
- Update python dependencies in setup.py
- Decoupled the number of workers from the number of replicas.
- You can now use any combination that divides evenly
- e.g. 64 replicas on 16 workers
launch_remd_multiplexhas been removed- use
launch_remdwith one worker instead
- use
- You can now use any combination that divides evenly
- RDCs now work with peak mapping
- Ability to freeze atoms with
freeze_atoms - Ability to remove the potential from the force field with
remove_potential- This can be useful when trying to infer assignments using peak mapping
- Added helper methods to simplify setting up simulations
setup_replica_exchangesetup_data_store
- Previous versions of MELD could give erroneous values for forces and energies
- Due to a data race in the CUDA kernels, different threads could calculate different values for the cutoff energy within each collection. This would result in groups occasionally being activated when they shouldn't be and vice versa.
- We now use nvidia's cub library to handle the sort operation required for groups and collections.
- The new version of the CUDA kernel agrees well with the CPU verion.
- Compiling cub library requires the use of nvcc, which must be forced by
setting
OPENMM_CUDA_COMPILER.
- Previous versions could sometimes incorrectly negotiate device ids between nodes which would lead to failure at runtime.
- Amber-specific system buiding tools are in their own modules
- RDCs no longer add particles to the system
- Energies are saved to different energy groups. The current mapping is:
- Base forcefield energy: 0
- MELD force: 1
- RDC force: 2
- Confinement restraints: 3
- COM restraints: 4
- Cartesian restraints: 5
- Log Prior for parameter sampling: 6
- Most user facing modules are now imported into the root
meldnamespace. - Most user-facing functions now require units, which will help to eliminate errors.
- You can now specify the platform to use.
- Options are: "CUDA", "CPU", "Reference"
- E.g.
launch_remd --platform CPU
- Many changes to indexing.
- All atom indexing is zero-based.
- By default, residue and chain indexing is zero-based, but individual lookups can specify one-based indexing.
system.index.atom(resid, atom_name, chainid=None, one_based=False)- Can find the index of an atom.
- Absolute index if
chainid=None. - Relative index within a chain if
chainidis specified. - Both
residandchainidare zero-based by default, but can be one based ifone_based=True. - Returns an
AtomIndex, which is a zero-based absolute atom index.
system.index.residue(resid, chainid=None, one_based=False)- Can find the index of a residue.
- Absolute index if
chainid=None. - Relative index within a chain if
chainidis specified. - Both
residandchainidare zero-based by default, but can be one based ifone_based=True. - Returns a
ResidueIndex, which is a zero-based absolute residue index.
- All functions defining restraints now take
AtomIndexrather sets of residues and atom names. - "follower" changed to "worker"
- No longer use network logging.
- Previous versions used network logging to aggregate all log messages into a single remd.log file.
- This required the leader task to spawn an additional process to run the log server. Getting this right was difficult and was the source of many developer headaches.
- This feature has now been removed and all worker nodes write to their own log
file in the
/Logsdirectory.
- Implented MELD plugin for CPU and Reference plugins
- This should allow for improved minimization. When large forces are encountered
during minimization, OpenMM switches to the CPU platform. Previously, this would
cause calculations using forces in
meldpluginto crash. With this version,meldpluginforces are now supported on the CPU platform, so the minimizer may succeed. - This implmentation is incomplete, as it does not (yet) allow for the whole meld calculation to run on the CPU, but that should be possible in the future.
- Only the basic distance and torsion restraints are currently implemented. Use of other restraint types will cause an error on the CPU platform.
- This should allow for improved minimization. When large forces are encountered
during minimization, OpenMM switches to the CPU platform. Previously, this would
cause calculations using forces in
- Changed negotiation of device ids, which previously caused problems with startup on some HPC systems.
- Improved minimization with RDCs
- Fixed race condition in logging.
