Protein purification analysis method based on optical detection and artificial intelligence system

Abstract

The invention provides a protein purification analysis method based on optical detection and an artificial intelligence system, which relates to the technical field of artificial intelligence, and comprises the following steps: injecting a protein sample to be analyzed into a chromatographic separation system for purification treatment; a four-channel parallel light path structure is adopted to collect multi-dimensional spectral data such as absorbance, fluorescence intensity, light scattering intensity and circular dichroism spectrum signals in real time, and crosstalk between channels is eliminated by using a self-adaptive signal compensation algorithm; inputting the compensated multi-dimensional spectral data into a pre-trained deep neural network model, extracting spectral features and time sequence features through a convolution layer and a circulation layer, and generating a purified feature vector; and on the basis of the vector, a layered reinforcement learning method is adopted to carry out real-time optimization adjustment on process parameters such as flow velocity, gradient, pH value and the like of the chromatographic separation system. According to the invention, accurate monitoring and intelligent control of the protein purification process can be realized, and the purification efficiency and the product quality are improved.

Description

Technical Field

[0001] This invention relates to artificial intelligence technology, and more particularly to a protein purification and analysis method based on optical detection and artificial intelligence systems. Background Technology

[0002] Protein purification is a crucial step in biopharmaceutical, bioengineering, and life science research, aiming to isolate and obtain high-purity target proteins from complex biological samples. Traditional protein purification methods primarily rely on chromatography, purifying proteins by setting fixed separation conditions. Chromatographic separation systems are typically equipped with single optical detection devices such as ultraviolet detectors or fluorescence detectors to monitor the protein during elution in real time. Operators determine the elution status based on the detector signals and adjust process parameters such as flow rate, gradient, and pH accordingly to achieve optimal purification results.

[0003] Traditional optical detection devices can only provide single-dimensional or limited-dimensional detection signals, failing to comprehensively characterize the multiple property changes of proteins during purification. For example, they cannot simultaneously acquire multi-dimensional information such as protein concentration, fluorescence properties, aggregation state, and secondary structure changes, leading to inaccurate and incomplete judgments of protein purification status. Existing methods primarily rely on operator experience for process parameter adjustments, lacking systematic and intelligent decision support. This makes it difficult to dynamically optimize the purification process based on real-time detection data, resulting in low purification efficiency and poor consistency and reproducibility of purification results. Especially when dealing with complex samples or novel proteins, extensive trial-and-error experiments are often required to determine suitable purification conditions. Traditional methods lack the ability to deeply mine and analyze spectral data, failing to extract implicit protein state characteristics from multi-dimensional spectral signals or establish correlation models between spectral features and optimal process parameters. This hinders intelligent control of the purification process, limiting the development of protein purification technology towards automation and intelligence. Summary of the Invention

[0004] This invention provides a protein purification and analysis method based on optical detection and artificial intelligence systems, which can solve the problems in the prior art.

[0005] A first aspect of the present invention provides a protein purification and analysis method based on optical detection and artificial intelligence systems, comprising: The protein sample to be analyzed is injected into the chromatographic separation system for purification. The absorbance, fluorescence intensity, light scattering intensity and circular dichroism signal of the protein sample to be analyzed are collected in real time by an optical detection device set at the outlet of the chromatographic separation system. The optical detection device adopts a four-channel parallel optical path structure, and multidimensional spectral data are collected synchronously through a tunable filter array. An adaptive signal compensation algorithm is used to eliminate crosstalk between channels to obtain compensated multidimensional spectral data. The compensated multidimensional spectral data is input into a pre-trained deep neural network model. The deep neural network model extracts spectral features through convolutional layers and captures temporal features through recurrent layers to generate a purification feature vector of the protein purification state. Based on the purification feature vector, a hierarchical reinforcement learning method is used to optimize the process parameters of the chromatographic separation system in real time. The adjustment direction of flow rate, gradient and pH value is determined according to the purification feature vector, and the specific adjustment amount of each parameter is accurately calculated based on the adjustment direction. The process parameters of the chromatographic separation system are adjusted according to the specific adjustment amount to complete the purification analysis of protein.

[0006] The absorbance, fluorescence intensity, light scattering intensity, and circular dichroism signal of the protein sample to be analyzed are acquired in real time by an optical detection device installed at the outlet of the chromatographic separation system, including: A fluid connection channel is provided between the outlet of the optical detection device and the outlet of the chromatographic separation system. The inner diameter and length of the fluid connection channel are designed to match the flow rate and diffusion coefficient of the protein sample to be analyzed, so as to ensure that the protein sample to be analyzed maintains its separation state in the chromatographic separation system when it flows from the outlet of the chromatographic separation system through the fluid connection channel to the detection area of ​​the optical detection device. During real-time acquisition, the optical detection device measures the absorbance, fluorescence intensity, light scattering intensity, and circular dichroism signal of the protein sample to be analyzed through the spectral detection module.

[0007] The optical detection device adopts a four-channel parallel optical path structure, synchronously acquires multidimensional spectral data through a tunable filter array, and uses an adaptive signal compensation algorithm to eliminate inter-channel crosstalk, obtaining compensated multidimensional spectral data including: The optical path structure of the optical detection device is configured. The optical detection device adopts a four-channel parallel optical path structure. Each channel is equipped with a tunable filter array. By adjusting the transmission band of the tunable filter array, selective acquisition of spectral signals of different wavelengths can be achieved. The optical detection device is activated to acquire spectral data, and multidimensional spectral data is acquired synchronously through the four-channel parallel optical path structure. The multidimensional spectral data is processed by signal processing, and an adaptive signal compensation algorithm is used to eliminate crosstalk between channels. The adaptive signal compensation algorithm calculates the crosstalk coefficient matrix between channels in real time and performs feedback correction on the signal of each channel according to the crosstalk coefficient matrix to obtain the compensated multidimensional spectral data.

[0008] The optical detection device is activated to acquire spectral data, and multidimensional spectral data is acquired synchronously through the four-channel parallel optical path structure; signal processing is performed on the multidimensional spectral data, and an adaptive signal compensation algorithm is used to eliminate inter-channel crosstalk, including: Multidimensional spectral data is acquired synchronously through a four-channel parallel optical path structure. The four-channel parallel optical path structure splits the incident light into four parallel beams through a beam splitter. Each beam passes through an independent filter and collimating lens and is received by a photodetector, thus realizing the synchronous acquisition of spectral information from the four channels. The multidimensional spectral data is processed by signal processing. An adaptive signal compensation algorithm is used to establish the crosstalk transfer function matrix between channels. The degree of crosstalk influence is calculated in real time based on the spectral information of the four channels. The compensation coefficient of each channel is determined by iterative optimization. The data of each channel is compensated accordingly. At the same time, the compensation coefficient is dynamically adjusted based on the compensation effect to achieve adaptive elimination of crosstalk. The multidimensional spectral data is corrected according to the compensation coefficient to obtain spectral data after eliminating inter-channel crosstalk.

