Introduction

Personalized medicine aims to treat every patient optimally on the basis of each individual's disease and biological characteristics.1-4 Truly personalized therapy is standard in many areas of medicine, eg, surgery, where the straightforward interpretation of a wealth of data for the individual patient guides therapy choice. In contrast, drug therapies act on the complex molecular networks inside cells, as well as on the interactions between cells and organs. Historically, doctors have lacked both information on the detailed biological network in a patient and the capability to correctly predict the response of this complicated system to the complex action of a drug, essentially ruling out true personalization of drug therapy with current techniques.

To improve the chance that a specific therapy will help a patient, we have used stratification approaches based on statistical correlations between specific biomarkers5 or signatures and a positive or negative outcome of therapy, albeit with serious limitations. Despite intensive research, few biomarkers have shown real clinical benefit,6,7 often with relatively modest sensitivity/specificity values. In this regard, we have to distinguish between “causal” and “network” biomarkers. Causal biomarkers directly recognize a unique molecular state that is directly addressed by a selective drug; cancer-specific examples include the presence of a fusion protein against which a specific drug is available, in the case of crizotinib (Xalkori),8,9 and the use of ERBB2 (formerly HER2 or HER2/neu) amplification status as a biomarker fortrastuzumab , a drug targeting human epidermal growth factor receptor 2 (IIER2).10 Such directly targetable mechanisms are not very common, leaving many patients without this option. Network biomarkers are not directly linked to a drug target, and uncertainties are often much greater, further reducing the correlation between the biomarker result and drug response. Even when we have a biomarker, serious problems exist, including uncertainty in the prediction due to the enormous complexity and heterogeneity of diseases—eg, the usually unknown heterogeneity of the tumor and the rapid recurrence of tumors in resistant form even after (temporary) “miracle” cures.11-13 True personalization of drug therapy will therefore require much more than current attempts at stratification.

In similarly complex situations with dangerous and/ or expensive consequences, we characterize the situation in sufficient detail to make predictions possible and essentially rebuild the situation as exactly as necessary in mechanistic models of reality in the computer, allowing us to make unavoidable mistakes safely, quickly, and inexpensively on the computer rather than in reality. Such mechanistic models require three basic components to make accurate predictions of the future development of complex systems (with or without disturbances): (i) information about the detailed rules governing the processes to be modeled; (ii) the necessary computing power; and (iii) a detailed characterization of the situation at the start of the modeling.

From decades of research, there is a growing knowledge base on the processes relevant for a number of diseases (some more than others), with cancer being a particularly good example.14,15 Computing power has continued to increase by a factor of roughly a thousand every 10 years, providing rapidly increasing computing resources.16 However, every model is only as good as the data available to initialize it1—data that up to a short time ago simply have not been available on individual patients and their diseases. However, this has changed, primarily due to developments in DNA-sequencing technologies. Whereas the first human genome sequence took more than 10 years and billions of dollars to complete, we can now determine multiple genomes per machine run at a cost close to $1000 per genome.18 DNA sequencing can be used to decipher the patient genome, transcriptome, epigenome, and metagenome and thus directly determine much of the information we need to characterize the relevant biological networks in any disease.19-23 Sequencing techniques can also be applied to analyze other types of biological information (eg, levels of proteins, protein modifications, protein complexes, status of the immune system) by appropriate experimental strategies.24-27

This type of information about a disease in an individual patient allows us, for the first time, to develop personalized computer models of many different diseases. For example, in oncology, we generate computer models of the individual tumor and patient on the basis of a comprehensive omics analysis of both, allowing us to predict effects and side effects of (mechanistic) drugs. However, the approach is not restricted to cancer as long as the same three requirements above are fulfilled, albeit with different challenges; eg, information on molecular mechanisms of disease and drug therapy are, in general, available in cancer but often missing in neuropsychiatry diseases. Moreover, a sufficiently detailed molecular characterization of the disease in an individual patient is usually possible in cancer (eg, via biopsies) but may be more difficult in brain diseases due to sampling issues. Nevertheless, there is very encouraging progress in these disease areas, including from a number of large research projects that are characterizing basic mechanisms (eg, the Human Brain Project), as well as the development of innovative solutions for sampling issues.28-30 Thus, the virtual patient model outlined here, exemplified using cancer because of the advanced knowledge in the field on potential disease mechanisms, holds promise as a model for neuropsychiatry and other complex diseases, as well as for prevention and wellness applications.