- Changed default minimum value for restraint scalers
- Previously, restraints would be scaled to 0.0 by default when alpha=1.0
- This could cause problems with some restraints, like RDCs
- The new default scales the restraints down to 1e-3, which is effectively off, but prevents problems with restraints like RDCs
- The previous behaviour can be obtained by setting
stregnth_at_alpha_max=0.0
- Combined license into a single file
- Fixed error that caused simulations to hang upon completion or upon error
- Minor changes to plugin build process.
- Continuous integration now uses github actions.
quadratic_cutoption ofget_secondary_structure_restraintsis now interpreted correctly. Previously, this value was interpreted as tens of nanometers, instead of tenths of nanometers.
- Implemented flat-bottom / quadratic / linear RDC restraints
- Much faster handling of RDCs following the approach of Habeck and co-workiers. DOI 10.1007/s10858-007-9215-1
- You can now specify quadratic_cut for RDC restraints. The energy will increase quadratically when |d_obs - d_calc| is less than quadratic cut and linearly when it is larger.
- Fixed a bug where RDC energies were incorrect in mixed precision.
- Fixed updating of RDC forces. Previously, the RDC forces were not correctly updated with alpha and timestep.
- Major improvement in simulation startup time with large numbers of restraints.
- Fixed inconsistencies in RDC calculations
- Added support for implicit ion concentration and solute/solvent dielectric constants.
- Fixed major bug in handling of groups where the number of
restraints to be satisfied was less than the total size
of the group. On GTX 1080 Ti, this resulted in frequent
crashes with
CUDA Error (700). It is unknown what effect this would have on other platforms. Simulations with groups that were 100% satisfied are unaffected.
- Support for building explicit solvent systems
- Support for periodic boundary conditions
- Started working on documentation
- Fixed incorrect tests for GMM restraints
- Exceptions now flush buffers which fixes issue with some MPI implmentations dropping diagnostic information
- Long amino acid sequences no longer crash
tleap
- Drop osx support
- There really isn't a compelling use case and most recent Macs don't have NVIDIA GPUs
- Fixed bug in positioners with distance restraints that caused runs to fail immediately.
- Added support for positioners in distance restraints
- Changed handling of GMM scaling factors
- Enhanced numerical stability of GMM simulations
- Change the way the scale factor for GMMs is passed to the GPU. The previous version gave incorrect results.
- Added Gaussian Mixture Model (GMM) restraints
- Fixed bug in timeout handling
- Fixed an issue that could cause ImportErrors
meld.comm.MPICommunicatorobjects now have a timeout. The default is 600 seconds, or 10 minutes. This should prevent hangs that were occasionally observed.
- Virtual spin labels now use a restricted angle potential that avoids 180 and 0 degree angles that can cause the dihedral potential to blow up.
- Completed implementation of virtual spin label sites
- Fixed a bug where OpenMM would re-order identical chains involved in MELD forces. This could only occur if identical molecules were involved in a MELD force. MELD now correctly specifies that any molecules involved in a MELD force are unique and should not be reordered.
- (CRITICAL) The fix introduced in 0.3.7 was incorrect and sometimes gave incorrect results due to a data race. This version uses a work efficient, in place, parallel algorithm. It passes several hundred iterations of the plugin test-suite, whereas the previous version would fail every few trials.
- Added improved diagnostics to aid in future debugging of collections.
- (CRITICAL) Fixed a bug in collection handling introduced in e71Daa0 that caused undefined behaviour. The previous version gave correct results on test systems with GTX980 cards, but incorrect results on Tesla m2070. The status on other architectures is unknown.
- (CRITICAL) Fixed bug in mixed precision introduced in a727279. The energies reported in previous versions were undefined.
- (CRITICAL) Fixed bug in secondary structure handling that was introduced in 0.3.2.
- Added ability to specify arbitrary leap commands in header
- TranslationMover for translational Monte Carlo moves
- Support for reading in secondary structure from multiple chains
- (CRITICAL) Fixed critical bug where MELD forces were not being updated on GPU.
- This is the first version where we are tracking changes.
- The plugin has been merged into the main MELD repository.
- We now build a single anaconda package (meld-test) on each commit.