Since the early applications of artificial intelligence (AI) in the 1960s (Gasteiger, 2020), in silico methods have advanced at a remarkable pace, shaping what is now recognized as the modern era of computer-aided drug discovery (CADD) -or, as Dr. Gisbert Schneider aptly describes it, computer-aided bioactive compound discovery. More than ever, computational tools are being intensively employed at virtually every stage of the drug-discovery pipeline. Their use spans from the rapid screening and analysis of scientific literature in basic research to all subsequent stages, including regulatory approval and post-approval follow-up.

Five years have passed since the publication of “Grand Challenges of Computer-Aided Drug Design: The Road Ahead” in Drug Discovery Today, within the section In silico Methods and Artificial Intelligence for Drug Discovery (Medina-Franco, 2021). The objective of the present manuscript is to revisit and update those grand challenges considering the rapid and transformative developments that have taken place in the field since then. In doing so, we comment on recent advances, highlight persistent and emerging hurdles, and identify new areas of opportunity that have arisen because of this accelerated progress.

The manuscript is organized into four main sections. It begins with 1. chemical and biological spaces, followed by 2. methodologies and technologies, and 3. education and human factors, which also play a critical role in drug-discovery efforts. The final section presents 4. summary conclusions. Similar to the previous opinion paper (Medina-Franco, 2021), the challenges discussed and the views expressed on future directions in CADD reflect the authors’ perspectives, informed by the scientific literature and their own experience. The primary aim of this manuscript is to stimulate discussion within the community, recognizing that the viewpoints presented are open to ongoing revision and future updates. Deliberately, the discussion does not focus on AI itself. Over the past 5 years, a substantial body of literature has already addressed both the promises and limitations of AI in drug discovery. Numerous examples - by no means exhaustive - have been reported elsewhere (Jiménez-Luna et al., 2021; Jacobson, 2025; Zhang et al., 2025). Indeed, in recent years several scientific journals have emerged that are devoted exclusively to documenting and critically assessing advances in AI-driven drug discovery.

Chemical space (Reymond, 2025), together with its multiple representations using diverse descriptor sets constitutes a central pillar of modern drug discovery, with growing influence extending well beyond the realm of classical bioactive small molecules. In hit identification, molecular design, and hit-to-lead optimization, the concept of biologically relevant chemical space (Medina-Franco et al., 2025) has continued to expand to encompass regions not traditionally covered by organic small molecules. These regions include, but are not limited to, metal-containing compounds, peptides, macrocycles, molecules targeting protein–protein interactions, and proteolysis-targeting chimeras (PROTACs), among other representative examples that lie outside the conventional “drug-like” space. More broadly, there is increasing emphasis on exploring and populating chemical subspaces that fall within the “beyond rule of five” paradigm (Doak et al., 2016). In parallel, drug-discovery initiatives at the intersection of traditional small organic molecules with peptides and natural products continue to gain momentum, a trend that has been further accelerated by the application of computational approaches (Saldívar-González et al., 2022; López-López et al., 2026).

In recent years, a major development in chemical research has been the enumeration of large and ultra-large chemical libraries using a wide variety of computational approaches (Warr et al., 2022; Korn et al., 2023; Corrêa Veríssimo et al., 2025; Neumann and Klein, 2025). A continuing challenge in this area is the design and enumeration of libraries that are synthetically accessible and experimentally feasible (Korn et al., 2023). Equally important, however, is the need to focus not only on library size but also on library “quality,” encompassing meaningful structural and physicochemical diversity, particularly when the goal is to explore novel or underexplored regions of biologically relevant space.

The identification of therapeutically relevant molecular targets continues to be an active area of research. Regarding the modeling of molecular targets in general, it remains difficult but important to bridge the gap between computationally generated models of molecular targets and their behavior within the intrinsic biological complexity of biomacromolecules (e.g., binding-site plasticity, allostery, induced fit, membrane environments, intrinsically disordered regions, and the influence of cellular context). In parallel, network pharmacology which seeks to validate target combinations and optimize multiple structure-activity relationships or structure-property associations (vide infra) while maintaining drug-like properties (Hopkins, 2008; Zhang et al., 2023) as well as multi-target drug design and drug-discovery efforts focused on neglected and rare diseases, remain topics of significant interest. In addition, precision medicine continues to gain relevance as a guiding paradigm in contemporary drug discovery (Dugger et al., 2018). Modeling and handling toxicity, metabolism, stability under real biological conditions, and the overall developability of drug candidates also remain major challenges. This is particularly true for drug candidates that fall outside the traditional Rule of Five.

