Historically, drug development has been rooted in the identification of molecular targets—proteins whose modulation promises therapeutic benefits for diseases. This methodology, while effective in many contexts, falls short when diseases lack well-characterized targets or involve complex phenotypes that defy simplistic molecular intervention. Addressing these limitations, the IRB Barcelona team, under the guidance of Dr. Patrick Aloy, developed a strategy grounded in phenotypic discovery, where the molecule’s desired biological effect guides its design rather than adherence to a fixed molecular target.

This paradigm shift capitalizes on observable cellular responses as the blueprint for molecule generation. By prioritizing differential impact on cell populations—for instance, targeting pancreatic cancer cells with minimal effect on healthy controls—the researchers unlock avenues for precision therapeutics that evade the pitfalls of conventional target-centric drug design. The challenge, however, lies in predicting and crafting chemical entities capable of such nuanced cellular specificity, a problem exquisitely suited for artificial intelligence methodologies.

To establish a robust foundation for their AI-driven platform, the team initiated an extensive experimental campaign. Over 11,000 chemical compounds were systematically screened across eight distinct cell models, including six pancreatic cancer-derived cell lines and two normal control lines. This exhaustive bioactivity dataset formed the cornerstone for the construction of predictive algorithms, outperforming classical methods reliant on chemical structural similarity by learning direct correlations between molecular features and cell-type-specific responses.

Integrating these predictive models into a generative AI architecture, the researchers equipped the system to propose novel chemical compounds that satisfy a dual criterion: potent activity against target cancer cells and a diminished effect on normal cells. Beyond mere activity prediction, the system navigates the expansive chemical space to innovate structurally unique entities, transcending known compound libraries and expanding the potential for first-in-class drug candidates.

Crucially, the AI-designed molecules underwent rigorous experimental validation in the laboratory, confirming selective efficacy in the intended cell models. Several compounds not only met but exceeded the performance benchmarks of traditional screening-derived molecules, demonstrating both enhanced selectivity and biological activity. This success underscores AI’s potential to invert the conventional discovery funnel, enabling a more efficient and targeted generation of therapeutics without prior dependence on established molecular targets.

This approach also represents a significant leap toward addressing diseases that have historically been refractory to drug development. By circumventing the necessity for predefined targets—which may be undiscovered, non-druggable, or involved in complex biological networks—the AI framework offers a flexible, scalable solution applicable to a broad spectrum of pathological contexts, particularly those with heterogeneous cellular landscapes.

From a technical standpoint, the integration of predictive bioactivity models with generative chemistry leverages machine learning techniques to capture intricate molecular-biological interactions. Predictive models utilize multi-dimensional chemical descriptors and cellular response data to forecast activity profiles, while generative models employ neural network architectures to synthesize candidate molecules iteratively optimized for the desired phenotypic effect. This dual-layer framework embodies an adaptive learning system capable of refining compound design based on theoretical and empirical feedback loops.

Moreover, the resultant molecules exhibit structural novelty, often diverging significantly from known chemical scaffolds, thereby enriching the diversity of drug-like candidates and mitigating intellectual property challenges common in drug development. This structural innovation is critical, as novel scaffolds can display improved pharmacokinetics, reduced off-target effects, and enhanced efficacy, qualities essential for advancing new therapeutic agents into clinical pipelines.

The research team’s strategy not only accelerates the identification of bioactive compounds but also enhances the precision of therapeutic targeting. This is particularly beneficial in oncology, where selective cytotoxicity against tumors while sparing healthy tissue remains a paramount objective. Implementation of such AI-powered frameworks could revolutionize personalized medicine approaches by tailoring molecular interventions to specific cellular phenotypes observed in individual patients.

This study, published in Communications Chemistry, reflects a milestone in merging computational intelligence with experimental pharmacology. While still at an early stage, the methodology holds promise for reshaping drug discovery workflows, reducing reliance on exhaustive high-throughput screening campaigns, and fostering more directed, efficient therapeutic innovation, particularly in complex and poorly understood diseases such as pancreatic cancer.

Looking forward, this research opens pathways for further refinement of AI-guided compound generation, including integration with multi-omics data, incorporation of 3D structural considerations, and adaptation to dynamic cellular environments. By continuously enhancing model fidelity and expanding experimental validation, such frameworks could bridge the gap between in silico predictions and clinical reality, ultimately leading to safer and more effective therapies.

In summary, IRB Barcelona’s innovative combination of predictive and generative AI for molecule design marks a visionary shift from target-centric to effect-driven drug discovery. This approach represents not just an incremental advance but a fundamental reimagining of how molecules are conceived, designed, and validated, potentially accelerating the arrival of new medicines to patients with unmet medical needs.

Subject of Research: AI-driven phenotypic molecule design for selective cellular targeting in drug discovery.

Article Title: AI-Enabled Design of Selective Molecules Based on Cellular Phenotypes Without Predefined Targets.

News Publication Date: June 26, 2026.

Web References:
DOI: 10.1038/s42004-026-02071-x

Image Credits: IRB Barcelona

Keywords

Artificial intelligence, Generative AI, Drug design, Drug discovery, Machine learning, Pancreatic cancer, Artificial neural networks

The research team synthesized and meticulously tested 31 ketamine-derived compounds by strategically modifying specific molecular sites: the chloro substituent on the aromatic ring, the methylamino group linked to the cyclohexanone ring, and the sixth position on the cyclohexanone ring, which serves as a primary locus for metabolic hydroxylation. This comprehensive chemical redesign aimed to pinpoint analogues that outperform ketamine, the current gold standard in rapid-acting antidepressant treatment, by enhancing adenosine modulation.