The way forward: what information do we need?

In many diseases, there is an enormous complexity of processes we are trying to affect by drugs, which in turn have complex mechanisms of action.31 Oncologyis an obvious first model for this approach as, unlike for other diseases, a large amount of information on the biological networks and molecular processes acting in tumors (and other tissues) is available, helping us to understand the disturbed processes in tumors. Also, we often have access to the diseased tissue—or, in the form of a liquid biopsy, to tumor-derived material32-35—and can therefore analyze its biology by high-throughput (omics) techniques. Such analyses often identify somatic changes that have dramatic effects, which either lead to already well-characterized (and published) consequences or induce functional changes, which can often be deduced by predicting changes in protein structure and the expected functional consequences.24,25,36,37

We have made huge leaps forward in the diagnosis and treatment of cancer, but with the significant molecular heterogeneity apparent between individual patients, their tumors, and even between cells within a tumor, the enormity of the task of “curing cancer” is becoming apparent. The possibility to generate individual models of cancer patients has, in particular, been driven by the development of next-generation sequencing (NGS) techniques. NGS can be used to analyze the genome/exome of both tumor and patient, providing information on sequence variants arising by a range of mechanisms, eg, mutation, deletions/insertions, or rearrangements. Sequencing or chip-based analyses are typically used to characterize genome methylation patterns, whereas RNA sequencing is the best method to analyze the abundance and processing of the different types of RNA molecules (coding, noncoding, microRNA).23,36,37 Similarly, a number of approaches can be used to analyze proteins and protein modification states (eg, mass-spectrometry techniques, reverse-phase protein arrays, proximity extension/ligation assays), as well as metabolite levels, providing key information on the results of posttranscriptional regulation in the biological networks.25,26,38-42

Spatially resolved and single-cell analysis techniques

Nevertheless, even in tumors, bulk omics analyses only provide part of the information we would need. Tumors can differ from the normal cell by hundreds to thousands of potentially relevant changes, in combinations that are unique to the tumor and often even to tumor cells. Combined with differences in the underlying germ-line genome, as well as in the epigenome and transcriptome of the different somatic cell types from which the tumor has originated, we have an irreducible complexity that can only be ignored if we accept correspondingly lower success rates in selecting an optimal therapy for the individual patient. Moreover, many tumors are heterogeneous, with differences affecting the predicted therapy response, rendering smaller or larger fractions of tumor cells inherently resistant to a therapy.11,12 If only a fraction of the tumor cells stop growing, resistant cell populations will continue to multiply, counteracting the effect of the treatment.13

In an ideal situation, all analyses mentioned above would be carried out either in a spatially resolved fashion or on the single-cell level by disaggregating the tumor. Protocols for sequencing and imaging-based spatially resolved transcriptome sequencing have been developed.43-46; moreover, genome and transcriptome analyses and protein detection by proximity extension assays have been successfully carried out at the single -cell level.46-50

Immune therapies, immune status

With recognition of the key role the immune system plays in many disease processes, the field of immunotherapeutics has recently emerged.51 We are still in the discovery phase in identifying potential biomarkers predictive of response to these drugs beyond obvious candidates, such as the number and expression levels of mutations altering peptide sequences potentially displayed by human leukocyte antigen (HLA),as well as the possible effect of programmed death-ligand 1 (PDL1) expression on the response to anti-PDLl treatment.52,53 Characterization of an individual's immune system would also be important in attempts to truly personalize therapies in many diseases, including cancer53 and neuropsychiatric disorders such as depression and schizophrenia; neuropsychiatry disorders are a particular focus given the increasing interest in the role of the immune system in these disorders.

As a step in this direction, we have developed a new approach to identify the combination of antibody (heavy and light chain) or T-cell-receptor α and β chains in a blood sample.27 This method is based on the amplification of RNA from microsomes in an emulsion polymerase chain reaction system, under conditions in which the primary amplification products (heavy and light chains for B cells, α and β chains for T cells) join during the amplification step, providing a single, sequenceable DNA molecule containing the relevant information from both chains. Deep sequencing (sequencing of the transcriptome at sufficient depth to provide information on both the sequence and frequency of the messenger RNAs [mRNAs] present) of this material will not only provide information on the sequence of the mRNAs coding for relevant antibodies/T-cell receptors but also on the combination of chains present in the analyzed cells.

