Chromatographic method development has always occupied a large and persistent share of analytical workload, and the scale of the problem is growing. A typical reversed-phase LC method development campaign involves systematic exploration across stationary phase chemistry, mobile phase composition, gradient profile, column temperature, flow rate, and pH—a multidimensional parameter space that no single experiment can adequately characterize. Column screening alone remains a standard, resource-intensive step in most workflows,¹ and in pharmaceutical chemistry, manufacturing, and controls (CMC) environments, the same project may require multiple methods across different laboratories and instruments, with frequent method bridging as development progresses.² The burden intensifies further for complex biotherapeutics: monoclonal antibodies (mAbs), antibody-drug conjugates (ADCs), multispecific antibodies (MsAbs), and cell and gene therapy (CGT) products present heterogeneous, high-molecular-weight structures with dozens of critical quality attributes (CQAs) that no single technique can fully characterize.^3,4 The multi-attribute method (MAM)—an LC–MS-based workflow for simultaneous quantification of multiple product quality attributes—exemplifies the resulting tension: it offers substantial information richness and sensitivity advantages over conventional ultraviolet (UV)-based assays,⁵ but generates high-dimensional datasets that demand expert interpretation and whose complex sample preparation workflow has limited adoption as a routine quality control (QC) release assay.⁶ These converging pressures—multidimensional parameter spaces, growing data volumes, and increasing molecular complexity—define the conditions under which autonomous and AI-assisted chromatographic systems are now emerging. Early machine learning (ML) approaches to retention time prediction demonstrated the potential for data-driven models to reduce the experimental burden of method development,⁷ and the sections that follow examine how those foundations have grown into a broader transformation of analytical practice.

Robotics, AI, and Closed-Loop Method Development

The most tractable early entry point for automation has been sample preparation. A fully automated multi-attribute method (MAM) peptide mapping procedure using a high-throughput robotic liquid handling system has been shown to process up to 96 samples while achieving reproducibility comparable to the manual protocol across sequence coverage, digestion completeness, and post-translational modification (PTM) quantification—using only commercially available components, making adoption straightforward.⁸ Robotic sample preparation removes a major source of inter-laboratory variability, but it does not change the logic of method development. That shift has come from the application of AI to experimental design itself. The AutoLC framework, developed at the University of Amsterdam, operates as a genuinely closed loop: an open-source algorithm interfaces directly with the LC instrument, performs scouting runs, builds a mechanistic retention model, simulates chromatograms computationally, and selects the next experiment to run without analyst intervention.⁹ Extended to comprehensive two-dimensional LC×LC–MS separations and applied to a complex monoclonal antibody (mAb) digest, the algorithm iteratively improved the separation within a small number of experimental cycles,¹⁰ demonstrating that computer-driven optimization is tractable even in the high-dimensional parameter spaces of biopharmaceutical analysis. Underpinning many of the most effective closed-loop systems is Bayesian optimization—a class of adaptive algorithms that uses a Gaussian process surrogate model to identify the most informative next experiment to run. A multi-objective Bayesian approach has been shown to optimize resolution, peak count, and method duration simultaneously, with acceptance criteria adjustable without restarting the search.¹¹ In a fully operator-free platform, optimal high-pressure liquid chromatography (HPLC) conditions were reached in as few as 13 experiments—a fraction of what conventional design of experiment (DoE) or linear solvent strength approaches require.¹² In manufacturing contexts, online LC integration allows this logic to extend into real-time process control: sampling cycle times of 1.30–2.35 min have been demonstrated across multiple biotherapeutic modalities, functioning as a genuine process analytical technology (PAT) tool,¹³ while explainable deep learning models have been shown to monitor a Protein A capture step reliably under realistic sensor fouling conditions—with interpretable outputs that satisfy process understanding requirements.¹⁴ Bayesian optimization has also been applied to the closed-loop refinement of adeno-associated virus (AAV) affinity chromatography parameters, improving yield and purity for clinically relevant serotypes AAV2, AAV5, and AAV9 within a quality-by-design (QbD)-aligned design space.¹⁵ At the level of experimental architecture, an LLM-powered platform has been described that integrates literature retrieval, stationary phase screening logic, DoE design, and parameter optimization within a single conversational workflow¹⁶ — a qualitatively different mode of AI involvement, reasoning about which experimental structure to adopt rather than optimizing within one already specified.