[0009] The compensated multidimensional spectral data is input into a pre-trained deep neural network model. The deep neural network model extracts spectral features through convolutional layers and captures temporal features through recurrent layers to generate a purification feature vector of the protein purification state, including: The compensated multidimensional spectral data is input into a pre-trained deep neural network model. The deep neural network model extracts spectral features through convolutional layers. The convolutional layers adopt a multi-scale convolutional kernel structure to extract spectral features in different frequency ranges in a hierarchical manner. The spectral features are subjected to time-series analysis. The time-series features are captured by a recurrent layer. The recurrent layer adopts a long short-term memory network structure to establish the time-series correlation of the spectral features and generate a purification feature vector of the protein purification state.

[0010] Based on the purification feature vector, a hierarchical reinforcement learning method is used to optimize the process parameters of the chromatographic separation system in real time. The adjustment directions for flow rate, gradient, and pH are determined according to the purification feature vector, and the specific adjustment amounts for each parameter are accurately calculated based on these adjustment directions. Based on the purification feature vector, a hierarchical reinforcement learning method is used to optimize the process parameters of the chromatographic separation system in real time. The hierarchical reinforcement learning method includes an upper decision layer and a lower execution layer, wherein the upper decision layer determines the adjustment direction of flow rate, gradient and pH value according to the purification feature vector. Based on the adjustment direction, a parameter optimization objective function is constructed through the lower execution layer. The parameter optimization objective function comprehensively considers the stability and convergence speed of the process parameter adjustment, performs iterative calculations on each parameter, and accurately calculates the specific adjustment amount of each parameter.

[0011] A second aspect of the present invention provides an electronic device, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0012] A third aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0013] The beneficial effects of this application are as follows: This invention employs an optical detection device with a four-channel parallel optical path structure, enabling the simultaneous acquisition of multi-dimensional spectral data such as absorbance, fluorescence intensity, light scattering intensity, and circular dichroism signals, thus achieving real-time monitoring of protein samples from multiple angles. The combined use of a tunable filter array and an adaptive signal compensation algorithm effectively eliminates inter-channel crosstalk, significantly improving the accuracy and reliability of the detection data and providing high-quality foundational data support for subsequent data analysis.

[0014] This invention introduces a deep neural network model into the protein purification analysis process. Through the synergistic effect of convolutional and recurrent layers, it can effectively extract feature information from spectral data and accurately capture the temporal changes during purification, generating a feature vector that comprehensively reflects the protein purification status. This intelligent data processing method overcomes the limitations of traditional manual analysis, significantly improving the accuracy and efficiency of protein purification status identification. Attached Figure Description

[0015] Figure 1 This is a schematic flowchart of a protein purification and analysis method based on an optical detection and artificial intelligence system according to an embodiment of the present invention. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0018] Figure 1 This is a schematic flowchart of a protein purification and analysis method based on an optical detection and artificial intelligence system according to an embodiment of the present invention. Figure 1 As shown, the method includes: The protein sample to be analyzed is injected into the chromatographic separation system for purification. The absorbance, fluorescence intensity, light scattering intensity and circular dichroism signal of the protein sample to be analyzed are collected in real time by an optical detection device set at the outlet of the chromatographic separation system. The optical detection device adopts a four-channel parallel optical path structure, and multidimensional spectral data are collected synchronously through a tunable filter array. An adaptive signal compensation algorithm is used to eliminate crosstalk between channels to obtain compensated multidimensional spectral data. The compensated multidimensional spectral data is input into a pre-trained deep neural network model. The deep neural network model extracts spectral features through convolutional layers and captures temporal features through recurrent layers to generate a purification feature vector of the protein purification state. Based on the purification feature vector, a hierarchical reinforcement learning method is used to optimize the process parameters of the chromatographic separation system in real time. The adjustment direction of flow rate, gradient and pH value is determined according to the purification feature vector, and the specific adjustment amount of each parameter is accurately calculated based on the adjustment direction. The process parameters of the chromatographic separation system are adjusted according to the specific adjustment amount to complete the purification analysis of protein.

[0019] In one optional embodiment, the absorbance, fluorescence intensity, light scattering intensity, and circular dichroism signal of the protein sample to be analyzed are acquired in real time by means of an optical detection device disposed at the outlet of the chromatographic separation system, including: A fluid connection channel is provided between the outlet of the optical detection device and the outlet of the chromatographic separation system. The inner diameter and length of the fluid connection channel are designed to match the flow rate and diffusion coefficient of the protein sample to be analyzed, so as to ensure that the protein sample to be analyzed maintains its separation state in the chromatographic separation system when it flows from the outlet of the chromatographic separation system through the fluid connection channel to the detection area of ​​the optical detection device. During real-time acquisition, the optical detection device measures the absorbance, fluorescence intensity, light scattering intensity, and circular dichroism signal of the protein sample to be analyzed through the spectral detection module.

[0020] To ensure that the protein sample remains separated as it flows from the chromatographic separation system outlet through the fluid connection channel to the detection area of ​​the optical detection device, precise matching of the fluid connection channel's structural parameters is required. The fluid connection channel is a capillary structure made of polyetheretherketone (PEEK), with its inner wall polished to reduce surface roughness to below 0.2 micrometers, preventing non-specific adsorption of protein molecules onto the tube wall. When the flow rate of the protein sample is 0.5 mL / min, the inner diameter of the fluid connection channel is set to 0.18 mm, and the length is controlled within the range of 80 mm to 120 mm. For immunoglobulin samples with a diffusion coefficient of 5 × 10⁻⁷ cm² / s, an inner diameter of 0.15 mm and a length of 90 mm are selected for the fluid connection channel, which can control peak broadening within 2% of the theoretical plate number.

[0021] The connection between the fluid flow channel and the outlet of the chromatographic separation system uses a zero-dead-volume connector. The internal channel of the connector is designed with a tapered transition structure, with the cone angle controlled between 15 and 25 degrees to ensure that the Reynolds number of the liquid flow in the transition region remains laminar. The connection seal uses a double O-ring design. The inner ring is made of fluororubber to withstand organic solvents, while the outer ring is made of silicone rubber to provide elastic cushioning. The distance between the two rings is 1.5 mm. The tightening torque of the connector is controlled between 0.8 Nm and 1.2 Nm using a torque wrench to avoid over-tightening that could deform the fluid flow channel or damage the sealing ring.

[0022] The optical detection device comprises an ultraviolet-visible absorption detection unit, a fluorescence detection unit, a light scattering detection unit, and a circular dichroism detection unit. These detection units are arranged in a ring around a central flow cell, which is made of quartz with an optical path length of 10 mm and an internal volume controlled between 8 and 12 microliters. The inlet and outlet of the flow cell are positioned on the same horizontal plane, offset from each other by 120 degrees, causing the liquid flow to form a spiral trajectory within the cell, extending the residence time to 2.5 to 3.5 seconds and improving the stability of the detection signal.

