take a small optimization step for the torch rsa calculation for mode…

…l selection

alphafold3_pytorch/alphafold3.py

@@ -5707,6 +5707,36 @@ def _compute_unresolved_rasa(
 
         return unresolved_rasa.mean()
 
+    @typecheck
+    def calc_atom_access_surface_score_from_structure(
+        self,
+        structure: Structure,
+        **kwargs
+    ) -> Float['m']:
+
+       # use the structure as source of truth, matching what xluo did
+
+        structure_atom_pos = []
+        structure_atom_type_for_radii = []
+        side_atom_index = len(self.atom_type_index)
+
+        for atom in structure.get_atoms():
+
+            one_atom_pos = list(atom.get_vector())
+            one_atom_type = self.atom_type_index.get(atom.name, side_atom_index)
+
+            structure_atom_pos.append(one_atom_pos)
+            structure_atom_type_for_radii.append(one_atom_type)
+
+        structure_atom_pos: Float[' m 3'] = tensor(structure_atom_pos)
+        structure_atom_type_for_radii: Int[' m'] = tensor(structure_atom_type_for_radii)
+
+        return self.calc_atom_access_surface_score(
+            atom_pos = structure_atom_pos,
+            atom_type = structure_atom_type_for_radii,
+            **kwargs
+        )
+
     @typecheck
     def calc_atom_access_surface_score(
         self,
@@ -5749,7 +5779,7 @@ def calc_atom_access_surface_score(
             lat.sin()
         ), dim = -1)
 
-        # first get atom relative positions + distance
+        # get atom relative positions + distance
         # for determining whether to include pairs of atom in calculation for the `free` adjective
 
         atom_rel_pos = einx.subtract('j c, i c -> i j c', atom_pos, atom_pos)
@@ -5762,13 +5792,23 @@ def calc_atom_access_surface_score(
             (atom_rel_dist_sq > atom_distance_min_thres)
         )
 
+        # max included in calculation per row
+
+        max_included = include_in_free_calc.long().sum(dim = -1).amax()
+
+        include_indices = include_in_free_calc.long().topk(max_included, dim = -1).indices
+
+        include_in_free_calc = einx.get_at('i [m], i j -> i j', include_in_free_calc, include_indices)
+        atom_rel_pos = einx.get_at('i [m] c, i j -> i j c', atom_rel_pos, include_indices)
+        target_atom_radii_sq = einx.get_at('[m], i j -> i j', atom_radii_sq, include_indices)
+
         # overall logic
 
         surface_dots = einx.multiply('m, sd c -> m sd c', atom_radii, unit_surface_dots)
 
         dist_from_surface_dots_sq = einx.subtract('i j c, i sd c -> i sd j c', atom_rel_pos, surface_dots).pow(2).sum(dim = -1)
 
-        target_atom_close_to_surface_dots = einx.less('j, i sd j -> i sd j', atom_radii_sq, dist_from_surface_dots_sq)
+        target_atom_close_to_surface_dots = einx.less('i j, i sd j -> i sd j', target_atom_radii_sq, dist_from_surface_dots_sq)
 
         target_atom_close_or_not_included = einx.logical_or('i sd j, i j -> i sd j', target_atom_close_to_surface_dots, ~include_in_free_calc)
 
@@ -5831,28 +5871,10 @@ def _inhouse_compute_unresolved_rasa(
             chain_atom_mask,
         )
 
-        # use the structure as source of truth, matching what xluo did
-
-        structure_atom_pos = []
-        structure_atom_type_for_radii = []
-        side_atom_index = len(self.atom_type_index)
-
-        for atom in structure.get_atoms():
-
-            one_atom_pos = list(atom.get_vector())
-            one_atom_type = self.atom_type_index.get(atom.name, side_atom_index)
-
-            structure_atom_pos.append(one_atom_pos)
-            structure_atom_type_for_radii.append(one_atom_type)
-
-        structure_atom_pos: Float[' m 3'] = tensor(structure_atom_pos)
-        structure_atom_type_for_radii: Int[' m'] = tensor(structure_atom_type_for_radii)
-
         # per atom rsa calculation
 
-        per_atom_access_surface_score = self.calc_atom_access_surface_score(
-            structure_atom_pos,
-            structure_atom_type_for_radii,
+        per_atom_access_surface_score = self.calc_atom_access_surface_score_from_structure(
+            structure,
             **rsa_calc_kwargs
         )
 

pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "alphafold3-pytorch"
-version = "0.5.48"
+version = "0.5.49"
 description = "Alphafold 3 - Pytorch"
 authors = [
     { name = "Phil Wang", email = "[email protected]" },