Digital Twins and Self-Optimizing Chromatography

Digital twins—dynamic computational models that mirror a physical system, update in real time, and can predict performance under untested conditions—are beginning to reshape how chromatographic process development is structured. A white-box hybrid digital twin for mAb purification combining high-throughput screening of 29 resins with mechanistic modeling has been shown to reduce the experimental burden of early development while building the process understanding required for a QbD approach.¹⁷ In continuous manufacturing, where pooling decisions must be made on timescales that preclude off-line analysis, a digital twin of a continuous downstream train integrated with an online high performance liquid chromatography process an analytical technology (HPLC–PAT) tool has been demonstrated; with model states updated in real time from charge variant data, model-predicted purity of 99.1% was achieved versus experimentally observed 99.3%.¹⁸ For analytical method development, quantitative structure-retention relationship (QSRR) frameworks trained on molecular descriptors now enable retention time prediction across chromatographic conditions before a single experiment is run. Structure-based random forest models have been shown to predict retention times with a median error below 4% and can seed in-silico condition selection without additional screening,¹⁹ while a generalizable framework incorporating stationary phase physicochemical descriptors enables retention time transfer across 20 commonly used reversed-phase columns without pre-existing data on the target column—directly addressing the method bridging burden in pharmaceutical CMC settings.² These foundations have been extended further with a physically grounded in-silico HPLC development framework combining quantitative structure-property relationship (QSPR), linear solvation energy relationships, and linear solvent strength theory, requiring no experimental data to generate predictions.²⁰ Beyond prediction, ML is now enabling automated troubleshooting and predictive maintenance: hybrid partial least squares-artificial neural network (PLS-ANN) models have been shown to detect subtle changes in column packing quality before they propagate into product quality deviations,²¹ and AI/ML-enabled continued process verification (CPV) of cation exchange chromatography has been demonstrated using soft sensors and deep neural networks to maintain consistent charge variant composition across manufacturing runs.²² A multi-company consortium has published consensus recommendations for AI application in CPV covering algorithm selection, data quality requirements, and model transparency, signaling that autonomous chromatographic capabilities are moving from research demonstration into the structured quality systems of commercial manufacturing.²³ The logical endpoint of these developments is a continuously learning system that improves its own performance with each analytical run—a transition already visible in commercial ML models for peak deconvolution that update as new data is incorporated,²⁴ and at the research frontier in reinforcement learning-based process control where the boundary between method development and autonomous manufacturing operation begins to dissolve.²⁵

