In the September 2019 issue of Nature Biotechnology, a collective of scientists mostly from Insilico Medicine, led by Alex Zhavoronkov, CEO and founder of the company, published a landmark article.
Deep learning enables rapid identification of potent DDR1 kinase inhibitors  is not just an article championing the benefits of AI in the search for molecules for use in pharmaceutics.
It is the story of taking a new molecule to market in 46 days, including its synthesis and in-vitro testing. The upstream design phase took just 21 days. A process like this would normally take 2 to 3 years by traditional means.
What’s more, the Machine Learning technology used is no small feat: it is called GAN, Generative Adversarial Networks, a method that was thought to be confined to image processing and generation.
Why is this success story attracting our attention in particular? Among the innumerable applications of AI, in the field of industrial control, voice recognition, object detection in image processing or customer relations, what makes this story special?
The fundamental tripartition of AI domains
As discussed in previous articles, an important dividing line must be considered for any AI application: are we dealing with a semantically closed domain, a semantically opened domain, or a domain in between?
A semantically closed domain is one in which the rule system that defines it is:
- Formalised using language that leaves no room for ambiguity
All the great strategy games meet this definition: Chess, Go, Bridge, Checkers, Awélé…
Diagnostic activities using precise and formal technical language also fall into this camp: Diagnosis of breakdowns in technical industrial fields, medical diagnosis, or legal analysis in commercial law.
In semantically closed domains, AI is relentlessly efficient. In the great strategy games, not only does it far surpass man, beating the world champions of the discipline, but it also outperforms algorithms based on raw combinatorial computation .
Prestigious professions, such as doctors and business lawyers, were very upset when AI demonstrated that their expertise was nothing more than the synthetic compilation of a large number of rules   .
However, performance is mitigated where medical diagnosis is concerned, as psychological elements or elements related to the patient’s life situation can have a major effect on the practitioner’s judgement. AI outperforms the human physician only in cases where the relevant clinical observations are completely formulaic. Sometimes these formal observations are sufficient, sometimes much more implicit elements are needed for diagnosis.
The impressive performance of AI in semantically closed domains has fed the fears of a superhuman AI, storming the fortress of consciousness, leading to the trans-humanist fantasies of techno-prophets. The discourse skilfully combines the positive dream of immense progress with the fear of machines outstripping us and threatening to take power, a cocktail intended more for media success than for scientific analysis.
In many other areas of knowledge, those that can be described as “semantically open”, AI not only does far less well than humans, but even my cat could do better in terms of adaptability. In a nutshell, AI will always do much better than humans in any given game, but real life consists of playing a multitude (an infinity?) of very varied games, without us being given any explicit indication that we are moving from one game to another.
Until now, human consciousness and even animal consciousness have had the privilege of knowing how to operate these shifts, that is to say, of being able to work outside their conditioning while still being entirely within it. The human consciousness is capable of deciding for itself to position itself “out of the box”, which no AI can do.
Finally, a fairly wide field of knowledge mixes semantically open and semantically closed natures. In this scenario, AI performs well and is making significant inroads, but remains dependent on human supervision, particularly to make the final decision resulting from treatment. The relationship between man and machine is in this case one of cooperation or “decision support”. These are often the most interesting areas, where AI can demonstrate its full value compared to classical algorithms, moving into areas that cannot be handled by the latter.
The distribution of the various fields where AI can be applied are summarised in the following table:
The combination of biological inhibitors
Molecular biology teaches us that many diseases and cures are a matter of one molecule taking the shape of another molecule. The archetypal model of this mechanism is that of the catalytic action of an enzyme on another molecule, the ligand or substrate, which the enzyme will help transform:
The enzyme-substrate complex can be formed by a “lock and key” type action (Fischer model), in which the molecules retain their shape rigidly. In this case, only stereometric matching of the molecules will allow the complex to be formed: the substrate simply fits into the active site of the enzyme like a key in a lock.
More elaborate Enzyme-Substrate complexes can be obtained, either by a prior deformation of the active site of the enzyme which adapts to its substrate before binding (Koshland induced-fit), or by an ex-post deformation of the enzyme (conformational selection), which will be preferentially chosen by the substrate because its active site is more compatible than other competing sites, even if the binding is not a perfect fit.
