A tool for automated detection of monoclonal proteins and other patterns in clinical electrophoresis GEL scans

Abstract

An exemplary automated computer vision tool system and method are disclosed for the automated interpretation of a routine diagnostic analysis gel, such as serum protein electrophoresis (SPE) or immunofixation electrophoresis (IFE), based on machine learning and artificial intelligence (AI). The automated computer vision tool system can provide highly accurate identification of diagnostic monoclonal protein bands and saliency maps of such identified bands for clear interpretability and transparency.

Description

Attorney Docket No. 11258-048WO125018-01 A TOOL FOR AUTOMATED DETECTION OF MONOCLONAL PROTEINS AND OTHER PATTERNS IN CLINICAL ELECTROPHORESIS GEL SCANSRelated Application

[0001] This application claims priority to, and the benefit of, U. S. Provisional Patent Application No. 63 / 779,845, filed March 28, 2025, entitled “COMPUTER VISION TOOL FOR THE AUTOMATED DETECTION OF MONOCLONAL PROTEIN BANDS IN IMMUNOFIXATION ELECTROPHORESIS GEL SCANS,” which is incorporated by reference herein in its entirety.Background

[0002] Gel electrophoresis is a laboratory technique used to separate DNA, RNA, or proteins based on their size and charge. Electrophoresis is the migration of charged molecules in solution in response to an electric field. Their migration rate depends on the field strength, the net charge, the size and shape of the molecules, and the ionic strength, viscosity, pH, and temperature of the medium in which they are moving. As an analytical tool, electrophoresis is simple, rapid, and can be highly sensitive in certain configurations. It is used analytically to study the properties of a single charged species and as a separation technique.

[0003] Serum protein electrophoresis (SPE), as an example of gel electrophoresis, is a diagnostic method for separating and quantifying blood proteins, e.g., those altered by disease. Immunofixation electrophoresis (IFE), a type of SPE, can be used to determine disease type, seventy, and prognosis for monitoring disease progression. IFE involves adding anti-sera to gel lanes to identify7antibody heavy and light chains corresponding to specific monoclonal immunoglobulins.

[0004] Monoclonal gammopathy, characterized by monoclonal immunoglobulins or light chains in the blood or urine, ranges from benign conditions (e.g., monoclonal gammopathy of undetermined significance (MGUS)) to malignant diseases (e.g., multiple myeloma). IFE can be a screening tool for plasma cell disorders and Waldenstrom Macroglobulinemia, especially in patients with anemia, bone lesions, or renal dysfunction.

[0005] Interpreting SPE and IFE, among other electrophoresis gels, often requires expertise and is subject to variability. Automation of IFE interpretation can reduce turnaround time, labor costs, and errors, but no automated interpretation systems are commercially available. There is a benefit to developing an automated interpretation system.Attorney Docket No. 11258-048WO125018-01 Summary[0006| An exemplary automated computer vision tool system and method are disclosed for the automated interpretation of a routine diagnostic analysis gel, such as serum protein electrophoresis (SPE), based on machine learning and artificial intelligence (Al). The automated computer vision tool can provide highly accurate identification of diagnostic bands and corresponding saliency maps for clear interpretability and transparency.

[0007] In some embodiments, the exemplary7system and method are used for the automated interpretation of Immunofixation Electrophoresis (IFE), a type of serum protein electrophoresis (SPE), e.g., for clinical laboratories for the diagnosis of multiple myeloma, other diseases, and / or blood cell cancer. The exemplary system and method can identify the presence of diagnostic proteins in IFE gel, which can require significant time, effort, and expertise from medical staff to triage patients, thereby allowing interpreters more time to spend on difficult cases most pertinent to patient care.

[0008] Conservatively assuming that 70% of clinical IFE cases could be automatically signed out by the automated computer vision tool at a large clinical laboratory' processing -1,000 samples per weekday, this would free a substantial proportion of pathologist effort currently devoted to routine case review (estimated at ~$2.73 million annually in equivalent labor), enabling redistribution of expert time to higher-value activities. This is based on an assumed labor rate of $150 / hr for clinical pathologists or clinical chemists, and that it takes -2 hours to visually review, interpret and sign out 20 IFE cases on a daily basis.