Over the past 5 years, data-driven methodologies have gained increasing prominence in drug discovery, particularly those related to AI. However, as noted in the Background section, AI-based approaches fall outside the scope of the present discussion, as they are being extensively and continuously reviewed elsewhere (Jiménez-Luna et al., 2021; Jacobson, 2025; Zhang et al., 2025). Nevertheless, classical approaches, including chemoinformatics, molecular dynamics, and a broad range of other in silico strategies encompassed within CADD continue to play a central role. Benefiting from the growing availability of large, high-value public datasets and advances in machine learning, many of these classical techniques are being substantially enhanced. Notable examples include the docking of massive chemical libraries through so-called deep docking (Gentile et al., 2020), as well as virtual screening workflows augmented by deep learning (Zhang et al., 2024). In parallel, novel methodologies aimed at rapidly generating chemical matter such as DNA-encoded libraries and related approaches, have been developed and reviewed in recent literature (Korn et al., 2023; Hönig et al., 2024). As discussed further below, one of the overarching grand challenges lies not only in increasing the quantity and accessibility of data, but also in ensuring their reliability and overall quality.

Novel paradigms for exploring and characterizing the relationship between chemical structures and their responses, namely, structure–property associations (SPAs) (López-López et al., 2026), have recently been proposed. It is anticipated that, in the coming years, these approaches will significantly influence the development of new analytical frameworks and predictive methodologies.

At the interface with education, discussed in the next section, a key challenge lies in identifying effective combinations of methodologies within drug-discovery projects, rather than pursuing a single “best” approach. As recently discussed, the integration of multiple methodologies - often referred to as consensus or data-fusion approaches in specific contexts - can lead to more robust and reliable outcomes (Medina-Franco et al., 2024). Accordingly, it is advisable to identify appropriate ensembles of methods that strike a balance between established, traditional approaches and more recent methodological developments.

Beyond AI-related technologies, quantum computing continues to hold promise for advancing drug discovery. As recently reviewed in detail elsewhere (Blunt et al., 2022; Santagati et al., 2024; Sood and Pooja, 2024), quantum computing remains in its infancy with respect to practical applications in the pharmaceutical industry, although it is being actively pursued (Santagati et al., 2024). It is widely anticipated that, over the next few years, quantum computing will gain increasing practical relevance in drug-discovery research.

Undoubtedly, scientific progress continues to be driven by effective scientific communication, encompassing traditional publication in peer-reviewed journals as well as conferences, seminars, and online webinars. Online webinars, many of which remain freely accessible in support of the democratization of science, became widespread during the COVID-19 pandemic and have since remained a standard mode of knowledge exchange. In parallel, the scientific community has witnessed a notable increase in the dissemination of preprints (Coudert, 2020), which provide a faster and preliminary means of sharing data, information, and, ultimately, knowledge.

Closely associated with peer-reviewed scientific publications, computational studies are increasingly accompanied by the sharing not only of raw data but also of the code used to generate computational results and implement analytical methodologies. These practices are intended to promote reproducibility or replication, which remains a cornerstone of scientific research (Calnan et al., 2024; Udesky, 2025). Accordingly, scientific journals, particularly those operating under open-access models, continue to align their policies with the FAIR Data Principles (Wilkinson et al., 2016), which state that research outputs should be Findable, Accessible, Interoperable, and Reusable.