[0023] The UV-Vis absorption detection unit is equipped with a deuterium lamp light source, with an emission wavelength range covering 190 nm to 800 nm. The light beam emitted by the light source is focused by a collimating lens and then incident on a grating beam splitter. The grating has a line density of 1200 lines per millimeter, and the blaze wavelength is set at 280 nm. The split monochromatic light illuminates the flow cell through a 2 nm wide slit, and the light beam passing through the sample is received by a silicon-based photodiode array. The photodiode array contains 512 pixels, each with an effective photosensitive area of ​​25 μm × 500 μm and a pixel pitch of 50 μm. The detection system monitors the absorbance at 280 nm in real time, with a sampling frequency set to 10 Hz. The analog signal is converted into a digital signal by a 16-bit analog-to-digital converter. When the protein concentration is 1 mg / mL, the absorbance value at 280 nm is typically in the range of 0.8 to 1.2, and the linear response range of the detection system covers 0.001 to 3.0 absorbance units.

[0024] The fluorescence detection unit uses a xenon lamp as the excitation source. The excitation light is filtered through a bandpass filter with a center wavelength of 450 nm and a full width at half maximum (FWHM) of 20 nm before being incident on the flow cell at a 45-degree angle. The fluorescence generated by the excitation of tryptophan and tyrosine residues in the sample is transmitted through a 520 nm long-pass filter perpendicular to the excitation light and then received by a photomultiplier tube (PMT). The PMT's operating voltage is set between 800 and 1000 volts, achieving a gain of 10 to the power of 6. The integration time for fluorescence detection is set to 100 milliseconds. Under these conditions, when detecting a bovine serum albumin solution with a concentration of 0.1 μg / mL, the fluorescence intensity signal output value is 2500 relative fluorescence units, and the signal-to-noise ratio (SNR) exceeds 200.

[0025] The light scattering detection unit is equipped with a 632.8 nm helium-neon laser with a stable output power of 5 mW. The laser beam, after being processed by a spatial filter and beam expander, forms a parallel beam with a diameter of 1 mm, which passes through the central region of the flow cell. The scattered light detector is positioned at a 90-degree angle to the incident light and uses an avalanche photodiode as the photosensitive element, with a response time of less than 10 nanoseconds. A 50 mm focal length collecting lens is installed at the front end of the detector, with an aperture angle controlled between 5 and 10 degrees to ensure that the collected scattered light comes from an effective scattering volume of approximately 200 μL at the center of the flow cell. For a monoclonal antibody with a molecular weight of 150 kDa, at a concentration of 2 mg / mL, the intensity of the scattered light at a 90-degree angle is approximately 0.003% of the incident light intensity. The detection system processes the signal using a correlator, calculating a hydrodynamic radius of 5.8 nm to 6.2 nm.

[0026] The circular dichroism spectroscopy detection unit uses a high-pressure xenon lamp in conjunction with a monochromator to generate linearly polarized light in the range of 190 nm to 260 nm. This linearly polarized light is converted into left-handed and right-handed circularly polarized light by a photoelastic modulator at a modulation frequency of 50 kHz. The two types of circularly polarized light alternately pass through a flow cell and are detected by a UV-enhanced photomultiplier tube. The detection system measures the absorption difference between the left-handed and right-handed circularly polarized light and converts it into an ellipticity parameter. For a protein sample with an α-helix content of 60%, the ellipticity value is -15,000 degrees square centimeters per decimal at a wavelength of 222 nm and -18,000 degrees square centimeters per decimal at a wavelength of 208 nm. The circular dichroism spectroscopy scanning speed is set to 100 nm per minute, the data point interval is 0.5 nm, and the accumulation time for each data point is 0.3 seconds.

[0027] Signals generated by each detection unit are synchronously transmitted to the data acquisition system, which employs a multi-channel synchronous sampling architecture with a time deviation of less than 1 millisecond between channels. Absorbance signals are converted using a logarithmic amplifier, fluorescence intensity and light scattering intensity signals are processed by a transimpedance amplifier, and circular dichroism signals are extracted using a lock-in amplifier. All analog signals are digitized using a 24-bit high-precision analog-to-digital converter with a sampling rate set to 100 Hz to meet the detection requirements where chromatographic peak widths are typically greater than 10 seconds. During data acquisition, the elution time, absorbance value, fluorescence intensity value, scattered light intensity value, and ellipticity value corresponding to each detection point are recorded in real time, forming a multi-dimensional spectral data matrix for subsequent analysis.

[0028] In one optional implementation, the optical detection device employs a four-channel parallel optical path structure, synchronously acquires multidimensional spectral data through a tunable filter array, and uses an adaptive signal compensation algorithm to eliminate inter-channel crosstalk, obtaining compensated multidimensional spectral data including: The optical path structure of the optical detection device is configured. The optical detection device adopts a four-channel parallel optical path structure. Each channel is equipped with a tunable filter array. By adjusting the transmission band of the tunable filter array, selective acquisition of spectral signals of different wavelengths can be achieved. The optical detection device is activated to acquire spectral data, and multidimensional spectral data is acquired synchronously through the four-channel parallel optical path structure. The multidimensional spectral data is processed by signal processing, and an adaptive signal compensation algorithm is used to eliminate crosstalk between channels. The adaptive signal compensation algorithm calculates the crosstalk coefficient matrix between channels in real time and performs feedback correction on the signal of each channel according to the crosstalk coefficient matrix to obtain the compensated multidimensional spectral data.

[0029] The optical path structure of the optical detection device includes four independent optical channels, each equipped with a light source module, a collimating lens group, a tunable filter array, a focusing lens group, and a photodetector. The light source module uses a broadband light-emitting diode with an emission wavelength range covering 400 nm to 1000 nm. The light source is driven by a regulated power supply with an operating current set at 120 mA to ensure output light intensity stability better than 0.5%. The collimating lens group consists of two aspherical lenses: the first lens has a focal length of 25 mm, and the second lens has a focal length of 50 mm. When combined, they form a parallel beam with a divergence angle of less than 0.1 degrees.

[0030] The tunable filter array is the core component of each channel, employing a Fabry-Perot interferometer structure. Continuous tunability of the transmission band is achieved by adjusting the resonant cavity length through the application of different voltages. The first channel's transmission band is adjustable from 450 nm to 550 nm, corresponding to a control voltage of 0 V to 10 V; the second channel's transmission band is from 550 nm to 650 nm, also with a control voltage of 0 V to 10 V; the third channel covers the 650 nm to 750 nm band; and the fourth channel covers the 750 nm to 850 nm band. The filter's transmission bandwidth is set to 5 nm, with a peak transmittance greater than 85% and an out-of-band rejection ratio better than -40 dB. Each filter is equipped with an independent voltage control circuit, achieving a voltage adjustment accuracy of 1 mV and a response time of less than 2 milliseconds.

