Rationality over fashion and hype in drug design

Computer-aided drug discovery

Computer-aided drug discovery (CADD) has become a key technology in drug discovery, providing guidance to experimentalists on which compounds and experiments to focus on next. The capacity of CADD has further increased by the development of powerful machine learning approaches, deep learning in particular.¹ In recent years, software for CADD has become more widely accessible. Today, a range of software packages are available that are open-source or free to use for academic research.^2–5 Together with significant alterations of the scientific landscape induced by the COVID-19 pandemic and the ensuing re-orientation of some early-career but also established research groups towards the use of computational tools, this has boosted the use of CADD methods in particular in academic research environments. However, low-barrier access to computational tools and computing power increasingly leads to the use of CADD techniques also by scientists who have not received formal training on these methods. Many newcomers to CADD employ easy-to-use software without realizing the complexities involved and without being aware of the many potential pitfalls. Improper use of CADD techniques can have a contrary effect on research than intended. The risk of generating meaningless or false predictions is high. Flawed predictions can lead to dedicating significant resources to futile experiments. In particular, publishing invalid predictions can lead to error propagation, eventually resulting in a loss of confidence in CADD. Moreover, uncritical or naive use of artificial intelligence (AI) methods that are being heavily promoted in many research fields has similar negative effects, working against the credibility and acceptance of CADD as a scientific discipline.

Good practices to conduct and report studies in computational medicinal chemistry and chemoinformatics have been outlined in several articles.^3,6,7 Similarly, best practices in different stages involved in CADD have been discussed, for instance, in quantitative structure-activity relationship (QSAR) analysis,⁸ data curation,⁹ molecular docking^10,11 and virtual screening.¹²

In this contribution, we discuss common misconceptions and false expectations associated with CADD, especially in the AI era, and make recommendations on how to avoid common pitfalls when using CADD software. We aim to stimulate an open discussion within the community to help improve our perception and practice of CADD and contribute to shaping its future.

CADD and related fields

Experts from various disciplines involved in drug discovery, such as chemical synthesis and biochemistry, are increasingly making use of computational tools to guide their experimental research and rationalize their observations. This is a positive trend, as developers of CADD tools have been aiming for a long time to make their software more widely accessible and intuitive to use. Prominent examples include web servers, and commercial, free and open-source software. But to develop further, it is of paramount importance to avoid confusion about the concepts and differences between the different areas in drug discovery such as molecular modeling, chemoinformatics, and theoretical chemistry. Conceptual and practical differences between these disciplines are clearly described in the literature.^13,14 Importantly, theoretical disciplines are an integral part of CADD, establishing its scientific foundations. The ability to apply such approaches using software does by no means guarantee that reasonable research is carried out. Therefore, any conclusions or claims should be carefully considered.

Common misconceptions and false perceptions when CADD is superficially viewed

The advent of more advanced computational tools, many of which are open source, freely accessible, and promoted as “easy-to-use,” also increases the widespread use of “buzz words,” and misconceptions among newcomers to CADD.

Table 1 gives examples of incorrect expressions and misconceptions frequently affecting students and researchers with little or no expertise in CADD. Readers and peer reviewers are welcome to comment openly on these points and modify or enrich the list according to their own experience.

When is a study “complete”?

Traditionally, there is a widespread belief in experimental sciences that experimental results represent reality, disregarding the different way in which natural phenomena can be represented and perceived and the relativity associated with varying representations. This ideological attitude works against “out of the box”, hinders intellectual progress, and indirectly de-values scientific disciplines such as CADD. With the rise of AI as one of the most heavily promoted approaches in contemporary society, the academic community has been encouraged to redirect its attention to computational tools to enhance its research impact and appeal. Nevertheless, unconditional trust in “experimental reality” reduces CADD to a “tool provider” and does not regard it as an independent scientific discipline in its own right. Consequently, computational models are often used without the necessary theoretical understanding and the rigor needed to apply them systematically.

