Comment by Barrera
4 years ago
If this were really a practical concern, machine learning would be designing drugs that fly through the clinic today. They aren't and so this paper, though click-grabbing, is probably of no practical consequence.
One reason is lack of data. Chemical data sets are extremely difficult to collect and as such tend to be siloed on creation. Synthesis of the target compounds and testing using uniform, validated protocols are non-trivial activities. They can only be undertaken by deep pockets. Those deep pockets are interested in return on investment. So, into the silo it goes. This might not always be the case, though.
For now, the paper does raise the question of the goals and ethics around machine learning research. But unintended and/or malevolent consequences of new discoveries have been a problem for a long time. Just ask Shelley.
A successful drug candidate must be useful in the treatment of human medical problems and not have harmful side effects that outweigh its benefits. A weaponized poison may have any number of harmful effects without diminishing its utility. A compound with really indiscriminate biochemical effects, like fluoroethyl fluoroacetate, makes a potent poison without any specific tuning for humans. It's much easier to discover compounds that genuinely harm people than those that genuinely help them.