[0031] The focusing lens group focuses the light signal passing through the filter onto the photosensitive surface of the photodetector. The lens employs a cemented doublet structure to eliminate chromatic aberration, with an effective focal length of 15 mm and a numerical aperture of 0.3. The photodetector uses a silicon-based photodiode with a photosensitive area of ​​5 mm x 5 mm, a response wavelength range covering 350 nm to 1100 nm, and a quantum efficiency peaking at 90% at 600 nm. The photocurrent output by the detector is converted into a voltage signal by a transimpedance amplifier, with the amplification gain adjustable from 1000 to 100000 times, and an output voltage range of 0 V to 5 V.

[0032] When the optical detection device is started, the control system simultaneously supplies power to the light source modules of all four channels, with a warm-up time set to 30 seconds to ensure stable light source output. The control system sends initialization commands to the filter control circuits of each channel via a serial communication interface, setting the first channel filter to transmit at a wavelength of 500 nm (corresponding to a control voltage of 5 volts); the second channel to transmit at a wavelength of 600 nm (corresponding to a control voltage of 5 volts); the third channel to transmit at a wavelength of 700 nm; and the fourth channel to transmit at a wavelength of 800 nm. After the filters complete wavelength setting, a 10-millisecond wait is allowed for the resonant cavity to stabilize.

[0033] During data acquisition, the output signals from the photodetectors of all four channels are simultaneously fed into a multi-channel analog-to-digital converter (ADC). The converter uses 16-bit resolution and a sampling rate of 100,000 times per second. Within a complete spectral scan cycle, the filter of each channel is adjusted according to a preset wavelength sequence, with a wavelength step interval of 2 nanometers. Taking the first channel as an example, the filter starts at 450 nanometers and is sequentially adjusted to 452 nanometers, 454 nanometers, up to 548 nanometers and 550 nanometers, for a total of 51 wavelength points. The dwell time at each wavelength point is 5 milliseconds, with the first 2 milliseconds used for filter stabilization and the last 3 milliseconds for signal acquisition. 300 data samples are collected at each wavelength point, and the average value is calculated as the spectral intensity value for that wavelength.

[0034] By operating four channels in parallel, the device completes a spectral scan across the entire 400 nm to 1000 nm wavelength band within 0.255 seconds, obtaining spectral data at 204 wavelength points. Actual measurements revealed optical crosstalk between adjacent channels due to the compact layout of the optical path. Specifically, when the first channel measures a 500 nm wavelength signal, its output value not only includes the actual 500 nm signal from that channel but also contains a portion of the optical signal from the second channel. Single-channel excitation tests under anechoic conditions showed that even when only the first channel light source is activated, the second channel detector can still detect an interference signal equivalent to 3.2% of the first channel signal strength.

[0035] An adaptive signal compensation algorithm is used to eliminate this inter-channel crosstalk. During algorithm execution, the control system sequentially excites each channel individually and records the response values ​​of all channels. When the first channel is excited alone and its nominal output value is 1000 millivolts, the detector on the first channel measures 982 millivolts, the second channel measures 32 millivolts, the third channel measures 15 millivolts, and the fourth channel measures 8 millivolts. When the second channel is excited alone and its output is 1000 millivolts, the first channel measures 28 millivolts, the second channel measures 975 millivolts, the third channel measures 35 millivolts, and the fourth channel measures 12 millivolts. This process continues until the third and fourth channels are tested.

[0036] Based on the test data above, a crosstalk coefficient matrix is ​​constructed. This matrix is ​​a 4x4 data table, where each element represents the coupling coefficient between corresponding channels. The diagonal elements of the matrix represent the self-coupling coefficient of each channel: 0.982 for the first channel, 0.975 for the second, 0.968 for the third, and 0.971 for the fourth. The off-diagonal elements represent the cross-coupling coefficients between channels. For example, the element in the first row and second column is 0.032, representing the crosstalk coefficient of the second channel to the first channel.

[0037] During signal compensation processing, the crosstalk coefficient matrix is ​​inverted to obtain the compensation coefficient matrix. The inversion process uses Gaussian-Jordan elimination, transforming the crosstalk coefficient matrix into an identity matrix through a series of row operations. The same row operations are then performed on the identity matrix to obtain the compensation coefficient matrix. The calculated value of the first element in the first row and first column of the compensation coefficient matrix is ​​1.025, and the element in the first row and second column is -0.034. This indicates that when compensating for the first channel signal, its measured value needs to be multiplied by 1.025 and subtracted by 0.034 times the measured value of the second channel.

[0038] In the actual compensation calculation, it is assumed that the measured values ​​of the four channels at a certain moment are 800 mV, 1200 mV, 950 mV, and 600 mV, respectively. The compensated signal of the first channel is equal to 800 mV multiplied by 1.025, minus 1200 mV multiplied by 0.034, minus 950 mV multiplied by 0.016, and minus 600 mV multiplied by 0.008, resulting in 768 mV. The compensated signals of the second, third, and fourth channels are calculated using the same method, yielding 1231 mV, 982 mV, and 618 mV, respectively. The compensated signal data eliminates the influence of inter-channel crosstalk and reflects the true spectral intensity received by each channel.

[0039] The adaptive function is reflected in the real-time update mechanism of the crosstalk coefficient matrix. After each power-on warm-up and every two hours of continuous operation, the device automatically executes a crosstalk coefficient calibration process. The calibration process control system records the ambient temperature; when the temperature change exceeds 5 degrees Celsius, an additional calibration process is triggered. By updating the crosstalk coefficient matrix in real time, the compensation algorithm can adapt to changes in crosstalk characteristics caused by environmental factors such as temperature and vibration in the optical path structure, ensuring that the compensation accuracy remains within 0.5% over a long period.

[0040] In one optional implementation, the optical detection device is activated to acquire spectral data, and multidimensional spectral data is acquired synchronously through the four-channel parallel optical path structure; signal processing is performed on the multidimensional spectral data, and an adaptive signal compensation algorithm is used to eliminate inter-channel crosstalk, including: Multidimensional spectral data is acquired synchronously through a four-channel parallel optical path structure. The four-channel parallel optical path structure splits the incident light into four parallel beams through a beam splitter. Each beam passes through an independent filter and collimating lens and is received by a photodetector, thus realizing the synchronous acquisition of spectral information from the four channels. The multidimensional spectral data is processed by signal processing. An adaptive signal compensation algorithm is used to establish the crosstalk transfer function matrix between channels. The degree of crosstalk influence is calculated in real time based on the spectral information of the four channels. The compensation coefficient of each channel is determined by iterative optimization. The data of each channel is compensated accordingly. At the same time, the compensation coefficient is dynamically adjusted based on the compensation effect to achieve adaptive elimination of crosstalk. The multidimensional spectral data is corrected according to the compensation coefficient to obtain spectral data after eliminating inter-channel crosstalk.