Haplotype sequencing

We have access to molecular data on tens of thousands of human genomes (healthy and diseased), but most of these are generated as “mixed diploid” sequences. Because of the diploid nature of the human genome, if we really want to understand how genetic variation is linked to gene and genome function, phenotype, and disease, then determining each of the haplotype sequences of a genome is essential. Anticipating the importance of analyzing “phase-sensitive” variation (ie, reflecting the inherently diploid biology of genes and genomes) for personal genomics and individualized medicine, investigators have carried out a recent analysis of multiple haplotype-resolved genomes, revealing the extent of sequence variation that exists between the two alleles of human individuals, with many genes harboring sequence variants that are apparent at the amino acid level, increasing the repertoire of proteins that are expressed.54

Model-driven personalized therapy in other disease areas

After cancer, it is expected that many other diseases (and prevention of diseases) will be addressed. For neuropsychiatry disorders and neurodegenerative diseases, challenges to establishing individual patient models remain. One major bottleneck is a lack of information on disease mechanisms, almost completely missing for disorders like schizophrenia, and partially missing for many neurodegenerative diseases, eg, Huntington and Alzheimer disease. Even for Huntington disease (the first neurodegenerative disease for which the primary cause was identified, approximately 20 years ago55), we still do not understand the translation of the primary cause into the disease phenotype. A much deeper understanding of the molecular basis of such diseases is required before we can apply the same technologies. To this end, large-scale initiatives—eg, the Human Brain Project, the Blue Brain Project, The Human Connectome Project, and the Virtual Brain—have been initiated, and efforts to generate omics data, an integral part of disease characterization, are gaining traction, albeit not yet on the scale we see for cancer. Until more mechanistic information is available, we should be able to use hybrid models combining a mechanistic core with statistics/machine learning components to compensate for the current lack of mechanistic knowledge.

A more serious problem could be the difficulty of accessing diseased tissues, eg, the brain, to generate molecular data. If data from other, more easily accessible tissues such as blood are not sufficient, alternative strategies are required. Owing to rapid progress in single-cell sequencing, molecular data from hard-to-sample areas could be generated from the few cells clinging to tools used to introduce electrodes into affected areas, eg, in brain-stimulation therapies.28 In parallel, surrogate tissues could be developed from individual patients by reprogramming induced pluripotent stem (iPS) cells; eg, skin cells could be developed from neuronal cells for further molecular analysis.29,30 Such data would be complemented by imaging and increasingly powerful sensor data, the latter providing large amounts of information, possibly on a continuous basis.

Despite these hurdles, the knowledge base is substantially increasing, thus expanding the possible uses of modeling to develop a mechanistic understanding—the first crucial step in the personalization of therapy in these diseases.

Data, data, data

Although in oncology much of the mechanistic information we consider in virtual patient models will be molecular (simply because drugs act as molecular entities), there is also other mechanistic information (eg, on blood flow and blood pressure) of high relevance to modeling patients in other disease areas (eg, heart disease). Comprehensive clinical data—including information on factors that may influence disease trajectory and response to drugs, such as lifestyle factors, comorbidities, and the microbiome—all act to complement the existing molecular-level data to construct a more comprehensive picture of an individual's disease. Imaging data, an ever more relevant data source in neuroscience, will become increasingly incorporated into virtual patient models. Neuroimaging and sensor techniques are providing a route, not only toward measuring and monitoring important functions within the body and brain, including behavioral responses, but also for assessing functional changes resulting from disease processes, injury, environmental stimuli, or response to treatment, for example.56,57 In oncology, imaging technologies are generating data that can be easily associated with molecular data.58; for instance, metabolic activity data from specific tumors generated using positron emission tomography scans can be straightforwardly linked to molecular data generated both on bulk tumor and single cells, eg, spatially resolved transcript-tome analyses from biopsies from the same tumor (see above). The merging of data from different sources will help to improve the mechanistic basis of the computational model and to optimize the predictions made.

The “virtual patient/in-silico self” model

The best (and probably only) way to integrate the very large data sets, necessary to be generated for every single patient as a prerequisite for truly personalized therapy and prevention, will involve the generation and use of sufficiently detailed computational models of the disease process in the individual patient. These computational models can then be “virtually” treated with all drugs or drug combinations. For this, we have developed PyBiosS (http://pybios.molgen.mpg.de),59,60 a modeling engine written in Python, an object-oriented modeling language, with objects corresponding to the components of the biological networks we are trying to model. Using PyBios3, we have developed a large and comprehensive model of cell signaling transduction and associated processes—ModCellTM.60-63

This system simplifies the construction of the very large models we need to represent the complex processes in the patient by allowing the assembly of larger models from submodels representing individual pathways or sections of pathways. Models initially represent the biological networks of normal cells and have two major types of components: (i) the basic structure of the model, eg, protein species and their biochemical complexes, phosphorylation, activation/inhibition, transcriptional induction; and (ii) the appropriate kinetic laws and (if available) parameter values derived from, eg, pathway databases and/or the scientific literature.