[0009] In an aspect, a system is disclosed comprising: a processor; and a memory having instructions stored thereon, wherein execution of the instructions causes the processor to: receive, via the processor, an electrophoresis gel image having one or more gel lanes; determine, via one or more trained Al classifiers, via the processor, one or more probability7, score, or indication values each for a presence or non-presence of a respective band in a respective gel lane of the one or more gel lanes, wherein the one or more trained Al classifiers were generated using training data derived from a set of electrophoresis gel training images; and output, via the processor, the one or more probability, score, or indication values, wherein the outputted probability7, score, or indication values are used in subsequent clinical or research for detection of a protein, antibody, nucleic acid, or disease or condition associated with a protein, antibody, nucleic acid.

[0010] In some embodiments, the one or more trained Al classifiers comprises: a coarse Al model configured to detect one or more bands in the electrophoresis gel image; and a setAttorney Docket No. 11258-048WO125018-01 of fine Al model configured to detect a band in the respective band in the electrophoresis gel image.

[0011] In some embodiments, the one or more trained Al classifiers each includes a neural network (e.g., residual neural network).

[0012] In some embodiments, the one or more trained Al classifiers each include a convolutional block attention model (CBAM) in each neural network (e.g., residual neural network),

[0013] In some embodiments, the one or more trained Al classifiers were generated using training data formatted as a tensor, and wherein the one or more trained Al classifiers operates on the electrophoresis gel image converted to a tensor.

[0014] In some embodiments, the one or more trained Al classifiers were generated using augmented training data derived from a set of electrophoresis gel training images, wherein the augmented training data each includes at least one of: image artifacts, ghost, speckle, shaped image elements, or a combination thereof.

[0015] In some embodiments, the electrophoresis gel image was acquired using immunofixation electrophoresis for a set of antibodies.

[0016] In some embodiments, the instructions are performed for a batch of electrophoresis gel images to identify one or more proteins, antibodies, nucleic acids, or disease or condition, for each respective electrophoresis gel image of the batch.

[0017] In some embodiments, the one or more gel lanes in the gel image include an antibody selected from the group consisting of IgG, IgM, IgA, K, and.

[0018] In some embodiments, the electrophoresis gel image was acquired using at least one of: SDS-PAGE (Sodium Dodecyl Sulfate-Polyacrylamide Gel Electrophoresis), Native PAGE, Two-Dimensional (2D) Gel Electrophoresis, Blue Nati e PAGE (BN-PAGE), Clear Native PAGE (CN-PAGE), Isoelectric Focusing (IEF), Immunoelectrophoresis, Rocket Immunoelectrophoresis, or Western Blotting (Immunoblotting).

[0019] In some embodiments, the execution of the instructions further causes the processor to: in response to the one or more probability, score, or indication values exceeding a predefined threshold value (e.g., 0.8), generate saliency maps, each having the respective band in the respective gel lane of the one or more gel lanes; and output the generated saliency maps, wherein the outputted generated saliency maps are subsequently used for clinical or research for detection of a protein, antibody, nucleic acid, disease, or condition associated with a protein, antibody, lipid, or nucleic acid.Attorney Docket No. 11258-048WO125018-01

[0020] In some embodiments, the execution of the instructions, prior to determining the one or more probability, score, or indication values, further causes the processor to: determine, via the processor, a presence or non-presence of the one or more gel lanes in the received electrophoresis gel image; determine, via the processor, a rotation or croppness of the one or more gel lanes in the received electrophoresis gel image; and m response to the one or more gel lanes having the non-presence, the rotation, or the croppness, generate, via the processor, an alert indicating the non-presence, the rotation, or the croppness of the one or more gel lanes.[00211 In some embodiments, the execution of the instructions further causes the processor to: display, via a user interface, the received electrophoresis gel image; adjust, via the user interface, contrast, brightness, tone curve, and scale of the displayed received electrophoresis gel image; receive, via the user interface, a first user-interpretation value (e.g., negative, positive, cannot interpret) indicating a presence or non-presence of the one or more bands in the electrophoresis gel image; receive, via the user interface, a second userinterpretation value (e.g., confident, faint, equivocal) indicating the presence or non-presence of the one or more bands in respective gel lanes; receive, via tire user interface, a third user¬ interpretation value (e.g., location marking) indicating positions of the one or more bands in the respective gel lanes; and output, via the user interface, the received first, second, and third user-interpretation values, wherein the outputted received first, second, and third user¬ interpretation values are used for validating the determining of the one or more probability, score, or indicati on values for the presence or non-presence of bands in respecti ve gel lanes.