[0041] After the optical detection device is activated, the incident light beam enters a beam splitter with a four-channel parallel optical path structure. This beam splitter uses a multilayer dielectric film beam splitter group to divide the incident light into four parallel beams according to wavelength range and energy ratio. The first beam corresponds to the blue light band with a wavelength range of 400 nm to 500 nm, the second beam corresponds to the green light band with a wavelength range of 500 nm to 600 nm, the third beam corresponds to the red light band with a wavelength range of 600 nm to 700 nm, and the fourth beam corresponds to the near-infrared band with a wavelength range of 700 nm to 900 nm. The splitting ratio of the beam splitter is set according to the light intensity characteristics of each band, specifically allocating 25% of the light energy to the blue light band, 30% to the green light band, 25% to the red light band, and 20% to the near-infrared band.

[0042] Each parallel light beam passes sequentially through its own independent filter assembly. The first channel's filter has a center wavelength of 450 nm and a bandwidth of 100 nm, achieving a transmittance of over 95%. The second channel's filter has a center wavelength of 550 nm and a bandwidth of 100 nm, also achieving a transmittance of over 95%. The third channel's filter has a center wavelength of 650 nm and a bandwidth of 100 nm, and the fourth channel's filter has a center wavelength of 800 nm and a bandwidth of 200 nm. The filters employ a multilayer dielectric film coating process to achieve high transmittance within the target wavelength range and a suppression effect with a cutoff depth greater than 4 times the optical density value within the non-target wavelength range.

[0043] The filtered light beam enters the collimating lens system. Each channel is equipped with an independent aspherical lens collimating lens with a focal length of 50 mm and a numerical aperture of 0.2. The collimating lens converts the diverging beam into a parallel beam, with the collimation controlled within 0.5 milliradians. The collimated beam is then incident on the photosensitive surface of the photodetector, which uses a silicon-based photodiode array. Each detector contains 256 pixel units, with a pixel size of 25 μm x 25 μm and a response time of less than 10 microseconds. The photodetectors of the four channels synchronously acquire signals at a sampling frequency of 10 kHz. The analog-to-digital converter has a resolution of 16 bits, enabling simultaneous acquisition of spectral information from all four channels in a single acquisition.

[0044] In a specific measurement, the first channel collected 3250 counts of light intensity, the second channel collected 4180 counts, the third channel collected 3890 counts, and the fourth channel collected 2560 counts. This raw data is stored in a buffer memory, awaiting subsequent signal processing.

[0045] The signal processing module receives raw spectral data from four channels and establishes an inter-channel crosstalk transfer function matrix. This matrix is ​​a 4x4 square matrix, where diagonal elements represent the direct response coefficients of each channel, and off-diagonal elements represent the crosstalk coefficients between channels. Initial matrix parameters were obtained through a standard light source calibration experiment. The response coefficient of the first channel to its own band is 0.98, the crosstalk coefficient to the second channel band is 0.03, the crosstalk coefficient to the third channel band is 0.01, and the crosstalk coefficient to the fourth channel band is 0.005. The response coefficient of the second channel to its own band is 0.97, the crosstalk coefficient to the first channel band is 0.025, the crosstalk coefficient to the third channel band is 0.02, and the crosstalk coefficient to the fourth channel band is 0.008. The response coefficient of the third channel to its own band is 0.96, the crosstalk coefficient to the first channel band is 0.008, the crosstalk coefficient to the second channel band is 0.018, and the crosstalk coefficient to the fourth channel band is 0.025. The response coefficient of the fourth channel to its own band is 0.95, the crosstalk coefficient to the first channel band is 0.006, the crosstalk coefficient to the second channel band is 0.012, and the crosstalk coefficient to the third channel band is 0.022.

[0046] The adaptive signal compensation algorithm calculates the crosstalk impact of the four channels in real time. Taking the measured value of 3250 for the first channel as a reference, its crosstalk contribution to the second channel is calculated as 3250 multiplied by 0.03, resulting in 97.5 counts; its crosstalk contribution to the third channel is 3250 multiplied by 0.01, resulting in 32.5 counts; and its crosstalk contribution to the fourth channel is 3250 multiplied by 0.005, resulting in 16.25 counts. Similarly, the crosstalk contribution of the second channel to the other channels is calculated: its crosstalk contribution to the first channel is 4180 multiplied by 0.025, resulting in 104.5 counts; its crosstalk contribution to the third channel is 4180 multiplied by 0.02, resulting in 83.6 counts; and its crosstalk contribution to the fourth channel is 4180 multiplied by 0.008, resulting in 33.44 counts. The crosstalk contribution of the third channel to the first channel is 3890 multiplied by 0.008, resulting in 31.12 counts; to the second channel, it is 3890 multiplied by 0.018, resulting in 70.02 counts; and to the fourth channel, it is 3890 multiplied by 0.025, resulting in 97.25 counts. The crosstalk contribution of the fourth channel to the first channel is 2560 multiplied by 0.006, resulting in 15.36 counts; to the second channel, it is 2560 multiplied by 0.012, resulting in 30.72 counts; and to the third channel, it is 2560 multiplied by 0.022, resulting in 56.32 counts.

[0047] The iterative optimization process determines the compensation coefficients for each channel through multiple iterations. In the first iteration, the initial compensation value for the first channel is set to the original value of 3250 minus the total crosstalk from other channels. The total crosstalk is 104.5 + 31.12 + 15.36 = 150.98 counts, resulting in a first compensation result of 3099.02 counts for the first channel. The total crosstalk for the second channel is 97.5 + 70.02 + 30.72 = 198.24 counts, resulting in a compensated value of 3981.76 counts. The total crosstalk for the third channel is 32.5 + 83.6 + 56.32 = 172.42 counts, resulting in a compensated value of 3717.58 counts. The total crosstalk of the fourth channel is 16.25 plus 33.44 plus 97.25, which equals 146.94 counts. After compensation, the value is 2413.06 counts.

[0048] The second iteration recalculates the crosstalk effect using the values ​​from the first compensation. The new crosstalk contribution for the first channel is 3099.02 multiplied by the corresponding crosstalk coefficient, the new crosstalk contribution for the second channel is 3981.76 multiplied by the corresponding crosstalk coefficient, and so on. After five iterations, the change in the compensation value for each channel is less than one count value, indicating that the algorithm has converged. The final determined compensation coefficients are 1.048 for the first channel, 1.053 for the second channel, 1.047 for the third channel, and 1.061 for the fourth channel.

[0049] The compensation effect evaluation module monitors the signal-to-noise ratio changes and inter-channel correlation indicators of the data in real time. When changes in ambient temperature or fluctuations in light source intensity are detected, the compensation coefficient is dynamically adjusted. For every 5-degree Celsius increase in temperature, the compensation coefficient increases by 0.2%; for every 10% decrease in light source intensity, the compensation coefficient increases by 0.5%. By continuously monitoring the data stability for thirty sampling periods, the adaptive adjustment process ensures that the compensation algorithm is always in optimal working condition.