The Evolving Role of the Analytical Scientist

The capabilities described above do not automate scientific judgment—they automate the execution of tasks that scientific judgment has already specified. Bayesian algorithms do not decide what a good separation looks like; they search for conditions that satisfy a definition the scientist has provided. Robotic systems do not determine which quality attributes to monitor; they execute protocols that reflect upstream analytical strategy decisions. Digital twins do not determine what process knowledge is worth building; they simulate outcomes within a design space defined by engineers and scientists. It has been articulated directly that next-generation analytical platforms will require reduced offline testing, renewed emphasis on online monitoring, multi-attribute characterization, and extensive use of advanced data analytics—not a replacement of analytical expertise but a redefinition of where it is applied.²⁶ It has also been demonstrated that the analytical scientist's role in an AI-augmented laboratory is in part to interrogate model outputs: to verify that dependencies align with first-principles expectations and to assess reliability under conditions that may fall outside the training distribution.¹⁴ This requires deep domain knowledge—of chromatographic theory, molecular behavior, and process context—not the ability to perform the physical operations the system has taken over. Successful AI/ML implementation in biopharmaceutical manufacturing requires interdisciplinary teams combining domain expertise with data science and software engineering capabilities; the absence of this integration is itself a source of implementation risk.²⁷ Autonomous laboratories across materials science and chemistry demonstrate that human domain expertise remains essential for directing the search toward scientifically meaningful objectives, with the most effective configurations maintaining human strategic input at key decision points.²⁸ AI literacy—the ability to critically evaluate and interpret AI-generated outputs without necessarily building the systems from scratch—is emerging as a professional imperative, and the educational gap between AI's expanding role in analytical chemistry and scientists' ability to engage with it critically represents a significant workforce challenge that training programs have not yet fully addressed.²⁹ Data quality compounds these challenges: models are only as reliable as the data they learn from, and chromatographic datasets accumulated under varying conditions present heterogeneity that can undermine generalizability.³⁰ The laboratory of the future—where AI-native platforms make machine learning accessible to scientists who are not data specialists, supported by a workforce with genuine digital literacy³¹—is one in which the premium on scientific judgment has not diminished but risen. The chromatographic systems examined in this column are becoming more autonomous. The scientists who design, deploy, and interpret them are becoming more consequential. That combination is what the autonomous analytical laboratory actually looks like.

Conclusion

Chromatography has always been shaped by the interplay between instrumental capability and scientific understanding. The emergence of autonomous chromatographic systems does not break this pattern; it extends it. Robotic sample preparation, AI-guided experimental design, closed-loop Bayesian optimization, digital twins, and continuously learning LC–MS platforms are not replacing the analytical scientist's contribution but relocating it: from the bench to the objective function, from the instrument to the model, from the chromatogram to the interpretation of what the algorithm cannot yet see. The challenges that remain—data quality, interoperability, algorithm transparency, and the readiness of regulated environments to accommodate adaptive analytical systems—are real, but they are challenges of implementation and governance, not of fundamental feasibility. The scientific and technological foundations are in place. What is now required is a generation of analytical scientists as fluent in the language of machine learning and experimental design strategy as in the language of retention mechanisms and method validation—scientists who understand that supervising an autonomous workflow is itself a form of scientific practice, one that demands judgment, mechanistic literacy, and critical engagement with outputs that no algorithm can fully evaluate on its own. The autonomous analytical laboratory is not a future state to be anticipated; it is a present condition being assembled, laboratory by laboratory, workflow by workflow, one closed loop at a time.

References

(1) Snyder, L. R.; Dolan, J. W. Current Challenges and Future Prospects in Chromatographic Method Development for Pharmaceutical Research. TrAC Trends Anal. Chem. 2017, 95, 28–46. DOI: 10.1016/j.trac.2017.07.021

(2) Shi, Z. Q.; Yi, Y. Y.; Madrigal, E. et al.. A Generalizable Methodology for Predicting Retention Time of Small Molecule Pharmaceutical Compounds Across Reversed-Phase HPLC Columns. J. Chromatogr. A 2025, 1742, 465628. DOI: 10.1016/j.chroma.2024.465628

(3) Heinrich, K.; Hoelterhoff, S.; Oezipek, S. et al. Establishing Multi-Dimensional LC-MS Systems for Versatile Workflows to Analyze Therapeutic Antibodies at Different Molecular Levels in Routine Operations. Pharmaceuticals 2025, 18 (3), 401. DOI: 10.3390/ph18030401

(4) Di Ianni, A.; Cowan, K. J.; Riccardi Sirtori, F. et al. Unlocking the Complexity of Antibody-Drug Conjugates: A Cutting-Edge LC-HRMS Approach to Refine Drug-to-Antibody Ratio Measurements with Highly Reactive Payloads. Int. J. Mol. Sci. 2025, 26 (7), 3080. DOI: 10.3390/ijms26073080