To assess these compounds’ functional impact, the researchers employed fiber photometry—a cutting-edge technique allowing real-time monitoring of extracellular adenosine fluctuations directly within the mPFC of living mice. This innovative use of adenosine dynamics as a biomarker enabled the identification of analogues capable of triggering robust and sustained adenosine surges. Among the compounds tested, two dechlorinated derivatives, deschloroketamine (DCK) and deschloro-N-ethyl-ketamine (2C-DCK), stood out by significantly amplifying adenosine release at doses as low as 2 and 5 mg/kg, surpassing ketamine’s effects observed at 10 mg/kg doses.

Notably, the superior adenosine-modulating properties of DCK were evident even at the lowest tested dose of 2 mg/kg, marking a substantial leap in potential therapeutic efficiency. This dose responsiveness underscores the compound’s promising pharmacodynamic profile, suggesting that effective antidepressant action could be achieved with markedly diminished systemic exposure, potentially minimizing side effects.

To investigate the functional consequences of heightened adenosine release, the study utilized behavioral paradigms widely accepted in psychiatric research: the forced swim test (FST) and the sucrose preference test (SPT). These assays, performed in mice subjected to chronic restraint stress to model depression-like states, revealed that DCK exhibited robust antidepressant-like effects at doses significantly lower than those required for ketamine. Specifically, DCK administered at 2 mg/kg elicited comparable amelioration of depressive behaviors relative to 10 mg/kg ketamine, with heightened efficacy observed at 5 mg/kg.

Parallel evaluations of 2C-DCK mirrored these findings, demonstrating potent antidepressant efficacy at 5 mg/kg, while 3’-chloro-ketamine, a structurally distinct analogue that failed to evoke substantial adenosine surges, showed no behavioral improvement even at the highest doses. This clear correlation between adenosine modulation and antidepressant efficacy solidifies the role of extracellular adenosine dynamics as a predictive biomarker for therapeutic potential in novel ketamine derivatives.

Crucially, the study also addresses safety considerations by evaluating the propensity of these analogues to induce hyperlocomotion, a behavioral proxy for dissociative side effects commonly associated with ketamine. DCK, at its effective antidepressant dose of 2 mg/kg, produced only mild increases in locomotor activity, contrasting the significant hyperlocomotion induced by 10 mg/kg ketamine. This finding suggests a wider therapeutic window and a possibly improved side effect profile for DCK, enhancing its clinical appeal.

In dissecting the mechanistic underpinnings of these observations, the research investigates the relationship between N-methyl-D-aspartate receptor (NMDAR) antagonism—a well-established mode of action of ketamine—and adenosine release. By systematically comparing the in vivo adenosine-inducing capacity of ketamine and six analogues with their corresponding in vitro NMDAR inhibitory IC50 values and brain pharmacokinetic profiles, the authors discovered a striking dissociation.

Specifically, no direct correlation emerged between the degree of NMDAR blockade and adenosine surge magnitude. This was exemplified by 3’-chloro-ketamine, which potently inhibited NMDARs without triggering adenosine release, in contrast to 3C-DCK, which elicited strong adenosine responses despite comparable NMDAR affinity. These results decisively indicate that NMDAR antagonism is not the primary driver of extracellular adenosine elevation.

Supporting this interpretation, prior parts of the study demonstrated that ketamine exerts direct modulatory effects on mitochondrial metabolism, a non-NMDAR pathway, which appears to orchestrate adenosine dynamics. This novel insight pivotally shifts the focus from classical glutamatergic hypotheses toward purinergic signaling as a central mediator of ketamine’s antidepressant actions.

Overall, this study exemplifies the power of integrating chemical synthesis, advanced in vivo neurochemical monitoring, and behavioral pharmacology to unravel complex therapeutic mechanisms. By identifying adenosine signaling as both a biomarker and a mediator of antidepressant efficacy, the researchers provide a compelling rationale for developing ketamine analogues with optimized purinergic profiles, offering hope for rapid-acting antidepressants with reduced side effects.

This research not only broadens our understanding of ketamine’s multifaceted pharmacology but also charts a promising course for next-generation antidepressant drug development. As depression remains a leading cause of global disability, breakthroughs that enhance treatment efficacy while minimizing adverse effects represent a transformative step forward in psychiatric medicine.

Future exploration will undoubtedly focus on further elucidating the interplay between mitochondrial function, adenosine signaling, and neuronal circuitry in mood regulation, while advancing these ketamine analogues toward clinical trials. The prospect of efficacious, fast-acting antidepressants with safer profiles could revolutionize care for millions suffering from treatment-resistant depression worldwide.

In conclusion, the identification of deschloroketamine and its derivatives as potent modulators of adenosine dynamics heralds a new paradigm in antidepressant pharmacotherapy. By integrating phenotypic screening and mechanistic insights, this work paves the way for innovative treatments rooted in a deeper understanding of brain metabolism and purinergic neurotransmission, marking a milestone in the quest to alleviate the global burden of depression.

Subject of Research: Development of ketamine-derived compounds enhancing antidepressant effects via adenosine signaling in the medial prefrontal cortex.

Article Title: Adenosine signalling drives antidepressant actions of ketamine and ECT.

Article References:
Yue, C., Wang, N., Zhai, H. et al. Adenosine signalling drives antidepressant actions of ketamine and ECT. Nature (2025). https://doi.org/10.1038/s41586-025-09755-9

DOI: https://doi.org/10.1038/s41586-025-09755-9