In the case of bacterial diseases, toxins are emitted that resemble enzymes in every way. Toxins take the place of healthy enzymes, replacing the natural function with their active site. They can prevent the catalytic reaction by simple inhibition or they can degrade the substrate by a harmful reaction instead of catalysis.
In the case of a viral disease, this mechanism does not seem to be at play because the viral aggression consists of the invasive molecule replicating itself by parasitizing a healthy cell. But in some cases, one can act indirectly against viral aggression from an enzyme/substrate complex, when the enzyme in question is found to facilitate the progression of the disease. This is the case when combatting influenza, where proteases – calpains – are found to facilitate the inflammatory cascade of the disease. Inhibiting calpains does not directly attack the influenza virus, but “dries” one of the fuels required for its spread .
Molecular forms are thus used as activators or inhibitors, both in the natural mechanisms of catalytic transformation, and in attacks by enemy molecules that attempt to replace their active sites, and finally in drugs that precisely block this invasion by hostile active sites.
The living world is like a giant Tetris game, with a struggle being fought to occupy free active sites, and with hostile organisms deflecting the original reaction when they are the first to occupy a site.
The struggle between our natural defences and their aggressors also resembles military decoys and insect camouflage techniques: the aggressor seeks to imitate the appearance and function of the helpful ally, to occupy his position, in a constant struggle for free spaces. The deadliest poisons are those that resemble in every way our natural catalysts for respiratory function or muscle transmission.
Activation or inhibition are the two opposing levers used by both friendly and unfriendly substances. A medicine can consist equally of fighting a toxin directly and of removing certain natural elements from its propagation, by voluntarily inhibiting them before it uses them.
The interplay of inhibition/activation is a formal semi-closed set, involving competition with tricks and deception.
Jacques Monod’s observation in “Le hasard et la nécessité” [Chance and Necessity] has not dated: in the living world, matter is equipped to serve as a set of logical keys. The material specificities of a molecule are only used for the stereometric interlocking combination, which itself is only the transmission vector of information and start or stop signals for production.
This giant Tetris is indeed a formal game, where you can identify the various combinations of pieces, with a finite number of them and the consequences of each move. Is it a semantically closed game? It would be tempting to answer in the affirmative, because after all it boils down to the wooden pieces of a large construction set and the positive or negative processing functions after embedding, which are well known.
However, the possibilities of adaptive deformation of the enzyme before or after engagement with the substrate (Koshland induced-fit or conformational selection) define a continuous palette of possible nuances, with infinite small variations around the fit. These margins for manoeuvre make the game semantically semi-closed / open.
The same applies to genetic engineering, whose formal rules seem to be perfectly defined. They are, but immersed in a continuous cloud of small possible modifications, introducing context dependence. We have already mentioned this fact on the matter of GMOs, with Jean-Paul Oury reminding us that they require specific experimental fields, genetic engineering that can work for a narrow range of given plants without being transposable to one of its seemingly close cousins. 
Let’s add to this that the little game of trick / deception between active sites and their substrates explains the success of AI techniques like GANs: they are exactly based on the principle of learning greater precision in distinguishing a genuine signal from the one produced by a forger. In the living world, our lethal enemies seek to resemble our best friends, with increasing exactness and subtlety, which is precisely what GANs are about.
Molecular pharmacology, the next field revolutionized from top to bottom by AI
For all these reasons, molecular pharmacology is the next area where AI will make invaluable breakthroughs. We hardly need to emphasise the huge stakes of these advances: research for molecules for use in medicine with detection reduced from two years to two months would enable massive victories to be won over a number of rare diseases, even those considered incurable. As explained above, the principle is not restricted to bacterial diseases, as some viral attacks can be indirectly combated by inhibition mechanisms.
As always, these breakthroughs will be manna for the sensationalist press: we will be treated to more bad editorials on humans thinking themselves god-like, capable of infallible self-care and in real time, why not, to extend the idea a little.
Armed with truly rational knowledge about the currently impassable barrier of changing contexts, we are able to appreciate the genuine progress made by these techniques without descending into mere congratulatory speeches.
Many human lives will be saved or greatly improved in the coming years thanks to the use of AI in molecular pharmacology, a far better reason to rejoice than any dreams of mankind made divine.
This post is also available in: FR (FR)