[0050] The final calibration process multiplies the original measured values ​​by the compensation coefficient to obtain the crosstalk-free spectral data. The calibration result for the first channel is 3250 multiplied by 1.048, yielding 3406 counts; for the second channel, it is 4180 multiplied by 1.053, yielding 4401 counts; for the third channel, it is 3890 multiplied by 1.047, yielding 4073 counts; and for the fourth channel, it is 2560 multiplied by 1.061, yielding 2716 counts. The calibrated data underwent cross-validation testing between channels, with each channel showing an independence coefficient exceeding 0.98 and a crosstalk suppression ratio better than 40 dB, meeting the technical requirements for high-precision spectral analysis.

[0051] In one optional implementation, the compensated multidimensional spectral data is input into a pre-trained deep neural network model. The deep neural network model extracts spectral features through convolutional layers and captures temporal features through recurrent layers to generate a purification feature vector representing the protein purification state, including: The compensated multidimensional spectral data is input into a pre-trained deep neural network model. The deep neural network model extracts spectral features through convolutional layers. The convolutional layers adopt a multi-scale convolutional kernel structure to extract spectral features in different frequency ranges in a hierarchical manner. The spectral features are subjected to time-series analysis. The time-series features are captured by a recurrent layer. The recurrent layer adopts a long short-term memory network structure to establish the time-series correlation of the spectral features and generate a purification feature vector of the protein purification state.

[0052] The compensated multidimensional spectral data is input into a deep neural network model for feature extraction and purified state recognition. This multidimensional spectral data has dimensions of 512×128×3, where 512 represents the number of spectral wavelength sampling points, 128 represents the time series length, and 3 indicates the inclusion of three channels: ultraviolet absorption spectrum, fluorescence spectrum, and scattering spectrum. The deep neural network model employs a hybrid architecture of convolutional neural networks and recurrent neural networks. The input interface receives spectral data in a three-dimensional tensor format, with the data format being floating-point and the numerical range normalized to between 0 and 1.

[0053] The convolutional layers employ a multi-scale convolutional kernel structure to extract spectral features hierarchically. The first convolutional layer uses three different sized kernels: a 3×1 kernel for capturing local peak features in the spectral data (64 kernels, stride 1, and uniform padding); a 5×1 kernel for extracting medium-range spectral patterns (also 64 kernels, stride 1); and a 7×1 kernel for identifying wide-range spectral trends (64 kernels). Each kernel performs convolution operations on the input multidimensional spectral data. The kernels are multiplied element-wise with the corresponding values ​​in the input data, summed, and then a bias term is added to obtain the feature mapping result.

[0054] Taking the ultraviolet absorption spectrum in the wavelength range of 280 nm to 400 nm as an example, this spectrum contains 128 sampling points. After processing with a 3×1 convolution kernel, 64 feature channels are extracted, with each feature channel retaining 128 values. The weight parameters of the convolution kernel are optimized during training using a backpropagation algorithm. The initial weights are initialized using a random normal distribution with a mean of 0 and a standard deviation of 0.02. The output data after the convolution operation undergoes batch normalization, making the mean of the numerical distribution of each feature channel close to 0 and the variance close to 1, thus improving the stability of model training.

[0055] The activation function is a modified linear unit function, which sets values ​​less than 0 in the feature map output of the convolutional layer to 0, while values ​​greater than 0 remain unchanged. The outputs of three different sized convolutional kernels are concatenated along the feature dimension to form a feature tensor of size 512×128×192. This feature tensor is input into a second convolutional layer for further feature extraction. The second convolutional layer has 128 3×3 kernels, performing feature fusion in both the spectral and temporal dimensions. The stride of the convolution operation is set to 2 in the spectral dimension and 1 in the temporal dimension. After processing by this layer, the feature tensor size becomes 256×128×128.

[0056] To avoid excessive computational complexity due to excessively large feature dimensions, a max pooling layer is added after the second convolutional layer. The pooling window size is 2×2, and the stride is 2×2. The pooling operation selects the maximum value within each 2×2 window as the output, which reduces the feature tensor size to 128×64×128. The pooled feature tensor retains the most significant feature information from the original spectral data while reducing the computational cost of subsequent processing.

[0057] The extracted spectral features need to undergo temporal analysis to capture the dynamic changes during protein purification. The feature tensors processed by convolutional layers are then globally averaged and pooled along the spectral dimension to obtain a 64×128 two-dimensional feature matrix, where 64 represents the time step and 128 represents the feature vector dimension. This feature matrix is ​​then input into the recurrent layers of a Long Short-Term Memory (LSTM) network structure in chronological order.

[0058] The Long Short-Term Memory (LSTM) network unit comprises four core components: an input gate, a forget gate, an output gate, and a cell state. At each time step, the input gate calculates a gating value between 0 and 1 based on the feature vector of the current time step and the hidden state of the previous time step. This gating value is generated using the sigmoid activation function. Specifically, the weight matrix of the input gate has a size of 128×128, and the bias vector has a size of 128. The feature vector of the current time step is multiplied by the weight matrix, plus the product of the hidden state of the previous time step and another weight matrix, plus the bias vector, and then input into the sigmoid function to obtain the activation value of the input gate.

[0059] The forget gate is calculated similarly to the input gate, generating a gating value between 0 and 1 to determine which information in the cell state needs to be forgotten. The weight matrix of the forget gate is also 128×128. Cell state updates are achieved by element-wise multiplying the cell state from the previous time step with the output of the forget gate, and then adding the element-wise product of the input gate output and the candidate cell state. Candidate cell states are generated using a hyperbolic tangent activation function, with values ​​ranging from -1 to +1.

[0060] The output gate controls the output of the hidden state at the current time step, and its calculation method is also based on the current input features and the hidden state at the previous time step. The activation value of the output gate is multiplied element-wise with the cell state processed by the hyperbolic tangent function to obtain the hidden state output at the current time step. This hidden state vector has a dimension of 128 and contains the accumulated feature information from the initial time step to the current time step.

[0061] The recurrent layer is configured as a two-layer structure, with the hidden state output of the first layer (Long Short-Term Memory network) serving as the input feature of the second layer. This two-layer structure enables the learning of more complex temporal dependencies and improves the ability to capture trends in protein purification states. After processing at all 64 time steps, the hidden state of the last time step is extracted as the representation vector of the temporal features, with a dimension of 128.

[0062] The spectral features extracted by the convolutional layer and the temporal features captured by the recurrent layer are fused to generate a purification feature vector representing the protein purification status. Feature fusion is implemented using a fully connected layer with an input dimension of 256, containing a concatenation of 128-dimensional spectral features and 128-dimensional temporal features. The weight matrix of the fully connected layer is 256×256. Each input feature is multiplied by the corresponding row of the weight matrix. All dot products are summed, a bias term is added, and then the result is processed through a modified linear unit activation function to obtain the 256-dimensional purification feature vector. This purification feature vector comprehensively reflects the spectral characteristics of the protein sample in different wavelength ranges and its temporal evolution during purification, providing a reliable feature representation for subsequent purification status determination.

