I worked for a company that tried to do exactly what this article is proposing, I was responsible for parsing the data from publications into exactly this sort of table.

The primary reason is that it is very hard to come up with a schema that even 5% of papers would adhere to. The vast majority of this knowledge is phrased as natural language.

There are databases that track compounds and the publications related to them, but those papers again are natural language and cannot be readily converted to tabular data. Our first basic approach involved POS tagging and then trying to associate proper nouns with numeric values. Again the issue became how do you interpret a sentence like "may lead to sudden death" as a symptom? Something like "may lead to symptom X in Y% of respondents" is a nightmare to consistently parse without heavy ML running over huge datasets of just text.

In the end we wound up having to shut down concluding that until papers are released with not only arbitrary XML tables/results we were not equipped to handle the task. And even worse, what if our models didn't interpret things correctly and a consumer got {symptom:"sudden death", chance:0%} and insisted on that compound for their indication only to later realize the paper stated "in lab setting 0% of animals didn't experience sudden death after being administered X after diabetes diagnosis". Paying a hundred students to work around the clock couldn't get the volume processed accurately for months, let alone getting a second army of validators to confirm each entry.