In the authors’ opinion, one of the first requirements a new computational practitioner needs to address is realizing that both experimental and computational results are constrained by the model or experimental framework applied to determine them and, in no case, an absolute account of reality. Among medicinal chemists, there is the frequent misconception that purely theoretical or computational studies are in principle “incomplete” because there are no “real” experiments. However, such views require reconsideration and correction, as pointed out above. Rigorous computational studies answer questions that are difficult to address without “in silico experiments”. As such, they are comprehensive and self-contained, regardless of whether the computational approach has led to experiments. “Complete” computational investigations are often consistent with prior experimental observations, but may also chart new scientific territory. Of course, new computational insights leading to experimental work trigger interdisciplinary research. This is a noted strength of CADD, if conducted properly. However, there are misconceptions at interfaces between computation and experiment. For instance, a common malpractice is trying to replace enzymatic inhibition assays with predictions based on molecular docking or dynamics simulations. Another misunderstanding is that black box predictions from machine learning would represent a form of “alchemy”. What we cannot understand is not necessarily incorrect and may have value. The catch is that we are left with making decisions in such cases, for example, about new experiments that go beyond our reasoning and hence require trust in computational work and prior experience. It is also false to believe that AI in its current state would provide solutions to questions that replace our judgment capacity. Data volumes quickly go beyond our comprehension but results of statistical analysis of pattern recognition do not replace human reasoning (algorithms and machines do not “think” -- at least so far). Furthermore, there is a severe misconception that computational predictions might demonstrate or “validate” the bioactivity of compounds. Notably, these and other misunderstandings may not be evident to researchers and students who are just beginning to use computational methods. We encourage the community to avoid judging a computational research project to be “incomplete” because it does not include experiments or to be “complete” just because it incorporates many different computational methods. The question of completeness is not separable from scientific rigor and adequate conduct of methodologies, be they computational or experimental in nature. Furthermore, let us not consider a computational analysis as a “luxury item” to decorate a project report, grant application, or scientific paper with “pretty pictures”.

Using methods for the right reasons

Mainstream media usually disseminate inaccurate or exaggerated reports about the capabilities of computational methods without also mentioning their limitations and flaws. Simultaneously, mass job search engines commonly offer job opportunities with extensive lists of different computational tools as requirements. These factors, among others, continuously put pressure on researchers to increase their productivity and academic credentials to further their careers at the expense of scientific rigor and the quality of research. CADD is not the exception of the increasing trend that disrupts the traditional academic structure in favor of a more market-oriented approach. One of the consequences of this phenomenon is that many young professionals and new CADD practitioners direct their efforts to increase the volume of their curriculum vitae rather than using CADD methods to answer relevant scientific questions.

The popularity of computational methods or tools often jeopardizes rational selection. Newcomers often turn to frequently used methods that are well-validated. However, the justification is questionable if the technique is merely used because it is “popular” (e.g., “follow the crowd” because “it should be right.”). Without properly addressing the question at hand, computational analysis applying irrelevant approaches is misleading or propagates errors. Arguably, one of the most misused guidelines in drug discovery is the Lipinski Rule of Five,¹⁵ which is often confused with assessing “drug-likeness”. Another common pitfall among newcomers to CADD is using docking to predict “real” protein-ligand complexes, given its popularity and easy-of-use. Practitioners should use methods for the right reasons and not just because everybody else is using them. This requires knowledge of underlying theories and sound scientific judgment.

General recommendations for the proper use of CADD resources

In the authors’ opinion, the following recommendations should be helpful to CADD novices and multidisciplinary research teams attempting or planning to use computational approaches to guide drug discovery projects. Similar to Table 1, the list is not exhaustive, but is also intended to stimulate an open discussion within the scientific community.

Concluding remarks

The current pandemic and related funding constraints in some countries and institutions have motivated many researchers at different levels to redirect their efforts from difficult to sustain experimental studies to easy-to-use computational tools that can be employed remotely. Also, reviewer panels of many current grant applications from academia, non-for profit, or the industry currently tend to give priority to research proposals that involve AI. Although this contributes to the popularity of CADD, it also comes at a cost. If uneducated CADD studies enter the realm of science fiction harm is done to this field, its credibility and acceptance, and further scientific development. This must be avoided at all costs. The methods used in CADD should not be applied as black boxes, which can be enabled by just a few hours of hands-on experience such as provided in workshops. Without sufficient understanding of the scope, complexity, and theoretical foundations of these computational methodologies such efforts will inevitably fail and discredit investigators and their work as well as the field as a whole. Newcomers to the area, including students, early-career scientists, and seasoned investigators attempting to re-focus their efforts should be fully aware that, similar to experiments, profound knowledge of CADD concepts and informed use of CADD tools is a must. A simple yet fundamentally important rule applies: “Don’t compute what you don’t understand”. In addition to the general recommendations outlined in this Opinion, we wish to encourage students, newcomers, and practitioners of CADD to use the computational tools and resources for the right reasons, not just because they are easily accessible. Similarly, we highly encourage the scientific community to avoid applying computational methods just because they are popular. Instead, it is strongly recommended to identify scientific questions that can be addressed appropriately using CADD approaches - and avoid others where computational efforts become questionable. In general, computational studies that cannot be reported in established peer-reviewed journals whose scope includes CADD are to be considered with appropriate caution, both by experts and novices to the field. This also applies to the use of modeling web servers. While the integrity of publicly accessible computational tools can be guaranteed by the developers, addressing ill-defined questions or tasks using these tools is beyond their control. Recognizing the benefits of the open post-publication review culture of F1000Research, we would be delighted if this contribution would catalyze open discussions among readers to raise further awareness of latent problematic issues in the CADD area and support its further scientific development.

Data availability

No data is associated with this article.