[0063] In one optional implementation, a hierarchical reinforcement learning method is used to optimize the process parameters of the chromatographic separation system in real time based on the purification feature vector. The adjustment directions for flow rate, gradient, and pH are determined according to the purification feature vector, and the specific adjustment amounts for each parameter are accurately calculated based on these adjustment directions, including: Based on the purification feature vector, a hierarchical reinforcement learning method is used to optimize the process parameters of the chromatographic separation system in real time. The hierarchical reinforcement learning method includes an upper decision layer and a lower execution layer, wherein the upper decision layer determines the adjustment direction of flow rate, gradient and pH value according to the purification feature vector. Based on the adjustment direction, a parameter optimization objective function is constructed through the lower execution layer. The parameter optimization objective function comprehensively considers the stability and convergence speed of the process parameter adjustment, performs iterative calculations on each parameter, and accurately calculates the specific adjustment amount of each parameter.

[0064] The system architecture is divided into two layers: an upper decision-making layer and a lower execution layer. The upper decision-making layer receives the purification feature vector acquired from the sensor acquisition module as input. This purification feature vector includes the current column pressure data, elution peak resolution data, target product recovery rate data, and impurity residue data. Internally, the decision-making layer constructs a state-space mapping mechanism to map each dimension of the purification feature vector to a corresponding state interval. For example, pressure data is mapped to low-pressure, normal, and high-pressure intervals; resolution data is mapped to under-separation, moderate separation, and over-separation states; and recovery rate data is mapped to inefficient, normal, and efficient regions.

[0065] The upper decision layer uses a trained policy network to determine the mapped state. This policy network employs a multi-layer neural network structure, comprising an input layer, three hidden layers, and an output layer. The input layer receives a normalized, purified feature vector. The first hidden layer contains 128 neurons, the second hidden layer contains 64 neurons, and the third hidden layer contains 32 neurons. The output layer generates adjustment direction instructions for three parameters: flow rate, gradient, and pH. The adjustment direction for each parameter is encoded as three discrete actions: increase, maintain, and decrease. During training, the policy network receives reward feedback signals from the lower execution layer. These reward signals comprehensively consider the improvement in target product purity, changes in processing flux, and system stability indicators.

[0066] In the specific decision-making process, when the purification feature vector shows a pressure value of 0.85 MPa, lower than the set normal operating pressure range of 0.90 to 1.10 MPa, and the separation index is 1.35, below the critical value of 1.50 for under-separation, and the recovery rate is 72%, lower than the target recovery rate of 85%, the policy network calculates the activation values ​​of each neuron node based on this set of input data. The input layer nodes normalize the pressure value to 0.45, the separation index to 0.38, and the recovery rate to 0.60. The nodes in the first hidden layer receive these normalized values ​​and perform a weighted summation operation. The weight coefficients of each node have been optimized and determined during training. The weighted sum is processed by the activation function and then passed to the second hidden layer. After the stepwise processing of three hidden layers, the output layer finally generates an adjustment instruction to increase the flow rate by 0.15, an adjustment instruction to increase the gradient by 0.08, and an adjustment instruction to decrease the pH value by 0.05.

[0067] After receiving the adjustment direction instructions from the upper decision-making layer, the lower execution layer initiates the process of constructing the parameter optimization objective function. This objective function needs to achieve rapid convergence while ensuring the stability of process parameter adjustments. Stability constraints are achieved by setting limits on the rate of change of parameters: the single adjustment of flow rate cannot exceed 12% of the current value, the single adjustment of gradient cannot exceed 8% of the current value, and the single adjustment of pH cannot exceed 0.3 units. The convergence speed is controlled by a dynamic adjustment mechanism of the iteration step size. The initial iteration step size is set to a larger value to accelerate the response speed, and the step size is gradually decreased as the number of iterations increases to improve accuracy.

[0068] The calculation of the objective function for parameter optimization is performed iteratively, with each iteration comprising three stages: prediction and evaluation, error correction, and constraint verification. The prediction and evaluation stage establishes a predictive model based on the current parameter values ​​and the adjustment direction. This model, based on parameter response curves derived from historical data, predicts the system's state after adjustment. Taking flow rate as an example, with a current flow rate of 2.8 ml / min and an increasing adjustment direction, the prediction model estimates that increasing the flow rate to 3.2 ml / min will improve the separation to 1.62, while simultaneously increasing the pressure to 0.98 MPa. The error correction stage compares the deviation between the predicted state and the target state. With the target separation set at 1.80, the predicted value of 1.62 deviates from the target value by 0.18. The correction mechanism adjusts the iteration step size based on the deviation; when the deviation exceeds 0.15, a larger step size of 0.25 is used for rapid approximation.

[0069] The constraint verification process ensures that the calculated adjustment amount meets the system's physical and process constraints. Physical constraints include a flow rate range of 1.5 to 5.0 mL / min, a gradient range of 5% to 45% eluent concentration, and a pH range of 4.5 to 8.5. Process constraints include a minimum time interval of 180 seconds between adjacent adjustments, pressure fluctuations not exceeding 0.15 MPa, and temperature changes not exceeding 2 degrees Celsius. When the calculated adjustment amount for a parameter exceeds the constraint range, the execution layer automatically corrects it to the constraint boundary value. For example, if the calculated pH adjustment amount is a decrease of 0.45 units, exceeding the single adjustment range limit of 0.3, the execution layer corrects the actual adjustment amount to a decrease of 0.3 units, and the remaining 0.15 units are deferred to the next iteration.

[0070] The iterative calculation process continues until the convergence criteria are met. Convergence criteria include: the change in the objective function value over five consecutive iterations being less than a set threshold of 0.02, or the adjustment of each parameter being less than the minimum controllable precision. In the above data example, after 13 iterations, the flow rate was finally determined to be 3.35 mL / min, an increase of 0.55 mL / min from the initial value; the gradient was finally determined to be 23.8% eluent concentration, an increase of 2.3 percentage points from the initial value; and the pH value was finally determined to be 6.25, a decrease of 0.35 units from the initial value. The execution layer converts these precisely calculated adjustments into control signals, which drive the pump valves and mixing device to complete the actual parameter adjustment operations. The response time of the entire optimization process is 47 seconds. After adjustment, the system's separation degree reaches 1.83, and the recovery rate increases to 87%, meeting the expected optimization objectives.