Model individualization

Models are individualized using omics (and other) data generated through a comprehensive, highly qualitative, and deep characterization of a patient's samples (for cancer patients, in particular the tumor). For the characterization of cell signal transduction pathways, measurements on multiple layers is helpful: whole-exome sequencing provides insight into loss- or gain-of-function mutations; the proteome offers a comprehensive picture of protein expression, whereas the phosphorproteome indicates the activity state of the signaling pathways; and transcriptome data helps to estimate synthesis and decay rates of mRNAs and proteins and will enable evaluation of turnover rates for different molecular species. Moreover, transcriptional signatures can be linked to modules of genes that are located downstream of a specific signaling pathway or indicate crosstalk effects between different upstream pathways.

However, the full range of data types mentioned above are not always available for every patient. For cancer patients, we typically use genomic data (eg, exome or whole-genome sequence data) from both tumor and germ-line samples, complemented by transcriptome data (ideally from RNA sequencing [RNA-Seq]experiments), allowing the expression of any identified mutations and other alterations to be cross-checked. The integration of other data types from the same patient and samples (eg, epigenome, proteome, and metabolome) further defines the individual disease at the molecular level and provides further data to initialize the models.

The molecular alterations identified within the tumor (by comparison, eg, with the patient's genome) are included in the model, resulting in objects with altered/ inactivated function; eg, if an activating ras mutation is identified in the tumor genome and the mutation is expressed in the RNA sequencing data set (and we have no protein data indicating otherwise), we assume that a mutant ras protein is present in the tumor, represented in the model by a corresponding object with appropriate changes in its function. Drugs are also objects that are added to the model on the basis of their known dissociation constants and target protein(s), resulting in, eg, inactive target-drug complexes.

To identify the optimal treatment for a specific patient, a variety of computational indicators are used; for instance, in cancer patients, the computed levels of MYC or phosphorylation status of tumor protein 53 (p53), cleavage of poly(ADP-ribose) polymerase 1 (PARP1), and GTP-loading status of Ras-related C3 botulinum toxin substrate 1 (RAC1)and cell division control protein 42 homolog (CDC42)serve as proxies for corresponding phenotypic effects, such as cell proliferation, senescence, migration, and apoptosis induction.

Beyond mechanistic information: hybrid models of the virtual patient

Although oncology provides a wealth of mechanistic information to incorporate into models, this is not sufficient for many applications. Although we can mechanistically predict the concentrations of protein and protein modification states involved in apoptosis, we do not have accurate mechanistic models describing the phenotypic response of cells to a specific molecular configuration. This is even worse in many other disease areas. For many neuropsychiatric diseases, we are still far from having molecular models describing the disease process and drug action. Many nondrug-based treatments, eg, electroconvulsive therapy and psychotherapy, do not act through well-defined molecular mechanisms; the drug therapies used in parallel act through known molecular mechanisms. To incorporate such nonmechanistic components into our mechanistic models, we can define “pseudo-objects” as part of the object-oriented modeling system, encapsulating, eg, neuronal or Bayesian networks trained to translate nonmechanistic components into mechanistic consequences. As our knowledge increases, these can be replaced by more robust mechanistic model components.

Model optimization and validation

To be able to represent relevant changes observed in the individual tumors and the different targets for many drugs, models have to be quite detailed. This information may not be directly perceivable or may be hard to obtain experimentally (eg, drug-binding affinities, dephosphorylation, and degradation).

As models grow in complexity, more parameters have to be specified; eg, the large-scale network represented by ModCellTM, covering over 45 signaling pathways and hundreds of genes and their interactions, leads to tens of thousands of parameters.60-63 For this, we are increasingly relying on reverse engineering/parameter optimization strategies, minimizing the difference between prediction and experimental results, eg, drug testing on experimental models (patient-derived organoids, xenografts, and iPS cells3,29,30), and as far as possible, patient treatment results. Results from such experimental models have a crucial role to play in refining the accuracy of and validating model predictions, generating invaluable data on disease mechanisms and response to drugs. This process is still at an early stage, but as time goes on will deliver increasingly accurate predictions in a self-learning model.