[0022] In some embodiments, the respective band corresponds to a monoclonal protein in the respective gel lane.

[0023] In another aspect, a method is disclosed comprising: receiving, via a processor, an electrophoresis gel image having one or more gel lanes; determining, via one or more trained Al classifiers executed by the processor, one or more probability, score, or indication values each for a presence or non-presence of a respective band in a respective gel lane of the one or more gel lanes, wherein the one or more trained Al classifiers were generated using training data derived from a set of electrophoresis gel training images; and outputting, via the processor, the one or more probability, score, or indication values, wherein the outputted probability, score, or indication values are used in subsequent clinical or research for detection of a protein, antibody, nucleic acid, or disease or condition associated with a protein, antibody, nucleic acid.Attorney Docket No. 11258-048WO125018-01

[0024] In some embodiments, the one or more trained Al classifiers comprises: a coarse Al model configured to detect one or more bands in the electrophoresis gel image; and a set of fine Al model configured to detect a band in the respective band in the electrophoresis gel image,

[0025] In some embodiments, the one or more trained Al classifiers each includes a neural network (e.g., residual neural network) (e.g., wherein the one or more trained Al classifiers each include a convolutional block attention model (CBAM) in each neural network (e.g., residual neural network)).

[0026] In some embodiments, the one or more trained Al classifiers were generated using training data formatted as a tensor, and wherein the method includes: converting the electrophoresis gel image to a tensor as input to the one or more trained Al classifiers.

[0027] In some embodiments, the one or more trained Al classifiers were generated using augmented training data derived from a set of electrophoresis gel training images, wherein the augmented training data each includes at least one of: image artifacts, ghost, speckle, shaped image elements, or a combination thereof (e.g., wherein the method further includes generating the augmented training data for the training).

[0028] In some embodiments, the electrophoresis gel image was acquired using immunofixation electrophoresis for a set of antibodies,

[0029] In some embodiments, the method is performed for a batch of electrophoresis gel images to identify one or more proteins, antibodies, nucleic acids, or disease or condition, for each respective electrophoresis gel image of the batch,

[0030] In some embodiments, the one or more gel lanes in the gel image include an antibody selected from the group consisting of IgG, IgM, IgA, K, and X.

[0031] In some embodiments, a non-transitory computer-readable medium having instructions stored thereon is disclosed, wherein execution of the instruction causes a processor to: receive, via the processor, an electrophoresis gel image having one or more gel lanes; determine, via one or more trained Al classifiers, via the processor, one or more probability, score, or indication values each for a presence or non-presence of a respective band in a respective gel lane of the one or more gel lanes, wherein the one or more trained Al classifiers were generated from a set of electrophoresis gel training images; and output, via the processor, the one or more probability, score, or indication values, wherein the outputted probability, score, or indication values are used in a subsequent clinical or research for detection of a protein, antibody, nucleic acid, or disease or condition associated with a protein, antibody, lipid, or nucleic acid.Attorney Docket No. 11258-048WO125018-01

[0032] In some embodiments, the electrophoresis gel images were generated by flatbed scanning on either a reflective document scanner, such as LED-based scanner illumination with contact image sensor (CIS) technology' or transparency film scanner, such as LED-based charge-coupled device (CCD) technology. Other technologies that may be used to generate electrophoresis gel images include optical density measurements by a photodiode densitometer system, UV or blue light transillumination, or complementary metal oxide semiconductor (CMOS) technology'.Brief Description of the Drawings

[0033] Figs. 1A, IB, 1C, and ID each show an example automated computer vision tool system for the automated identification and interpretation of a diagnostic analysis gel based on machine learning and artificial intelligence (Al) trained in accordance with an illustrative embodiment.

[0034] Figs. 2A, 2B, and 2C each show an example configuration 200 (shown as 200a, 200b, 200c) for a trained Al classifier module (e.g., 102, etc.) or a model in an Al train system (e.g., 110, 110’, etc.), e.g., as described in relation to Figs. IB,

[0035] Figs. 3A - 3G each show an example graphical user interface (GUI) of a validation module in the exemplary system, in accordance with an illustrative embodiment.

