The meaning of "interaction" is quite broad: 2 protein sequences 
occur in the same complex or they occur in the same reaction.

The 1st line begins with '#' and gives the number of unique 
interactions, i.e. unique pairs if UniProt accession numbers.

The rest of the file is tab delimited with columns having following 
meaning/content:

1. UniProt accession for interactor 1.

2. Ensembl gene id(s delimited by '|') for interactor 1. Note that one 
UniProt entry may be mapped to multiple Ensembl genes.

3. Entrez Gene id(s delimited by '|') for interactor 1. Note that one 
UniProt entry may be mapped to multiple Entrez Gene ids.

4. UniProt accession for interactor 2.

5. Ensembl gene id(s delimited by '|') for interactor 2. Note that one 
UniProt entry may be mapped to multiple Ensembl genes.

6. Entrez Gene id(s delimited by '|') for interactor 2. Note that one 
UniProt entry may be mapped to multiple Entrez Gene ids.

7. Type of interaction, either 'direct_complex', 'indirect_complex',
'reaction' or 'neighbouring_reaction'. Meanings of those below.

8. Interaction context. This is the class and internal identifier of 
Reactome instance via which this interaction happens. For
'neighbouring_reaction' type interaction this contains a pair of
class:identifier string separated by '<->'.

9. Comma-delimited list of PubMed identifiers supporting this 
interaction. For a reaction these are the PMIDs from the 
Reaction.literatureReference slot (and not from the "second tier" 
Reaction.summation.literatureReference).
For a complex these are from reations which produce the complex.
For 'neighbouring_reaction' type interaction this field is empty.


Interaction types:
------------------
In Reactome, complexes can be arbitrarily nested, i.e., contain
subcomplexes. Here we have complex C1 consisting of complexes C2 and
C3. C2 consist of molecules A and B. C3 consists of molecules C and D:

C1 --- C2 --- A
     |      |
     |      |- B
     |
     |- C3 --- C
            |
            |- D

Now really to the interaction types:
- direct_complex - interactors are directly in the same complex, i.e. 
w/o further nested complexes. From the example interactions A <->B and 
C<-D> are of this type.
- indirect_complex - interactors in different subcomplexes of a complex.
From the example above interactions A<->C, A<->D, B<->C and B<->D are 
of this type.
- reaction - interactors participate in the same reaction. Only those 
reactions are reported for which the intreactors are not complexed (with 
the exception being association dissociation reactions which are reported).
- neighbouring_reaction - interactor participate in 2 "consecutive"
reactions, i.e. when one reaction produces a PhysicalEntity which is either an 
input or a catalyst for another reaction. However, to avoid the "trivial" (i.e.
over ATP, ADP etc) interactions, the computation is done using the
'precedingEvent' attribute used to order reactions in a pathway.

The 'direct_complex' type doesn't really mean true pairwise interaction. 
For example, Reactome's representation of ribosomal S60  subunit consists
"directly" of 40+ proteins and has no sub-complexes. Obviously this can 
hardly mean that each one of those component proteins directly interacts 
with every other protein. Hence, the 'direct_complex' vs. 
'indirect_complex' distinction is quite meaningless.



Other notes:
------------
- If the same interaction (i.e same UniProte accession pair) happens in 
multiple contexts, each context is reported on a separate row and hence 
the row number in the file is bigger than the number of unique 
interactions reported on the 1st line.
- The order of interactors on one line has no meaning (they are just 
ordered according to the Reactome internal identifier...). Every 
interaction with context is reported once, i.e. A<->B is not reported as 
B<->A.