Among the challenges in promoting and sustaining the FAIR principles is the need to ensure or at least closely monitor the quality of shared research objects. Another major challenge concerns the long-term maintenance of free and openly accessible repositories capable of storing and preserving massive volumes of raw data, such as those generated by long-timescale molecular dynamics simulations, including multiple replicas. A further challenge is the continued promotion of the dissemination of negative results, which can ultimately enrich and strengthen predictive models. As discussed elsewhere, reporting negative data does not imply the dissemination of unreliable results arising from poorly designed analyses. Rather, it refers to data generated through robust experimental or computational (wet- or dry-lab) studies, grounded in well-formulated hypotheses that ultimately proved to be unsupported or “negative” (López-López et al., 2022). In addition, an ongoing challenge in this context is the sustained promotion of collaborative innovation and effective teamwork, as recently analyzed in the literature (Steinbeck, 2025). At the core of FAIR data, the reliability of publicly available predictions also remains challenging. It is imperative that scientific publications address the uncertainty of estimations, include prospective validation, and, when possible, perform community-assisted benchmarking, preferably supported by experimental validation.

Additional ongoing challenges concern the overall framework of scientific publishing in the era of open science and open access (Arita et al., 2024). These include the costs associated with scientific dissemination, understood here as the investments required to conduct research itself - such as funding for scientists, students, and technical staff - as well as the costs (and profits) incurred by publishers to manage peer review, a process largely supported by the academic community, and to maintain research outputs in an accessible form. These challenges are further compounded by the increasing role of AI in scientific publishing, including both its responsible use and its potential misuse (Salvagno et al., 2023).

Education and the proper training of newcomers to the field continue to represent a central pillar of drug discovery (Varnek et al., 2025). In light of the increasingly widespread - and, at times, exacerbated - use and misuse of AI in science, as well as in other sectors of society, there is a pressing need to educate current and future practitioners in the ethical and responsible application of AI.

Also driven by the current “AI hype,” there has been widespread misuse of the terminology itself. Today, AI in drug discovery is often incorrectly used as a synonym for CADD, computational chemistry, or molecular modeling. While these areas are closely related, they are not interchangeable. Undoubtedly, an overemphasis on “fashion” rather than conceptual rigor can have a negative impact on the precise use of scientific terminology. This represents a significant challenge in education and one that must be addressed from the earliest stages of training (Medina-Franco et al., 2021).

A closely related challenge lies in teaching students a broad spectrum of approaches, both classical and contemporary, while fostering critical thinking and the rational selection of methods based on their underlying principles and assumptions, rather than on their popularity or perceived trendiness. In this context, it is particularly important to encourage students, especially those at early stages of training, to avoid attempting to solve all problems with a single methodology and instead to tailor the problem formulation to the most appropriate analytical or computational approaches.

Another major challenge concerns training students and early-career researchers to articulate their ideas clearly in writing and to report scientific results rigorously. The growing reliance on large language models risks undermining the development of originality and the ability to clearly express novel ideas. Consequently, education should emphasize the cultivation of independent thinking, conceptual clarity, and scientific writing skills. AI tools should be used as supportive assistants not as substitutes for critical reasoning or creative thought.

Five years after the publication of the 2021 Grand Challenges of Computer-Aided Drug Design, the field has progressed substantially in methodological development, data availability, and computational scale. Yet, many of the core challenges identified then remain unresolved, while new ones have emerged. The expansion of chemical and biological spaces, particularly beyond traditional small molecules, has highlighted that size alone does not guarantee relevance; ensuring synthetic feasibility, data quality, and biological importance remains a central bottleneck. Methodological developments have reinforced the relevance of combining complementary approaches rather than pursuing single “best” methods, but issues of reproducibility, data quality, and over-interpretation remain. Also important are challenges beyond algorithms: sustaining FAIR, high-quality data and code sharing, prospective and experimental validation of computational predictions, rethinking scientific communication in the open-science era, and strengthening education to counteract the misuse of terminology and the uncritical adoption of fashionable methods. Looking ahead, addressing these interconnected challenges will require not only technological innovation but also sustained cultural and educational shifts that emphasize critical thinking, responsible tool usage, and scientific rigor across the CADD community. Although computational tools continue to play a substantial role in many steps of the drug-discovery process, as discussed recently, it is unlikely that in silico approaches will be the sole responsibility for the absolute total of the design in the next few years (Medina-Franco and López-López, 2024).