Holy Grails of computational organic and biochemistry and drug discovery
We have our work cut out
In 2017, Houk and Liu wrote an article laying out what they called the Holy Grails of computational organic and biochemistry. We have made significant advances in many of these areas, but there is still a significant amount of progress to be made, which makes them exciting fields of research. Here's the list as defined by Houk and Liu:
1. A universal and highly accurate density functional that applies to diverse chemical systems. Currently 1 kcal/mol is often the required accuracy target, but because an order of magnitude difference in equilibrium constants or rates is caused merely by a 1.4 kcal/mol change in free energy, the authors suggest the ambitious target of 1 kJ/mol.
2. Accurate polarizable force fields. While quantum chemical techniques will continue to make important advances, classical force field would still be required to accurately represent real systems that are studied experimentally: for instance even an invisible droplet of water has 10^18 atoms, too many for full DFT calculation.
3. New molecular dynamics techniques that do not depend on special hardware architectures or special techniques like Markov states.
4. Predicting accurate protein-ligand free energies to binding that are both accurate to within experimental numbers and are cheap and fast enough to prospectively guide experiment.
5. Predicting crystal structures and polymorphs. This is clearly a problem of enormous importance for areas ranging from drug design to materials science. Often the problem is limited by being able to discriminate between several significant crystal orientations in the unit cell that differ very little in energy (0.1 kcal/mol or less).
6. Reaction and catalyst design. Being able to predict reagents, reaction conditions and catalysts for for both liquid and solid-state systems would have a huge impact on industrial chemistry. There has been some important progress in the areas of HTE combined with machine learning, but there's a long way to go before such predictions become standard. Retrosynthesis is a big part of this area as well, and while the tools have gotten better, the lack of negative data in training sets and lower applicability to challenging multi-step synthesis leaves clearly defined areas for improvement.
7. Materials and device design. If we can accurately design the structures, dynamics and properties of multiscale assemblies of molecules, it might not be a stretch to think that we can some day design entire macroscopic systems: devices like batteries, chips and solar cells, and perhaps biological assemblies like cells and organoids that are used in environmental, medical and industrial applications.
To this list I would add a few Holy Grails specifically applicable to drug discovery (polymorph prediction and protein-ligand binding energy prediction would already be part of the list):
1. Accurate and chemistry-independent prediction of ADME properties. Current techniques are often limited by imperfect and dirty training set data and applicable to only specific series of compounds. They are also often poorly predictive of more advanced properties like PgP efflux and clearance.
2. Toxicity prediction. Late stage systemic toxicity prediction (kidney and liver in particular) is still poor, and surprises in these areas still lead drugs to be withdrawn in advanced clinical trials. DILI is one area where machine learning has had some impact; hopefully other areas benefit in the same way.
3. Target prediction: Target validation is still the most important and hardest part of preclinical drug discovery. The key problem is being able to situate a particular target in the complex milieu of multiple proteins of which it forms a small part. Often hitting a target fails because we cannot predict up or downregulation of compensatory mechanisms (this is why in vitro assays are almost always imperfect predictors of in vivo behavior). Being able to do this would be incredibly helpful.
4. Multi-protein-multi-ligand prediction. As progress in predicting protein-ligand conformations and energies has advanced, experimentally the field too has advanced to more and more challenging assembles: PROTACs, glues and other ternary, quaternary and oligomeric complexes. Better crystallography and cryo-EM techniques make resolution of such large assemblies feasible now. While techniques like protein design are having an important impact, being able to predict the structures and energetics of these multi-molecule assembles is still a hard problem.
5. Genetic variability modeling and predicting: Drugs often fail because even the population of patients in clinical trials does not accurately capture the genetic variability encoded in a larger population. This variability can come from ethnic, gender, hereditary and environmental influences. Being able to predict drug responses based on these factors is already a distinct part of what's called pharmacogenomics, but being able to do this accurately and comprehensively would definitely be a Holy Grail.
A related problem is to be able to predict human responses based on animal models. Already we know differences between pharmaceutically relevant animal species and human beings that can lead to drastically different and potentially misleading drug responses; for instance, rats don't vomit, dogs have a "leaky" BBB that can lead to much higher neurotoxicity compared to people, and mice express a very different set of genes in response to immunological drugs. Being able to capture these differences in predictive models so that we can predict the behavior of a drug in a human being based on its behavior in a chosen animal model would be consequential, to say the least.
Ultimately a kind of predictive "supermodel" that ties together everything would be the true Holy Grail of drug discovery. Using this Holy Grail, given a target gene and protein, we would be able to validate it, design a small molecule, predict the affinity and all ADMET properties as well as behavior in several animal models and human beings. This Holy Grail is of course an ideal construct, but making even 20% progress toward this goal in the next two decades would have a tangible impact on the rate and novelty of new drugs for critical health conditions.