Add function to check if there is a route with a given set of starting molecules #34
Conversation
There was a problem hiding this comment.
A very general question: what would happen if we just set the is_purchasable attribute to False for all nodes apart from those in starting_mols, and then computed the has_solution attribute using run_message_passing? Would that not give us what we want? Presumably this would give an affirmative answer even if a route exists that uses a subset of starting_mols, so a related question would be whether we only care about routes that use exactly that set of leaves instead of its subset?
(Even more generally, I wonder how useful this graph-based formulation of this metric is: if a graph contains a combinatorial number of routes, then stating that there exists a route with the right set of leaves seems not so practically useful. But I guess it doesn't hurt to have a function to compute this.)
| # Initialize to empty sets, except for nodes with purchasable mols | ||
| node_to_mols: dict[ANDOR_NODE, set[Molecule]] = {n: set() for n in graph.nodes()} | ||
| for n in graph.nodes(): | ||
| if isinstance(n, OrNode): | ||
| node_to_mols[n].add(n.mol) |
There was a problem hiding this comment.
The comment says "purchasable", but is this reflected in the code?
There was a problem hiding this comment.
Good catch: I used to check whether a molecule is purchasable but removed it to make the method more general (i.e. the nodes marked purchasable in the graph don't need to be the starting molecules). I will fix this by changing the comment.
You are right the I thought a bit about how to do this but could not think of a better way than what I coded: my impression is that a dynamic programming solution is not possible here because it there is no clear way to divide the problem into independent subproblems. |
|
I converted this to a draft PR since some experiments I am running are very slow. I might change the implementation. |
…es are solvable at all).
|
Just added some extra pruning code to increase the efficiency. I updated the main PR description to reflect this. |
|
I think I will close this PR because:
We can re-visit and re-open this PR later if we want to add this function. |
First fix: error message for converting an AND/OR graph to a synthesis route would say there was more than one node if in fact there were zero nodes. Second fix: reaction string in `conftest.py`. This was also fixed in #34, but since I am not merging that I decided to fix it here.
In the paper FusionRetro they proposed checking whether the starting molecules for a route found during multi-step planning match the starting materials used in an actual synthesis route. This is essentially a less strict version of checking whether an exact ground-truth route is found. This PR implements code to check for the existence of a route with a specific set of starting materials.
In FusionRetro this metric is called "set-wise exact match", however this name is very vague so I instead called it
is_route_with_starting_mols. I only wrote a version for AND/OR graphs because these typically contain a prohibitively large number of routes to explicitly enumerate them. The method is essentially a brute-force search with pruning of search branches. The inner workings of the methods are described reasonably-well in the docstrings (but I want to improve the description before merging)I also fixed a random typo in a
conftest.pydocstring that I found while writing tests.TODO: