Release 2026.2.0

CHANGELOG.md

@@ -1,6 +1,31 @@
 Changelog
 =========
 
+[2026.2.0](https://github.com/openbiosim/somd2/compare/2026.1.0...2026.2.0) - Sep 2026
+--------------------------------------------------------------------------------------
+
+* Add support for generating Boresch restraints for absolute binding free energy calculations [#166](https://github.com/OpenBioSim/somd2/pull/166).
+* Give alchemical ions their own plain morph lambda schedule so they interpolate correctly under non-standard lambda schedules [#169](https://github.com/OpenBioSim/somd2/pull/169).
+* Persist alchemical ion identity across restarts so the same molecule is reused regardless of GCMC state [#172](https://github.com/OpenBioSim/somd2/pull/172).
+* Use perisistent `ThreadPoolExector` objects within the main replica exchange dynamics block [#175](https://github.com/OpenBioSim/somd2/pull/175).
+* Allow `oversubscription_factor` to change on restart [#177](https://github.com/OpenBioSim/somd2/pull/177).
+* Restrict energy component decomposition to force groups that are used for integration [#180](https://github.com/OpenBioSim/somd2/pull/180).
+* Parallelise replica mixing [#181](https://github.com/OpenBioSim/somd2/pull/181).
+* Fixed the replica exchange GPU memory check querying the wrong device when `CUDA_VISIBLE_DEVICES` does not start at zero, since OpenMM numbers devices relative to the visible set whereas `pynvml` enumerates all of them [#183](https://github.com/OpenBioSim/somd2/issues/183).
+* Store GCMC sampling statistics per lambda value, converting those from earlier checkpoints on restart [#184](https://github.com/OpenBioSim/somd2/pull/184).
+* Link restart systems to the reference end state rather than the perturbed one, since that is the coordinate set that dynamics maintains. Perturbable molecules were otherwise resumed from the coordinates they were built with [#189](https://github.com/OpenBioSim/somd2/pull/189).
+* Add `max_contexts` to cap the number of OpenMM contexts used for replica exchange, re-using each across lambda values so that GPU memory no longer limits the number of replicas [#191](https://github.com/OpenBioSim/somd2/pull/191).
+* Skip minimisation on restart [#191](https://github.com/OpenBioSim/somd2/pull/191).
+* Pre-equilibrate the water with GCMC moves before minimising in the regular `Runner`, making it consistent with the `RepexRunner`, which already did so to stop the geometry relaxing into a dry pocket [#191](https://github.com/OpenBioSim/somd2/pull/191).
+* Add a `precision` option for GPU platforms, defaulting to `single` [#191](https://github.com/OpenBioSim/somd2/pull/191).
+* Add support for generating Morse restraints for ring-breaking perturbations [#194](https://github.com/OpenBioSim/somd2/pull/194).
+* Remove the unused `kappa` lever equations from the ring-breaking/making lambda schedules [#195](https://github.com/OpenBioSim/somd2/pull/195).
+* Accept stream file paths for the `restraints` and `lambda_schedule` configuration options, so they can be set from the command line [#198](https://github.com/OpenBioSim/somd2/pull/198).
+* Account for off-site charges (virtual sites) when computing the charge difference between the end states. They are held as a molecule property rather than on the atoms, so a charge-preserving perturbation could appear to change charge and be given spurious alchemical ions [#200](https://github.com/OpenBioSim/somd2/pull/200).
+* Handle `num_lambda=1`, which previously raised a `ZeroDivisionError` when generating the lambda values. The `RepexRunner` now rejects a single lambda window, since there is nothing to exchange with and the regular `Runner` is faster [#203](https://github.com/OpenBioSim/somd2/pull/203).
+* Detect the available GPUs once in the base runner and re-use the list, rather than the `RepexRunner` querying `CUDA_VISIBLE_DEVICES` regardless of the chosen platform. Replica exchange is now also permitted on the HIP platform [#206](https://github.com/OpenBioSim/somd2/pull/206).
+* Query the free memory of AMD GPUs with `CL_DEVICE_GLOBAL_FREE_MEMORY_AMD` rather than `CL_DEVICE_BOARD_NAME_AMD`, which returns the device name [#206](https://github.com/OpenBioSim/somd2/pull/206).
+
 [2026.1.0](https://github.com/openbiosim/somd2/compare/2025.1.0...2026.1.0) - Jun 2026
 --------------------------------------------------------------------------------------
 

pixi.toml

@@ -6,9 +6,9 @@ platforms = ["linux-64", "osx-arm64"]
 [dependencies]
 python = ">=3.10"
 # main
-biosimspace = ">=2026.1.0,<2026.2.0"
+biosimspace = ">=2026.2.0,<2026.3.0"
 # devel
-#biosimspace = "==2026.2.0.dev"
+#biosimspace = "==2026.3.0.dev"
 filelock = "*"
 ghostly = "*"
 loch = "*"