(5) Millán-Martín, S.; Jakes, C.; Carillo, S. et al. Multi-Attribute Method (MAM): An Emerging Analytical Workflow for Biopharmaceutical Characterization, Batch Release and cGMP Purity Testing at the Peptide and Intact Protein Level. Crit. Rev. Anal. Chem. 2024, 54 (8), 3234–3251. DOI: 10.1080/10408347.2023.2238058

(6) Jiang, T.; Kwofie, F.; Attanasio, N. et al. Exploring the Correlation between LC-MS Multi-Attribute Method and Conventional Chromatographic Product Quality Assays through Multivariate Data Analysis. AAPS J. 2024, 27 (1), 5. DOI: 10.1208/s12248-024-00973-z

(7) Fedorova, E. S.; Matyushin, D. D.; Plyushchenko, I. V. et al. Deep Learning for Retention Time Prediction in Reversed-Phase Liquid Chromatography. J. Chromatogr. A 2022, 1664, 462792. DOI: 10.1016/j.chroma.2021.462792

(8) Yang, F.; Zhang, J.; Buettner, A. et al. Fully Automated Peptide Mapping Multi-Attribute Method by Liquid Chromatography–Mass Spectrometry with Robotic Liquid Handling System. mAbs 2023, 15 (1), 2197668. DOI: 10.1016/j.jpba.2021.113988

(9) Bos, T. S.; Boelrijk, J.; Molenaar, S. R. A. et al. Chemometric Strategies for Fully Automated Interpretive Method Development in Liquid Chromatography. Anal. Chem. 2022, 94 (46), 16060–16068. DOI: 10.1021/acs.analchem.2c03160

(10) Molenaar, S. R. A.; Bos, T. S.; Boelrijk, J. et al. Computer-Driven Optimization of Complex Gradients in Comprehensive Two-Dimensional Liquid Chromatography. J. Chromatogr. A 2023, 1707, 464306. DOI: 10.1016/j.chroma.2023.464306

(11) Boelrijk, J.; Ensing, B.; Forré, P. et al. Closed-Loop Automatic Gradient Design for Liquid Chromatography Using Bayesian Optimization. Anal. Chim. Acta 2023, 1242, 340789. DOI: 10.1016/j.aca.2023.340789

(12) Dixon, T.; et al. Operator-Free HPLC Automated Method Development Guided by Bayesian Optimization. Digital Discovery 2024, 3, 1591–1601. DOI: 10.1039/d4dd00062e

(13) Graf, T.; Naumann, L.; Bonnington, L. et al. Expediting Online Liquid Chromatography for Real-Time Monitoring of Product Attributes to Advance Process Analytical Technology in Downstream Processing of Biopharmaceuticals. J. Chromatogr. A 2024, 1729, 465013. DOI: 10.1016/j.chroma.2024.465013

(14) Medl, M.; Leisch, F.; Dürauer, A. et al. Explainable Deep Learning Enhances Robust and Reliable Real-Time Monitoring of a Chromatographic Protein A Capture Step. Biotechnol. J. 2024, 19 (2), e2300554. DOI: 10.1002/biot.202300554

(15) Idanwekhai, K. P.; Shastry, S.; Minzoni, A. et al. Adaptive Machine Learning Framework Enables Unprecedented Yield and Purity of Adeno-Associated Viral Vectors for Gene Therapy. bioRxiv 2025.DOI: 10.1101/2025.05.23.655859

(16) Wan, X.; et al. A Chromatographic Process Design and Optimization Platform Powered by Large Language Models. arXiv 2025, arXiv:2601.03702. DOI: 10.48550/arXiv.2601.03702

(17) Silva, T. C.; Eppink, M.; Ottens, M. Digital Twin in High Throughput Chromatographic Process Development for Monoclonal Antibodies. J. Chromatogr. A 2024, 1717, 464672. DOI: DOI: 10.1016/j.chroma.2024.464672