To optimize the choice of parameter, statistical techniques are deployed, eg, memetic algorithms based on local search chains (MA-SW-Chain),64 to minimize the difference between predicted and experimental data, as well as derived predictions, improving the choice of parameter space and thus the accuracy of the model. To minimize the effect of overtraining, observed in many situations in which we have many free parameters but comparatively few observations, the available data are divided into training sets, used to identify optimized parameter vectors, and test sets, providing an unbiased estimate of the performance of these vectors on an independent data set. In tandem, uncertainty and sensitivity analyses can be conducted to assess how reliable model predictions are. Uncertainty analysis exploits statistical techniques, such as Bayesian and frequentist estimation methods, to assess the effects of lack of knowledge or potential errors in the model.65 Sensitivity analysis allows the identification of key data and assumptions that have most influence on model outputs.66

Evaluation of this strategy has revealed the potential of the mechanistic modeling approach, based on comprehensive molecular characterization of patient and tumor, to identify patient-specific responses to microRNA-based treatments for colon cancer61 and is generating promising results within clinical studies (eg, TREAT20[Tumor REsearch And Treatment - 20 patient pilot] and its follow-up project, TREAT20plus].

Treating the virtual tumor with a drug or drug combination

Within the model, a drug is represented as an object that binds to one or more (usually many) target objects with specific dissociation constants, forming a drugtarget object with specific changes. This change can be simple, eg, a protein kinase inhibitor binding to specific protein kinases forming inactive complexes, but can also be more complex, as in the case of a drug blocking activation of its target protein. Drug combinations, probably to become increasingly important in clinical practice,67 will result in the formation of the appropriate complexes for both drugs. Specific schedules of administering drugs in a combination therapy can similarly be represented in the modeling of the tumor response.

Role of the liver: pharmacogenomics and the tumor model

Individual patients react differently to drugs, due to differences in the metabolism or activation of the drug, determined by their cytochrome C genes68 and other variants in their genome. To represent this, we can introduce a “liver” model into the overall system, which will metabolize the drug appropriately, mimicking the multifactorial processes determining drug bioavailability and degradation. Drug and drug metabolite concentrations generated over time then interact with the “tumor” model.

Drugs also affect normal tissues, not just tumors

To estimate possible (side) effects of drugs or drug combinations on normal tissues of the body, we can model the different normal cell types of the patient by combining information on relevant variants, deletions, and copy number variants, etc, identified in the patient genome with information on tissue- or cell-specific gene expression patterns (eg, from the Illumina Body map). As with the tumor, these tissue/cell models can then be (virtually) exposed to the drug and drug metabolites in the virtual circulation to identify potentially unacceptable side effects of specific therapies.

Interactions of tumor cells with each other, soma cells, and the immune system

As knowledge accumulates on the molecular mechanisms underlying interactions between tumor cells, their surrounding soma cells, and infiltrating cells of the immune system, “virtual” tumors will increase in complexity, represented by individual cellular models interacting through the exchange of signals, etc. Also, different regions of tumors are exposed to different oxygen and nutrient levels, depending on proximity to blood vessels, and will therefore differ in their biology. As we learn more about interactions between the immune system and tumors, this knowledge can be represented in future models, based on the growing information base and the increasing use of immunomodulation in oncology.51,52

Synthesis A: a truly personalized medicine

We are confident that a truly personalized drug therapy in oncology (outlined in (Figure 1). based on a deep omics analysis of tumor and patient will rapidly become the standard for cancer patients and will serve as a model for applying similar techniques in many other therapeutic areas. This will not only allow the identification of unexpected drugs (not normally used for cancer treatment), but also the identification of patient-specific combination therapies. New, model-based tools can allow the doctor treating the patient to visualize and interactively explore the effects of drugs/drug combinations at the molecular level within individual patients (Figure 2). The modeling approach provides information on the underlying biological pathways relevant for many diseases, with potential benefits for identifying specific disease endophenotypes and associated biomarkers, as well as comorbidities that have a genetic rather than an environmental/lifestyle basis. An integrated approach is required, comprising information on multiple levels, from clinical data and lifestyle factors to imaging techniques that provide a more global view of interactions and help to refine the model's underlying knowledge base.

