Open bgyori opened 8 years ago
Along similar lines, when a molecular entity is mentioned in relation to a cell type, it is often extracted as the actual subject/object of an event. For instance, consider the synthetic example: BRAF inhibition in NF1 mutant cells causes apoptosis.
from which BRAF negatively activates NF1 is extracted. Here the statement is about NF1 mutant cells
and not NF1
itself.
An actual example from a paper is Combination of FGFR and AKT inhibition in an FGFR2 mutated endometrial cancer xenograft model [...]
, which yields two negative activation events: FGFR negatively activates FGFR2 and AKT negatively activates FGFR2.
@bgyori: so, on the first example, should we extract a negative activation between AR and docetaxel?
No, the simple solution would be to not extract anything at all. Figuring out what docetaxel resistance
is would be beyond the capabilities of the system so it's better to leave it out. For the NF1 mutant cells
example, in principle, NF1 mutant cells
could be extracted as context on the actual process that the sentence talks about but again it might be hard to do and it would be better not to extract anything for NF1.
Actually, for the first sentence, a legitimate extraction would be "AR increases p21 expression."
Yes, John is right.
Possible approaches:
nn
link. A disadvantage of this is that we'd have to make this edit everywhere -- in all rules that have an optional nn
before the theme, which is most of them.I like (1), but there is some small risk it could hamper the coref module's hunt for antecedent. Generally, though, I don't think we care about entity mentions unless they happen to be participating in events.
(2) could be handled as a variable defining a lookaround constraint on certain args of events, but this will complicate the templates a bit...
Some more examples that result in incorrect negative activation events:
@myedibleenso, @danebell: it looks like the actionable item here is to add 1 (or 2) grammar rules to capture: "NF1 mutant cells" and "Brca1 deficient cells" and "BRAF V600E -positive melanoma cells" as CELL_TYPE. Thus disabling events on the proteins included in these entities. @myedibleenso, @danebell: please decide between the two of you who should add these rules.
Foud another one:
Treatment of AI resistant cell lines with LBH589 repressed NF-kappaB1 mRNA and protein expression
Here AI
shouldn't be extracted as negatively activating NF-kappaB1
.
I found another one that is not about cell types but has a similar pattern in which a modifier is missed after the entity name: a new mechanism of oxygen independent activation of HIF-1 has been identified
from which oxygen activates HIF-1
was extracted and independent
was ignored.
@bgyori: syntactically, this looks similar but it has to be handled differently in our system. This fits under code that detects the polarity of activations/regulations. Except, in this case, if the activation is independent of the controller, we should discard the interactions. I will start a separate issue for this.
@myedibleenso: please add rules for "* cell"?
In the evaluation results we bump into incorrect explanations due to a frequently occurring error related to this issue. Consider these examples:
19244107 This is supported by a previous study showing that haplodeficiency of Akt1 dramatically inhibits prostate tumor development in Pten +/- mice (XREF_BIBR). 23786676 Akt1 and 2 deficiency is sufficient to markedly reduce the incidence of tumors in Pten (+/-) mice [XREF_BIBR] and Myc also cooperates with Akt1 in promoting prostate tumorigenesis [XREF_BIBR]. 20231902 Moreover, knockdown of Akt1 induces MST2 activation and enhances doxorubicin activated MST2 and apoptosis in PTEN mutated MDA-MB-468 cells (XREF_FIG). 19369943 PI3K and Akt signalling is also essential for oncogenic ErbB-2-induced transformation (XREF_BIBR), and Akt1 deficiency sufficiently suppresses tumour development in PTEN +/- mice (XREF_BIBR).
In each case AKT1 inhibits PTEN was extracted even though the sentences talks about a more complex process (e.g. tumor development) and only mentions PTEN as a contextual variable.
Thanks @bgyori! Unfortunately, these are multiple issues showing for the same interaction... We are discussing these.
The consensus at this point seems to be to try @danebell's possible approaches 1 (modifying the entity rules).
For the sentence
The consequences of increased AR function might then increase docetaxel resistance via increasing p21 expression.
, REACH extracts an event stating that AR positively activates docetaxel. The key issue here is that the object of the statement isdocetaxel resistance
and notdocetaxel
. To simplify the example, I tried the sentenceAR increases docetaxel resistance.
which yields the same extracted event.