FiberPro 1.0: Multiagent AI-Guided Design for High-Throughput Production and Conformal Deposition of Functional Protein Micro/Nanofibers

FiberPro organizes protein spinnability into testable hypotheses through a human-in-the-loop loop of one Central Agent plus four specialized agents for design, experimentation, analysis, and data management, updating its materials-design hypothesis after experimental failure. Unlike prior data-driven approaches that target a single property, require sizable curated datasets, and operate as open-loop recommenders, this framework integrates literature, characterization images, and newly generated experimental results into a traceable iterative decision path. Representative dialogues, formulation recommendations, and decision pathways are reported for SPI, keratin, and BSA (Table S1), with convergence reported at an average of about two iterations; this is case-based validation rather than a large-scale statistical comparison.

FRJS serves as the manufacturing platform with rapid feedback and high throughput: nascent fibers appear within about 1 s and a continuous network within 10 s, with a fiber production rate of approximately 30 g h⁻¹ at 20 wt% protein, nearly two orders of magnitude above single-needle electrospinning at the same concentration, scaling from 5 mL to 50 mL over 20 min and 500 mL over 3 h. It combines rapid spinnability assessment, scalable yield, and conformal deposition (random mats, aligned fibers, tubes, coatings on human appendages) in one platform, reducing subsequent cutting, transfer, and rolling steps. Based on time-lapse imaging, mats from increasing solution volumes, a production-rate comparison, and SEM of collagen, silk fibroin, and gelatin fibers; the rate comparison is for a single orifice.

FiberPro converted an initially unspinnable zein formulation into continuous Zein-PC fibers and provided mechanistic evidence for PEO as a bridging component. Neat zein formed only particles and short fiber fragments; adding 5–15 wt% PEO enabled continuous fibers, and reducing PEO content while increasing rotational speed suppressed fiber fusion. Rheology showed storage moduli of approximately 20–30 Pa for PEO-containing formulations, and all-atom MD showed the zein-chain diffusion coefficient decreasing from about 0.014 to 0.004 nm² ns⁻¹ as the ethanol/water ratio fell from 70:30 to 10:30, while zein–PEO hydrogen bonds per PEO group rose from about 0.004–0.007 to about 0.014. Combines SEM morphology, rheological measurements, and all-atom MD with three replicates per system (50 ns production trajectories, analysis over the final 30 ns), with mechanistic evidence consistent in direction with the experiments.

Zein-PC coatings demonstrated integrated antibacterial, preservation, colorimetric, and intelligent-prediction functions at real food interfaces. On aluminum foil the coating reduced E. coli from about 5.9 to 1.7 log CFU mL⁻¹ and achieved at least about 2.7 log reductions on all tested food interfaces; water-vapor transmission was about 14 g mm m⁻² d⁻¹ kPa⁻¹; the colorimetric response was about 4–6% between pH 5 and 7 and rose to about 39–45% between pH 9 and 11; a multitask ResNet-18 model trained on 42,936 images reached 99.49% classification accuracy on 7,226 held-out test images, with storage-time regression R² = 0.997 and mean absolute error of 1.40 h. Antibacterial testing used suspension and surface direct-contact assays with three independent replicates per condition; preservation covered 11 perishable foods over 10 days at ambient storage; the image model split data by individual specimen; prediction performance corresponds to controlled 21 °C and 5600 K illumination conditions.

The framework targets settings that require rapid screening of spinnable protein formulations and direct deposition of fibers onto complex, irregular surfaces, such as food-contact coatings and preservation; its conclusions rest on the selected protein systems and controlled application settings, and the authors state the study remains limited to a selected set of protein systems and controlled application settings. For readers, this means the method can serve as a starting point and decision-recording tool for exploring new protein systems rather than a universal rule predicting spinnability across all proteins.

The authors note that FiberPro's transferability across broader protein classes, systematic comparison with conventional optimization strategies, and robustness across laboratories require further evaluation; in addition, image-prediction performance corresponds to controlled 21 °C storage and 5600 K illumination, and behavior under other storage and imaging conditions remains to be seen. Readers may also note that this is a preprint that has not been certified by peer review, that data are available from the corresponding author upon reasonable request, and that the authors declare a related U.S. provisional patent application.