An AI System for Autonomous Algorithm Evolution in Drug Development

Zhou, Z.; Nan, Y.; Mou, M.; Qian, Y.; Liu, Y.; Zuo, Z.; Yang, H.; Xu, W.; Li, B.; Jiang, W.; Ren, Y.; Liao, Y.; Wang, Y.; Li, Y.; Yang, Q.; Xi, Z.; Mi, T.; Sun, H.; Liu, P.; Zhu, F.

doi:10.64898/2026.08.16.745117

bioRxiv · 10.64898/2026.08.16.745117

An AI System for Autonomous Algorithm Evolution in Drug Development

Zhou, Z.·Nan, Y.·Mou, M.·Qian, Y.·Liu, Y.·Zuo, Z.·Yang, H.·Xu, W.·Li, B.·Jiang, W.·Ren, Y.·Liao, Y.·Wang, Y.·Li, Y.·Yang, Q.·Xi, Z.·Mi, T.·Sun, H.·Liu, P.·Zhu, F.

View original source ↗Export metadata ↓

Abstract

Artificial intelligence (AI) is increasingly permeating the drug development pipeline. Numerous algorithms for accelerating this multi-stage and multi-task process have been constructed, which depends heavily on expert design and labor-intensive task-specific optimization. Given that AI-driven acceleration of drug development is recognized as a cumulative, often synergistic, effect across multiple stages, the autonomous evolution of existing algorithms across the entire pipeline is demanded to achieve a holistic advancement. Here, we present DrugEvolve, a multi-role large language model system for systematic and autonomous algorithm evolution in drug development. DrugEvolve realizes a closed-loop evolution process by incorporating Researcher, Engineer, and Analyst domains, and enables an iterative design, implementation, evaluation, and refinement of algorithm by leveraging scientific knowledge and accumulated evolutionary experience. Across eleven representative tasks spanning target identification, drug discovery, preclinical study, and clinical trial, DrugEvolve autonomously evolved the corresponding task-specific algorithms and achieved substantial performance enhancement on 120 benchmark test sets. Moreover, it showed robust generalizabilities across heterogeneous data modalities (ranging from biological sequence and graph to molecular topology and textual language), and realized gains in both predictive and generative tasks. Collectively, this AI system can serve not only as an algorithmic infrastructure for drug development, but also as a transferable paradigm for broader scientific domains.

Explore related subjects

pharmacology and toxicology Preprint

Keep this discovery

Explore connections, maps & timelines

BibTeX RIS

Zhou, Z., Nan, Y., Mou, M., Qian, Y., Liu, Y., Zuo, Z., Yang, H., Xu, W., Li, B., Jiang, W., Ren, Y., Liao, Y., Wang, Y., Li, Y., Yang, Q., Xi, Z., Mi, T., Sun, H., Liu, P., Zhu, F.. 2026-08-20. An AI System for Autonomous Algorithm Evolution in Drug Development. https://doi.org/10.64898/2026.08.16.745117

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related preprints

bioRxiv2026

A metastable GαS/β-arrestin complex sustains endosomal PTH1R signaling

G protein-coupled receptors (GPCRs) initiate cellular signaling by activating heterotrimeric G proteins at the plasma membrane and are silenced when beta-arrestin binds the phosphorylated receptor, preventing G protein binding. However, many receptors continue to generate cAMP after internalization, even while barr is bound, a paradox that lacked a molecular mechanism in live cells. A direct barrestin1-Gas interface in the PTH1R-barr1-Gas complex has recently been defined structurally, but whether the complex operates in endosomes, and how it behaves there, has not been established. Using intact cells, we examine the assembly at three spatial resolutions. Confocal imaging with a nanobody selective for nucleotide-free Gas places activated Gas, barr, and agonist-occupied PTH1R on the same endosomes. Single-molecule localization microscopy, with illumination modes that separate the plasma membrane from the endosomal membrane in single cells, shows that the agonist recruits Gas and barr at both membranes. FRET, however, reveals a specific interaction only on endosomes, where a Gas inhibitor restricted to that compartment abolishes the sustained cAMP phase without affecting receptor internalization. Weighted-ensemble-enhanced molecular simulations show that the complex is metastable, releasing Gas through a large rigid-body rotation away from barr while barr itself tightens its interactions with the receptor. These findings provide a molecular basis for endosomal Gs/cAMP signaling and open a new avenue for drug development targeting this pathway.