[0071] A second aspect of the present invention provides an electronic device, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0072] A third aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0073] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0074] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A protein purification and analysis method based on optical detection and artificial intelligence systems, characterized in that, include: The protein sample to be analyzed is injected into the chromatographic separation system for purification. The absorbance, fluorescence intensity, light scattering intensity and circular dichroism signal of the protein sample to be analyzed are collected in real time by an optical detection device set at the outlet of the chromatographic separation system. The optical detection device adopts a four-channel parallel optical path structure, and multidimensional spectral data are collected synchronously through a tunable filter array. An adaptive signal compensation algorithm is used to eliminate crosstalk between channels to obtain compensated multidimensional spectral data. The compensated multidimensional spectral data is input into a pre-trained deep neural network model. The deep neural network model extracts spectral features through convolutional layers and captures temporal features through recurrent layers to generate a purification feature vector of the protein purification state. Based on the purification feature vector, a hierarchical reinforcement learning method is used to optimize the process parameters of the chromatographic separation system in real time. The adjustment direction of flow rate, gradient and pH value is determined according to the purification feature vector, and the specific adjustment amount of each parameter is accurately calculated based on the adjustment direction. The process parameters of the chromatographic separation system are adjusted according to the specific adjustment amount to complete the purification analysis of protein.

2. The method according to claim 1, characterized in that, The absorbance, fluorescence intensity, light scattering intensity, and circular dichroism signal of the protein sample to be analyzed are acquired in real time by an optical detection device installed at the outlet of the chromatographic separation system, including: A fluid connection channel is provided between the outlet of the optical detection device and the outlet of the chromatographic separation system. The inner diameter and length of the fluid connection channel are designed to match the flow rate and diffusion coefficient of the protein sample to be analyzed, so as to ensure that the protein sample to be analyzed maintains its separation state in the chromatographic separation system when it flows from the outlet of the chromatographic separation system through the fluid connection channel to the detection area of ​​the optical detection device. During real-time acquisition, the optical detection device measures the absorbance, fluorescence intensity, light scattering intensity, and circular dichroism signal of the protein sample to be analyzed through the spectral detection module.

3. The method according to claim 1, characterized in that, The optical detection device adopts a four-channel parallel optical path structure, synchronously acquires multidimensional spectral data through a tunable filter array, and uses an adaptive signal compensation algorithm to eliminate inter-channel crosstalk, obtaining compensated multidimensional spectral data including: The optical path structure of the optical detection device is configured. The optical detection device adopts a four-channel parallel optical path structure. Each channel is equipped with a tunable filter array. By adjusting the transmission band of the tunable filter array, selective acquisition of spectral signals of different wavelengths can be achieved. The optical detection device is activated to acquire spectral data, and multidimensional spectral data is acquired synchronously through the four-channel parallel optical path structure. The multidimensional spectral data is processed by signal processing, and an adaptive signal compensation algorithm is used to eliminate crosstalk between channels. The adaptive signal compensation algorithm calculates the crosstalk coefficient matrix between channels in real time and performs feedback correction on the signal of each channel according to the crosstalk coefficient matrix to obtain the compensated multidimensional spectral data.

4. The method according to claim 3, characterized in that, The optical detection device is activated to acquire spectral data, and multidimensional spectral data is acquired synchronously through the four-channel parallel optical path structure. Signal processing of the multidimensional spectral data, including eliminating inter-channel crosstalk using an adaptive signal compensation algorithm, includes: Multidimensional spectral data is acquired synchronously through a four-channel parallel optical path structure. The four-channel parallel optical path structure splits the incident light into four parallel beams through a beam splitter. Each beam passes through an independent filter and collimating lens and is received by a photodetector, thus realizing the synchronous acquisition of spectral information from the four channels. The multidimensional spectral data is processed by signal processing. An adaptive signal compensation algorithm is used to establish the crosstalk transfer function matrix between channels. The degree of crosstalk influence is calculated in real time based on the spectral information of the four channels. The compensation coefficient of each channel is determined by iterative optimization. The data of each channel is compensated accordingly. At the same time, the compensation coefficient is dynamically adjusted based on the compensation effect to achieve adaptive elimination of crosstalk. The multidimensional spectral data is corrected according to the compensation coefficient to obtain spectral data after eliminating inter-channel crosstalk.

5. The method according to claim 1, characterized in that, The compensated multidimensional spectral data is input into a pre-trained deep neural network model. The deep neural network model extracts spectral features through convolutional layers and captures temporal features through recurrent layers to generate a purification feature vector of the protein purification state, including: The compensated multidimensional spectral data is input into a pre-trained deep neural network model. The deep neural network model extracts spectral features through convolutional layers. The convolutional layers adopt a multi-scale convolutional kernel structure to extract spectral features in different frequency ranges in a hierarchical manner. The spectral features are subjected to time-series analysis. The time-series features are captured by a recurrent layer. The recurrent layer adopts a long short-term memory network structure to establish the time-series correlation of the spectral features and generate a purification feature vector of the protein purification state.

6. The method according to claim 1, characterized in that, Based on the purification feature vector, a hierarchical reinforcement learning method is used to optimize the process parameters of the chromatographic separation system in real time. The adjustment directions for flow rate, gradient, and pH are determined according to the purification feature vector, and the specific adjustment amounts for each parameter are accurately calculated based on these adjustment directions. Based on the purification feature vector, a hierarchical reinforcement learning method is used to optimize the process parameters of the chromatographic separation system in real time. The hierarchical reinforcement learning method includes an upper decision layer and a lower execution layer, wherein the upper decision layer determines the adjustment direction of flow rate, gradient and pH value according to the purification feature vector. Based on the adjustment direction, a parameter optimization objective function is constructed through the lower execution layer. The parameter optimization objective function comprehensively considers the stability and convergence speed of the process parameter adjustment, performs iterative calculations on each parameter, and accurately calculates the specific adjustment amount of each parameter.

7. A protein purification and analysis system based on optical detection and artificial intelligence, used to implement the method of any one of claims 1-6, characterized in that, include: The first unit is used to inject the protein sample to be analyzed into the chromatographic separation system for purification. The absorbance, fluorescence intensity, light scattering intensity and circular dichroism signal of the protein sample to be analyzed are collected in real time by an optical detection device set at the outlet of the chromatographic separation system. The optical detection device adopts a four-channel parallel optical path structure, and synchronously collects multidimensional spectral data through a tunable filter array. An adaptive signal compensation algorithm is used to eliminate crosstalk between channels to obtain compensated multidimensional spectral data. The second unit is used to input the compensated multidimensional spectral data into a pre-trained deep neural network model. The deep neural network model extracts spectral features through convolutional layers and captures temporal features through recurrent layers to generate a purification feature vector of the protein purification state. The third unit is used to optimize the process parameters of the chromatographic separation system in real time using a hierarchical reinforcement learning method based on the purification feature vector. It determines the adjustment direction of flow rate, gradient and pH value according to the purification feature vector, and accurately calculates the specific adjustment amount of each parameter based on the adjustment direction. The process parameters of the chromatographic separation system are adjusted according to the specific adjustment amount to complete the purification analysis of protein.

8. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 6.

Images