MLE DG estimates are wrong for bidirectional RBFE calculations on ligand network.

[afroditi-zaki-cresset]

Describe the bug
I have run bidirectional RBFE calculations (lig1->lig2 and lig2->lig1 for all the transformations in the ligand network) on a dataset with a user-defined ligand network and when I compute and plot the predicted vs experimental dG values, I see that the predicted dG values are not calculated properly and I get very high values of MUE and RMSE (correct R2 though). This does not happen when I only use unidirectional calculations, in which case the pred. vs exp dG plot is very reasonable. I was wondering if this is known behaviour because the MLE algorithm is not designed to work on bidirectional ddG values?

To Reproduce
I attach the cinnaber_input.csv file that I used.

cinnabar_input.csv

I also attach the ligand network along with the predicted ddG and dG values.

Software versions

• My OS is Ubuntu 24.04
• For FEP calculations and analysis I used OpenFE v1.3

Output
The pred. vs exp. dG plot shows very high MUE and RMSE values and the ddG values do not add up to the estimated dGs (see ligand network image).

Expected behavior
If I only use unidirectional ddG values in my cinnaber-input.csv the pred. vs exp. dG plot and the statistics look much more reasonable and in line with what we expected.

Additional context
I also tried to exclude from the calculation the edges where the forward and reverse ddG values did not match and I still did not get a correct dG prediction.

[IAlibay]

Thanks for reporting this @afroditi-zaki-cresset - we'll try to triage this and get back to you as soon as we can.

From a very brief look at things, I suspect this is partly responsible:

cinnabar/cinnabar/stats.py

Lines 201 to 208 in 37cbc2b

	for (a, b) in graph.edges:
	i = node_name_to_index[a]
	j = node_name_to_index[b]
	if i == j:
	# The MLE solver will fail if we include self-edges, so we need to omit these
	continue
	F_matrix[i, j] = -df_ij[i, j] ** (-2)
	F_matrix[j, i] = -df_ij[i, j] ** (-2)

We will need to think it through more though.

[jthorton]

#178 Is a partial fix to this and now adds guardrails to advise that this doesn't work by default, for now its best if you combine the predictions into a single entry before using MLE and this should work!

[jthorton]

Confirmed this was fixed in #211, running the CLI on this input with the current main branch generates the following plot which agrees well with the unidirectional results above.