Figure 1.
Figure 1. Path to prediction. Tumor biopsies and control (usually blood) samples are taken from individual patients for deep molecular analysis. Complex omics data are generated (eg, genome/exome, transcriptome, possibly proteome) and analyzed comprehensively for tumor-specific alterations, eg, mutations and gene fusions, loss of heterozygosity, ploidy, and 3D protein modeling of protein mutations, etc. Data generated are mined for clinically relevant features, such as diagnostic or prognostic markers. The patient-specific data are mapped to cancer-relevant pathways within a large-scale mechanistic computational model of cell signaling transduction and associated processes (eg, ModCellTM60-63). Individualized models are used to predict in silico the response of individual tumors to drug(s), singly or in combination, to identify the optimal therapeutic strategy for a specific patient. LOH, loss of heterozygosity; NGS, next-generation sequencing.
Figure 2.
Figure 2. (Opposite and overleaf) Visualizing drug action. Illustration of the effect of afatinib on receptor tyrosine signaling (eg, EGFR) and major downstream pathways, such as the MAP kinase cascade and PI3K/AKT signaling. (A) Diagram of the model network, illustrating its complexity. Colors of nodes represent fold change in concentration of model species (see scale) due to inhibition by afatinib as predicted by the in silico model. (B) Afatinib's major target is the EGF/EGFR signaling pathway. Visualization of one branch of this pathway shows the strong inhibitory effect of afatinib on downstream network components (shown in green). ABL1, Abelson murine leukemia viral oncogene homolog 1; ADAM10, A Disintegrin and metalloproteinase domain-containing protein 10; AKT, protein kinase B; EGFR, epidermal growth factor receptor; ERBB2, formerly HER2 or HER2/neu; MAP, mitogen-activated protein; PI3K, phosphoinositide 3-kinase; PLCG1, phospholipase C γ 1 SHC1, SHC (Src homology 2 domain containing) transforming protein 1; STAT, signal transducer and activator of transcription.

However, a molecular model of even a single cell can be daunting in its complexity. It is therefore hard to imagine that we could, in the foreseeable future, represent the patient completely with an estimated 1014 interacting cells. We will have to compromise and model the patient as interacting molecular models representing the essential components that probably affect treatment success. In cancer, these would represent the different cell populations of the tumor (representing tumor heterogeneity, stroma, invading immune cells, blood vessels, etc); the liver as one single idealized cell carrying out drug metabolism and activation, depending on the cytochrome C alleles and other variants in the patient genome; and the most important cell types of the body (eg, neuronal cells, heart cells), to identify possible side effects of drugs in normal tissues. Ideally, the patient's immune system would also be represented, essential for autoimmune or infectious diseases, but also of probable relevance for predicting response of tumors to immunotherapeutics. As our knowledge and technical capabilities increase, such virtual patient models will increase in complexity. In the future, it could become routine practice to generate “preclinical” body maps from every patient (from iPS cells) for molecular analysis. Data of this type could become enormously important for further improvement of the individual patient models, but it would also provide access to “surrogate” brain samples for molecular analysis.

The virtual patient/in-silico self models could also find applications in addition to personalized therapy and prevention. If sufficiently predictive, they could become an important component of our health care system and also our lifestyle, ultimately developing into a “guardian angel,” accompanying every individual from before birth into old age, updated both from medical and molecular data generated at intervals at, eg, medical visits, but also more or less continuously through increasingly powerful sensor systems. Such in-silico self models have the potential to not only help doctors identify the optimal therapy and develop truly personalized prevention, but also to allow us to interact more intelligently with our bodies and possibly even our minds (see www.healthcarecompacteurope.eu and www.futurehealtheurope.eu for further information regarding potential impact on health care systems).

Synthesis B: virtual clinical trials, virtualization of drug development

The same technology can also be used to virtualize large parts of the drug development process, allowing virtual clinical trials as soon as molecular data on the binding specificity of a drug candidate are available (through the use of docking programs, possibly even on candidates before synthesis). These virtual clinical trials can increase the effectiveness of real clinical trials and improve the likelihood that drugs that have already failed during late stage clinical development due to lack of efficacy in a nonstratified clinical trial are approved rapidly and (relatively) cheaply in patients selected as responders to the drug.69 Again, virtualization based on detailed mechanistic models is likely to happen first in cancer and other diseases where we have detailed knowledge of molecular and cellular disease mechanisms; nevertheless, a similar approach to diseases other than cancer, such as neuropsychiatry disorders, will hopefully follow relatively soon given the enormous impact these diseases have on individuals, their families, and on society as a whole.