[0036] Figs. 4A - 4G, respectively, show' components of an example pipeline / operation flow' for an automated computer vision tool system, e.g., for detecting the presence of monoclonal proteins, and other molecules of interest as discussed herein. Specifically, Figs.4A - 4G show example operation flow, image segmentation operation, fine model training process, and saliency map generation process of the exemplary' system, where Fig. 4A show's an example Al model, Fig. 4B shows pre-processing operations, Fig. 4C shows neural network architecture improvements to the Al model that may be implemented, Fig. 4D shows an example training operation with improved training using combined loss and optimization. Fig. 4E shows example data augmentation operation that may be performed, and Figs. 4F and 4G show the saliency map processing operation to be used for the training. While the example is described in relation to an IFE gel scan, other gels and molecules of interest may be identified / classified using a trained Al model generated using a similar training and analysis operation, with saliency mask processing.

[0037] Figs. 5A - 51 show' a first set of test results and performance evaluation for the fine models and the final model used in an experimental computer vision tool system, as described in relation to Figs. 1 - 4. Fig. 5A shows the combined loss over epochs for trainingAttorney Docket No. 11258-048WO125018-01 the coarse model and five fine models for detecting monoclonal protein bands. Fig. 5B shows the coarse model results in test data for detecting monoclonal proteins in the image. Fig. 5C shows the final model results in test data for detecting monoclonal proteins in the IgG lane. Fig. 5D shows the final model results in test data for detecting monoclonal proteins in the IgA lane. Fig. 5E shows the final model results in test data for detecting monoclonal proteins in the IgM lane. Fig. 5F shows the final model results in test data for detecting monoclonal proteins in the K lane. Fig. 5G shows the final model results in test data for detecting monoclonal proteins in the lane. Fig. 5H shows the performance of the final model for identifying heavy / light chain combinations by saliency maps m test data. Fig. 51 shows the combined loss during quality control model training.

[0038] Figs. 6A - 6K show a second set of test results and performance evaluation for the fine models and the final models used in the experimental system. Fig. 6A shows a training and evaluation process of models in the experimental system. Fig. 6B shows the training process of a quality control model in the experimental system. Fig. 6C shows the final model results in the test data for detecting monoclonal proteins in the entire image, in the IgG lane, in the IgA lane, in the IgM lane, in the K lane, and in the A lane. Fig. 6D shows the final model results, in the non-equivocal labeled samples for the validation data, for detecting monoclonal proteins in the entire image, in the IgG lane, in the IgA lane, in the IgM lane, in the K lane, and in the X lane. Fig. 6G shows the performance results and probability distribution results, in the validation dataset of quality control images, of the quality control model. Fig. 6H shows saliency maps generated by the experimental system. Fig. 61 shows the distributions of the images in the validation and test datasets classified as definitive negatives, definitive positives, and equivocal. Fig. 6J shows an operational flow of the triaging logic for the experimental system. Fig. 6K shows 3 samples in the test dataset predicted as definitive false negatives for the presence of disease,

[0039] Figs. 7A - 7C show example model architecture and training for an example trained Al model configured in an automated computer vision tool system.Detailed Description

[0040] Some references, which may include various patents, patent applications, and publications, are cited in a reference list and discussed in the disclosure provided herein. The citation and / or discussion of such references is provided merely to clarify the description of the disclosed technology and is not an admission that any such reference is “prior art” to any aspects of the disclosed technology described herein. In temis of notation, “[n]” correspondsAttorney Docket No. 11258-048WO125018-01 to the nth reference in the list. For example, [1] refers to the first reference in the list. All references cited and discussed m this specification are incorporated herein by reference in their entirety and to the same extent as if each reference were individually incorporated by reference.

[0041] Definitions

[0042] Gel electrophoresis is a laboratory technique that separates DNA, RNA, or proteins based on their size and charge. Electrophoresis is the migration of charged molecules in solution in response to an electric field. Their migration rate depends on the field's strength, the net charge, the size and shape of the molecules, and the ionic strength, viscosity, pH, and temperature of the medium in which the molecules are moving. As an analytical tool, electrophoresis is simple, rapid, and highly sensitive. It is used analytically to study the properties of a single charged species, and as a separation technique.