pharmacology and toxicology↗

bioRxiv2026

Advancing AOPs for genotoxicity through quantifying KERs linking DNA alkylation to increase in mutations and chromosome aberrations

Quantitative Adverse Outcome Pathways (qAOPs) help translate mechanistic knowledge into quantitative toxicity predictions; however, their development is constrained by the limited availability and heterogeneity of quantitative data. Here, we investigated the feasibility of developing a qAOP for genotoxicity by modeling quantitative relationships linking DNA alkylation to increased mutations and structural chromosome aberrations through intermediate biological key events (KEs). We compiled and harmonized published concentration-response data from human lymphoblastoid TK6 cells exposed to ethyl methanesulfonate, methyl methanesulfonate, or N-ethyl-N-nitrosourea, covering eight genotoxicity assays representing five KEs with well-established dependencies. We evaluated sources of variability, compared points of departure (PoDs), and quantified relationships between successive events using response-response modeling. Substantial between-study variability could to a limited extent be explained through differences in experimental design while batch effects for individual studies and researchers were substantial. PoD estimates that were calculated from combined results of independent studies with similar experimental design, were often uncertain due to insufficient concentration-response characterization or conflicting results between studies. Nevertheless, response-response modeling suggested a threshold effect for certain key-event relationships (KERs). Specifically, DNA adducts seemed to be triggered at lower concentrations of alkylating agents compared to the subsequent KEs of DNA strand breaks and chromosome aberrations. Furthermore, these fitted response-response models can be used for prediction of downstream KEs from earlier ones. The general approach we developed for BMC analysis followed by response-response modeling is applicable to other AOPs as well. These findings highlight both the feasibility and limitations of using existing data for qAOP development.

pharmacology and toxicology↗

bioRxiv2026

Deoxycholic Acid (DCA) a Secondary Bile Acid, Mitigates Ovariectomy-Induced Bone Loss and promotes Bone Healing: A Novel Candidate for Orthopaedic Applications

Introduction Postmenopausal osteoporosis (PMO) is a metabolic disorder, characterized by accelerated bone loss, gut microbiota (GM) dysbiosis and altered microbial metabolites. Deoxycholic acid (DCA) is a secondary bile acid with an FDA-approved drug status for submental fat treatment in adults; however its effect on bone health remains underexplored. Materials and Methods This study investigated the effects of DCA on bone health using both preclinical animal model of PMO and clinical samples from PMO women. DCA (20 mg/kg) was administered to ovariectomised (OVX) mice for 45 days, and its effects on bone microarchitecture and serum cytokines were analysed. Osteoclastogenesis and osteoblastogenesis were evaluated using murine and human samples. Untargeted and targeted metabolomics analyses were performed using an ultra-high-performance liquid chromatographic system coupled with high-resolution Orbitrap mass spectrometry (HRMS). Secondary analysis of 16S rRNA sequencing data was performed using R programming. Results GM analysis and plasma metabolomics revealed an alteration in bacterial taxa associated with secondary BA synthesis and reduced circulating DCA levels. DCA administration improved impaired trabecular and cortical bone microarchitecture in the OVX mice by inhibiting osteoclastogenesis and enhancing osteoblastogenesis. Moreover, DCA altered circulating osteoclastogenic (IL-6, Il-9, IL-17A, IFN-{gamma}, RANKL) and osteoblastogenic (OPG, TGF-{beta}, IL-10) factors. Mechanically, DCA suppressed osteoclast differentiation and the expression of genes associated with osteoclast formation (NFATC1, NFKB, TRAF6, TREM2, SYNB), fusion (CSF1R, DC-STAMP, CD47), and resorption (CTSK, ATP6VOD1, ATP6VOD2, CAR2, SPP1) in human peripheral blood mononuclear cells (PBMCs), while promoting osteogenic differentiation and matrix mineralization in bone marrow-derived mesenchymal stem cells (BMSCs) from PMO women. Additionally, DCA enhanced the nascent calcium deposition at the fracture site and improved bone healing via inhibiting osteoclastogenesis. Conclusion Overall, reduced DCA availability was associated with PMO, while DCA treatment mitigated OVX-induced bone loss and improved bone healing. This study highlights the therapeutic potential of DCA for maintaining both bone homeostasis in PMO and fracture healing with immense orthopaedic applications.

pharmacology and toxicology↗