Living in the age dominated by science and technology, we are witness to unprecedented knowledge-based innovations in several areas of human activity. One such area is the application of artificial intelligence in drug discovery.
Recollecting the history, it is no denying that magic and religion played the most prominent role in prehistoric medicine. Nevertheless, with the progress of human civilization, disease-countering remedies gradually started to be based on evidence, albeit initially primitive.
As in other fields of human pursuits, dramatic changes came about in medicine, first, with the invention of writing and, later, that of printing. Now, information could be recorded and preserved.
Science and technology bring modern medicine
Developments in X-ray crystallography, nuclear magnetic resonance and cryo-electron microscopy helped decipher the molecular interactions in biological (including human) systems by resolving the structures of biomolecules at the atomic resolution. The technological revolution in medicine was further consolidated and boosted by incredible advances in computers – hardware and software.
It is well-known that biological systems, especially in the field of human diseases, are a complex and rich source of information. Systematic measurements and recording using various technologies have led to the accumulation of an awesome amount of biological information.
In the field of medicine, computers are being used not only for the storage of health records and patient databases but, more importantly, as an indispensable tool for molecular diagnosis and structure-based drug design.
Are computers indispensable?
To answer this question, let us look at the stages of drug discovery. The process begins with identification of potential targets for a specific disease. Biological data, such as genome data, transcriptome data and proteome data, obtained from normal and diseased individuals are accessed and analyzed by bioinformatic approaches. Patterns and other valuable information are uncovered from large data sets leading to identification and prioritization of a biomolecule (for example, a protein) or biomolecular complex that can be related to the disease.
Once the biomolecule (or a biomolecular complex) has been identified, effort is made to narrow down the target to a crucial site, either catalytic or regulatory, in the molecule. This requires a three-dimensional (3D) structure of the molecule obtained experimentally or, where experimental structure is not available, by what is known as “homology” modeling. The target is then validated in the laboratory using animal models.
Is all that “humanly” possible?
After all, humans possess the “intelligence” which, according to Collins dictionary is “the ability to think, reason, and understand, instead of doing things automatically or by instinct”. So, in principle, to extract meaningful pattern or other valuable information from biological data, they should be able to “read” the data, think and apply reasoning, and understand which biomolecule would be a good target for the disease under investigation.
Nevertheless, humans are limited by their physical ability – precision and speed in the present case – which cannot be decoupled from the “intelligent” (mental) activity. Unavoidably, they look for a “machine” (computer) which takes in data as input and dispense results as output in a lot more efficient, speedy and precise manner.
Humans make computers to work for them
It is true that the traditional computer does not possess the necessary “intelligence” for the input-to-output conversion. This “intelligence” is “donated” to the computer as an algorithm (an explicit, precise, unambiguous, mechanically-executable sequence of elementary instruction) by human programmers.
The necessity for the computer becomes even more pressing at the stage of structure determination. Understandably, it is not a question of a handful of atoms and their relative positions. Here, one seeks to transform X-rays-generated images from thousands of atoms into a plausible structure.
Identification of the drug target, be it around a catalytic or regulatory site, is followed by the next logical step to “explore the chemical space” by screening libraries of chemical compounds for possible binding. The search is carried out using powerful computers by a high-throughput “docking” method based on geometric complementarity, hydrogen bonds and hydrophobic contacts (incorporated in the algorithm) between the target and ligand.
Limitations recognized
Despite these advances of computer-aided drug design, the success of the approach is not absolute. “Hit” molecules, defined as chemical compounds having the desired activity demonstrated in the screening test, often end up in failures at the stages of validation and clinical trials. The underlying reasons may be several, one among which is the inability of the docking algorithm to produce more accurate results.
It should be remembered that even at the stage of biomolecular (protein) structure prediction there are limitations. Traditional computational approaches often produce predictions falling far short of experimental accuracy. This is seen particularly in the cases where the structure of a close homolog (a similar molecule with shared ancestry) of the protein has not been solved experimentally.
Then, what is the remedy?
A ‘naïve’ answer to the question could be – modification of the unsuccessful algorithm or try with a different algorithm. After all, a human “agent” has the necessary “intelligence” to perceive the environment (nature of input and output) and “learn” to make amends. Unfortunately, this strategy will be “humanly frustrating” if the number of required interventions becomes large.
Nevertheless, humans are intelligent enough to recognize this limitation of theirs. Wisely, they transfer the “drudgery”, together with some of their intelligence to a mechanical agent (built with hardware and software) – artificial intelligence (AI) is created.
Machine at the service of humankind
AI essentially refers to the simulation of human intelligence by a system or machine. It encompasses varied areas of investigations and applications – including machine learning (ML). The primary objective of AI is to develop a machine that would think like humans and mimic human behaviors such as perceiving, reasoning, learning, planning, predicting, etc.
What then is the difference between traditional computing systems and those working on ML principles? A traditional computer takes in input data, uses the algorithm/program stipulated by a human programmer to process the data, and subsequently outputs a result.
In the case of a computing system based on the concept of ML, the machine determines, without human intervention, how input data would be processed to predict outcome. Nevertheless, as the title indicates, the machine has to “learn” to do so.
Can a machine learn?
Surely – like animals or humans, machines can also learn through experience, if built with requisite hardware and provided with suitable “learning algorithms”. Such a machine acquires the capability to learn from examples, each example (data sets) containing input data and expected results.
Drawing inspirations from (human) brain science, a powerful machine learning algorithm was developed – artificial neural networks (ANN). ANN do not precisely mimic the function of actual human neuron. Instead, they are based on general mathematical principles that enable them to learn from examples to determine the process (algorithm) by which a given input would produce a given (accurate) output data. The process can then predict (with a very high degree of accuracy) the unknown output when new input data is provided.
Like many other disciplines, the emerging role of AI has brought about a revolution in the field of drug discovery. Starting from target identification, followed by biomolecular structure determination and ‘hit’ screening, to arrival of a successful drug in the market – at each stage AI has introduced unprecedented and almost unimaginable speed and efficiency.