[0043] A “protein”, "polypeptide", or “peptide” each refer to a polymer of amino acids and does not imply a specific length of a polymer of amino acids. Thus, for example, the terms peptide, oligopeptide, protein, antibody, and enzyme are included within the definition of polypeptide. This term also includes polypeptides with post¬ expression modification, such as glycosylation (e.g., the addition of a saccharide), acetylation, phosphorylation, and the like.

[0044] The term “amino acid,” includes but is not limited to amino acids contained in the group consisting of alanine (Ala or A), cysteine (Cys or C), aspartic acid (Asp or D), glutamic acid (Glu or E), phenylalanine (Phe or F), glycine (Gly or G), histidine (His or H), isoleucine (He or I), lysine (Lys or K), leucine (Leu or L), methionine (Met or M), asparagine (Asn or N), proline (Pro or P), glutamine (Gin or Q), arginine (Arg or R), serine (Ser or S), threonine (Thr or T), valine (Vai or V), tryptophan (Trp or W), and tyrosine (Tyr or Y) residues. The term “amino acid residue” also may include amino acid residues contained in the group consisting of homocysteine, 2- Aminoadipic acid, -Ethylasparagine, 3-Annnoadipic acid, Hydroxylysine, p-alanine, p-Amino-propionic acid, allo-Hydroxylysine acid, 2-Aminobutyric acid, 3-Hydroxyproline, 4-Aminobutyric acid, 4-Hydroxyproline, piperidinic acid, 6-Ammocaproic acid, Isodesmosine, 2-Aminoheptanoic acid, allo- Isoleucine, 2-Aminoisobutyric acid, N-Methylglycine, sarcosine, 3-Aininoisobutyric acid, N- Methylisoleucine, 2-Aminopimelic acid, 6-N-Methyllysine, 2,4-Diaminobutyric acid, N-Methylvaline, Desmosine, Norvaline, 2,2’-Diaminopimelic acid, Norleucine, 2,3-Diaminopropionic acid, Ornithine, and N -Ethylglycine. Typically, the amide linkages of theAttorney Docket No. 11258-048WO125018-01 peptides are formed from an amino group of the backbone of one amino acid and a carboxyl group of the backbone of another amino acid.

[0045] It should also be noted that amino acids and derivatives (except glycine) occur in two isomeric forms: L-forms or D-forms. The L- and D- forms represent the same atoms of an amino acid, however, the atoms can have different arrangements, which can impact the amino acid properties and functions. The two forms are similar in that they both occur naturally and comprise a central carbon atom, at least one hydrogen atom, a carboxylic group, an amine group, and a variable group. The two forms differ in that they are usually mirrored images of each other, wherein the location of the anime group varies. L-amino acids are used in protein synthesis, while D-amino acids are less common. L-amino acids rotate counterclockwise or to the left in a process known as levorotation. D-amino acids rotate clockwise or to the light in a process known as dextrorotation. L-amino acids are used to synthesize proteins, while D-amino acids are found in the cell walls of bacteria.

[0046] Generally, the sample is run in a support matrix such as paper, cellulose acetate, starch gel, agarose, or poly acrylamide gel. The matrix inhibits convective mixing caused by heating. It provides a record of the electrophoretic run: at the end of the run, the matrix can be stained and used for scanning, autoradiography or storage. In addition, the most commonly- used support matrices - agarose and polyacrylamide - provide a means of separating molecules by size, in that they are porous gels. A porous gel may act as a sieve by retarding, or in some cases completely obstructing, the movement of large macromolecules w hile allowing smaller molecules to migrate freely. Because dilute agarose gels are generally more rigid and easy to handle than polyacrylamide of the same concentration, agarose separates larger macromolecules such as nucleic acids, large proteins and protein complexes.Polyacrylamide, which is easy to handle and to make at higher concentrations, is used to separate most proteins and small oligonucleotides that require a small gel pore size for retardation. And therefore, this technique involves the preparation of a gel matrix (usually agarose for DNA / RNA or polyacrylamide for proteins), which is prepared and placed in a gel box, followed by a step wherein the DNA or protein samples are mixed with a loading dye and loaded into wells at one end of the gel. The gel is submerged in a buffer, and an electric current is applied. Since DNA is negatively charged, it moves towards the positive electrode. Smaller molecules migrate faster through the gel, while larger molecules move slower, resulting in size-based separation. The separated molecules are stained with a dye (e.g., ethidium bromide for DNA) and visualized under LTV light.Attorney Docket No. 11258-048WO125018-01

[0047] There are different types of gel electrophoresis methods, including those used to separate proteins and nucleic acids (DNA and RNA).

