From the abstract of the 2019 Nature article outlining the success of AlphaFold:
Here we show that we can train a neural network to make accurate predictions of the distances between pairs of residues, which convey more information about the structure than contact predictions. Using this information, we construct a potential of mean force4 that can accurately describe the shape of a protein. We find that the resulting potential can be optimized by a simple gradient descent algorithm to generate structures without complex sampling procedures
One implication of what is stated above is that AlphaFold provides an algorithm that maps sequence to experimental folds with high accuracy but does not attempt to provide a physically accurate multi-purpose interaction potential. It is a statistically generated potential rather than based on fine tuning terms with a physical interpretation (e.g. angular, torsional, steric, Coulombic). This results among other things in potential predictive limitations imposed by the training sequence domain. Like any model it can guarantee success only if relevant training data (or potential terms) are included in creating the model. The sequence-fold relations relevant to quaternary structure might not overlap well with those relevant to lower order folding. I would argue that a subset of the folding interactions (in particular contacts between distant residues not involved in secondary structure) are more relevant to quaternary structure. It is also not clear that the potential would be useful to perform MD simulations to probe non-equilibrium structures or sample local minima.
Despite that, if those potential roadblocks are not significant I do not see why your approach should not work, particularly if you pick an appropriate linker (which is easily tested).
This answer risks might seem opinion-based but is an interpretation based on what can be quickly inferred from information in the linked publication.