(18) Tiwari, A.; Masampally, V. S.; Agarwal, A. et al. Digital Twin of a Continuous Chromatography Process for mAb Purification: Design and Model-Based Control. Biotechnol. Bioeng. 2023, 120 (3), 748–766. DOI: 10.1002/bit.28307

(19) Fine, J.; Peterson Mann, A. K.; Aggarwal, P. Structure Based Machine Learning Prediction of Retention Times for LC Method Development of Pharmaceuticals. Pharm. Res. 2024, 41 (2), 365–374. DOI: 10.1007/s11095-023-03646-2

(20) Marchetto, A.; Tirapelle, M.; Mazzei, L. et al. In Silico High-Performance Liquid Chromatography Method Development via Machine Learning. Anal. Chem. 2025.DOI: 10.1021/acs.analchem.4c03466

(21) Das, M.; et al. Application of Advanced Machine Learning Algorithms for Anomaly Detection and Quantitative Prediction in Protein A Chromatography. J. Chromatogr. A 2023, 1689, 463741. DOI: 10.1016/j.chroma.2022.463486

(22) Rathore, A. S.; et al. Implementation of Machine Learning Tool for Continued Process Verification of Process Chromatography Unit Operation. J. Chromatogr. A 2025, 1742, 465642. DOI: 10.1016/j.chroma.2024.465642

(23) Stassen, M.; Leitão, C. S.; Manzano, T. et al. Recommendations for Artificial Intelligence Application in Continued Process Verification: A Journey Toward the Challenges and Benefits of AI in the Biopharmaceutical Industry. PDA J. Pharm. Sci. Technol. 2025, 79 (1), 68–87. DOI: 10.5731/pdajpst.2024.012950

(24) Abramowitz, D. Artificial Intelligence in Chromatography: Advancing Method Development and Data Interpretation. LCGC Int. 2026. https://www.chromatographyonline.com/view/artificial-intelligence-in-chromatography-advancing-method-development-and-data-interpretation

(25) Rathore, A. S.; Tiwari, A.; Pandey, A. K. Reinforcement Learning Based Optimization of Process Chromatography for Continuous Processing of Biopharmaceuticals. Comput. Chem. Eng. 2021, 152, 107399. DOI: 10.1016/j.ces.2020.116171

(26) Rathore, A. S.; Sarin, D. What Should Next-Generation Analytical Platforms for Biopharmaceutical Production Look Like? Trends Biotechnol. 2024, 42 (3), 282–292. DOI: 10.1016/j.tibtech.2023.08.008

(27) Rathore, A. S.; Nikita, S.; Thakur, G, et al. Artificial Intelligence and Machine Learning Applications in Biopharmaceutical Manufacturing. Trends Biotechnol. 2023, 41 (4), 497–510. DOI: 10.1016/j.tibtech.2022.08.007

(28) Bayley, O.; Savino, E.; Slattery, A. et al. Autonomous Chemistry: Navigating Self-Driving Labs in Chemical and Material Sciences. Matter 2024, 7, 2382–2398. DOI: 10.1016/j.matt.2024.06.003

(29) Schönbein, A.; et al. Artificial Intelligence in Chemistry Research—Implications for Teaching and Learning. J. Chem. Educ. 2025. DOI: 10.1021/acs.jchemed.4c01033

(30) Auclair, J.; Rathore, A. S. Biopharmaceutical Characterization in the Age of Artificial Intelligence. LCGC Int. 2025. https://www.chromatographyonline.com/view/biopharmaceutical-characterization-in-the-age-of-artificial-intelligence

(31) Lab Manager Editorial. Laboratory Automation and AI in the Modern Lab Era. Lab Manager 2025. https://www.labmanager.com/laboratory-automation-and-ai-in-the-modern-lab-era-34704