recipes/somd2/recipe.yaml

@@ -20,9 +20,9 @@ requirements:
     - versioningit
   run:
     # main
-    - biosimspace >=2026.1.0,<2026.2.0
+    - biosimspace >=2026.2.0,<2026.3.0
     # devel
-    #- biosimspace ==2026.2.0.dev
+    #- biosimspace ==2026.3.0.dev
     - filelock
     - ghostly
     - loch
@@ -37,6 +37,10 @@ tests:
       - python:
           imports:
             - somd2
+          # AmberTools can be pulled in transitively, and installs tools into
+          # site-packages whose metadata still pins numpy <2, so 'pip check'
+          # fails even though nothing here uses them.
+          pip_check: false
       - script:
           - PYTHONPATH=. pytest -vvv --color=yes --import-mode=importlib ./tests
         files:

src/somd2/__init__.py

@@ -34,10 +34,20 @@
 # Store the somd2 version.
 from ._version import __version__
 
-# Store the sire version.
+# Store the sire version. Unlike somd2/BioSimSpace/ghostly/loch (which use
+# versioningit and only append a "+g<revisionid>" local version segment
+# for non-release (".dev") builds, omitting it entirely for a clean tagged
+# release), sire exposes its version and revision id as separate attributes,
+# with __revisionid__ always set regardless of release status. Build a
+# composite string using the same "+g<revisionid>" convention as the other
+# packages, only appending it for non-release (".dev") builds, so that all
+# five version strings are formatted consistently.
 from sire import __version__ as _sire_version
 from sire import __revisionid__ as _sire_revisionid
 
+if ".dev" in _sire_version:
+    _sire_version = f"{_sire_version}+g{_sire_revisionid}"
+
 # Store the BioSimSpace version.
 from BioSimSpace import __version__ as _biosimspace_version
 
@@ -61,7 +71,7 @@ def get_versions():
     """
     return {
         "somd2": __version__,
-        "sire": f"{_sire_version}+{_sire_revisionid}",
+        "sire": _sire_version,
         "biosimspace": _biosimspace_version,
         "ghostly": _ghostly_version,
         "loch": _loch_version,

src/somd2/_utils/_schedules.py

@@ -27,7 +27,44 @@
 ]
 
 
-def annihilate(fix_epsilon=True):
+def _set_boresch_lever_equations(s, stage_dihedral, stage_distance_angle):
+    """
+    Set the equations for a "split" restraint_lever Boresch restraint (see
+    sire.restraints.boresch's restraint_lever parameter), reproducing the
+    RXRX protocol's staged restraint turn-on (Table S1 of the RXRX paper's
+    SI): within each of the two named stages, the corresponding restraint
+    group ramps from ~0 to 1 following a geometric progression, while the
+    other group is held fixed. 'stage_dihedral' is the stage over which
+    the dihedral restraint group ramps on (with the distance/angle group
+    held at 0); 'stage_distance_angle' is the stage over which the
+    distance/angle group ramps on (with the dihedral group held at 1,
+    already fully on).
+
+    Note: this aligns the restraint turn-on with SOMD2's own decharge/
+    annihilate(or decouple) stage boundaries, rather than reproducing the
+    RXRX paper's exact global 50/50 window split (which falls partway
+    through the annihilate/decouple stage, since the paper's own decharge
+    stage is only 21 of 64 total bound-leg windows) - the relative sizes of
+    SOMD2's stages are controlled by lambda_values weighting, not fixed.
+    """
+    from sire.legacy.CAS import Exp as _Exp
+    import math as _math
+
+    # Geometric progression from ~0.01 (fully off) to 1.0 (fully on), matching
+    # the ratio observed in the RXRX paper's published lambda schedule.
+    ramp_on = _Exp((1 - s.lam()) * _math.log(0.01))
+
+    s.set_equation(stage=stage_dihedral, lever="restraint_dihedral", equation=ramp_on)
+    s.set_equation(stage=stage_dihedral, lever="restraint_distance_angle", equation=0)
+    s.set_equation(stage=stage_distance_angle, lever="restraint_dihedral", equation=1)
+    s.set_equation(
+        stage=stage_distance_angle,
+        lever="restraint_distance_angle",
+        equation=ramp_on,
+    )
+
+
+def annihilate(fix_epsilon=True, restraint_lever="split"):
     """
     Build the ABFE lambda schedule using decharge → annihilate.
 
@@ -44,12 +81,28 @@ def annihilate(fix_epsilon=True):
         If False, epsilon is scaled normally from initial to final and the LRC
         follows naturally.
 
+    restraint_lever : str, optional
+        How the Boresch restraint is controlled by this schedule, matching
+        sire.restraints.boresch's restraint_lever parameter. Either "split"
+        (default), where the dihedral restraint terms are turned on during
+        decharge and the distance/angle terms are turned on during
+        annihilate, reproducing the RXRX protocol's staged restraint
+        turn-on, or "combined", where the whole restraint is turned on
+        together during the decharge stage. The Boresch restraint object
+        passed to the simulation must have a matching restraint_lever value.
+
     Returns
     -------
 
     schedule : sire.legacy.CAS.LambdaSchedule
         The lambda schedule.
     """
+    if restraint_lever not in ("combined", "split"):
+        raise ValueError(
+            "'restraint_lever' must be either 'combined' or 'split', "
+            f"got {restraint_lever!r}"
+        )
+
     from sire.cas import LambdaSchedule as _LambdaSchedule
 
     # Start with the standard decouple schedule and modify the stages and
@@ -65,14 +118,22 @@ def annihilate(fix_epsilon=True):
         lever="charge",
         equation=s.lam() * s.final() + s.initial() * (1 - s.lam()),
     )
-    s.set_equation(stage="decharge", lever="restraint", equation=s.lam() * s.final())
 
     s.add_stage(
         "annihilate",
         equation=(-s.lam() + 1) * s.initial() + s.lam() * s.final(),
     )
     s.set_equation(stage="annihilate", lever="charge", equation=s.final())
-    s.set_equation(stage="annihilate", lever="restraint", equation=s.final())
+
+    if restraint_lever == "split":
+        _set_boresch_lever_equations(
+            s, stage_dihedral="decharge", stage_distance_angle="annihilate"
+        )
+    else:
+        s.set_equation(
+            stage="decharge", lever="restraint", equation=s.lam() * s.final()
+        )
+        s.set_equation(stage="annihilate", lever="restraint", equation=s.final())
 
     if fix_epsilon:
         s.set_equation(stage="annihilate", lever="epsilon", equation=s.initial())
@@ -86,7 +147,7 @@ def annihilate(fix_epsilon=True):
     return s
 
 
-def decouple(fix_epsilon=True):
+def decouple(fix_epsilon=True, restraint_lever="split"):
     """
     Build the ABFE lambda schedule using decharge → decouple.
 
@@ -101,20 +162,35 @@ def decouple(fix_epsilon=True):
         ghost-LRC force is then explicitly scaled to zero over the stage.
         If False, epsilon is scaled normally and the LRC follows naturally.
 
+    restraint_lever : str, optional
+        How the Boresch restraint is controlled by this schedule, matching
+        sire.restraints.boresch's restraint_lever parameter. Either "split"
+        (default), where the dihedral restraint terms are turned on during
+        decharge and the distance/angle terms are turned on during decouple,
+        reproducing the RXRX protocol's staged restraint turn-on, or
+        "combined", where the whole restraint is turned on together during
+        the decharge stage. The Boresch restraint object passed to the
+        simulation must have a matching restraint_lever value.
+
     Returns
     -------
 
     schedule : sire.legacy.CAS.LambdaSchedule
         The lambda schedule.
     """
+    if restraint_lever not in ("combined", "split"):
+        raise ValueError(
+            "'restraint_lever' must be either 'combined' or 'split', "
+            f"got {restraint_lever!r}"
+        )
+
     from sire.cas import LambdaSchedule as _LambdaSchedule
 
     # Start with the standard decouple schedule and modify the stages and
     # equations as needed. This will be folded into Sire in future, but
     # we will use this approach for prototyping.
     s = _LambdaSchedule.standard_decouple()
 
-    s.set_equation(stage="decouple", lever="restraint", equation=s.final())
     s.set_equation(stage="decouple", lever="kappa", force="ghost/ghost", equation=0)
     s.set_equation(stage="decouple", lever="kappa", force="ghost-14", equation=0)
     s.set_equation(stage="decouple", lever="charge", equation=s.final())
@@ -142,7 +218,16 @@ def decouple(fix_epsilon=True):
     s.set_equation(
         stage="decharge", lever="kappa", force="ghost-14", equation=-s.lam() + 1
     )
-    s.set_equation(stage="decharge", lever="restraint", equation=s.initial() * s.lam())
+
+    if restraint_lever == "split":
+        _set_boresch_lever_equations(
+            s, stage_dihedral="decharge", stage_distance_angle="decouple"
+        )
+    else:
+        s.set_equation(stage="decouple", lever="restraint", equation=s.final())
+        s.set_equation(
+            stage="decharge", lever="restraint", equation=s.initial() * s.lam()
+        )
 
     return s