[0048] A “nucleic acid” is a chemical compound that is the primary information-carrying molecules in cells and makes up the cellular genetic material. Nucleic acids comprise nucleotides, monomers of 5-carbon sugar (usually ribose or deoxyribose), a phosphate group, and a nitrogenous base. A nucleic acid can also be a deoxyribonucleic acid (DNA) or a ribonucleic acid (RNA). A chimeric nucleic acid comprises two or more of the same kind of nucleic acid fused to form one compound comprising genetic material.

[0049] A “nucleotide” is a compound consisting of a nucleoside, which consists of a nitrogenous base and a 5-carbon sugar, linked to a phosphate group, forming the basic structural unit of nucleic acids, such as DNA or RNA, The four types of nucleotides are adenine (A), cytosine (C), guanine (G), and thymine (T), each of which is bound together by a phosphodiester bond to form a nucleic acid molecule.

[0050] Some exemplary' protein-separation gel electrophoresis methods are SDS-PAGE (Sodium Dodecyl Sulfate-Polyacrylamide Gel Electrophoresis), Native PAGE, Two- Dimensional (2D) Gel Electrophoresis, Blue Native PAGE (BN-PAGE), Clear Native PAGE (CN-PAGE), Isoelectric Focusing (IEF), Immunoelectrophoresis, Rocket Immunoelectrophoresis, Western Blotting (Immunoblotting), and agarose gel electrophoresis.

[0051] Agarose gel electrophoresis. Agarose gel electrophoresis is a technique that separates DNA fragments by size. Clinical 1FE Apically employs an agarose gel. Example clinical applications of agarose gel electrophoresis include Lipoprotein electrophoresis, Hemoglobin electrophoresis, and Von Willebrand factor multimer. Lipoprotein particles are a mixture of proteins, lipids (e.g., triglycerides), and cholesterol molecules. Hemoglobin and vWF are thus also examples of proteins.

[0052] SDS-PAGE. SDS-PAGE (Sodium Dodecyl Sulfate-Polyacrylamide Gel Electrophoresis) is used for protein separation based on molecular weight. Sodium dodecyl sulphate (SDS) is an anionic detergent that denatures proteins by “wrapping around” the polypeptide backbone, and SDS binds to proteins fairly specifically in amass ratio of 1.4:1. In so doing, SDS confers a negative charge to the polypeptide in proportion to its length. Further, it is usually necessary' to reduce disulfide bridges in proteins (denature) before they adopt the required random-coil configuration for separation by size; this is done with 2-mercaptoethanol or dithiothreitol (DTT). In denaturing SDS-PAGE separations, migration is determined not by the intrinsic electrical charge of the polypeptide, but by molecular weight.Attorney Docket No. 11258-048WO125018-01

[0053] Determination of molecular weight is done by SDS-PAGE of proteins of known molecular weight, along with the protein to be characterized. A linear relationship exists between the logarithm of the molecular weight of an SDS-denatured polypeptide, or native nucleic acid, and its Rf. The Rf is calculated as the ratio of the distance migrated by the molecule to that migrated by a marker dye-front. A simple way of determining relative molecular weight by electrophoresis (Mr) is to plot a standard curve of distance migrated vs. log10MW for known samples, and read off the log*Mr* of the sample after measuring distance migrated on the same gel.

[0054] Native PAGE. Native PAGE separates proteins without denaturation (based on charge and size).

[0055] Two-Dimensional (2D) Gel Electrophoresis. Two-Dimensional (2D) Gel Electrophoresis is used for High-resolution protein separation. In 2D gel electrophoresis, proteins are fractionated first based on one physical property, and, in a second step, based on another. For example, isoelectric focusing can be used for the first dimension, conveniently carried out in a tube gel and SDS electrophoresis in a slab gel, which can be used for the second dimension. One example of a procedure is that of O’Farrell, P. H., High Resolution Two-dimensional Electrophoresis of Proteins, J. Biol. Chem. 250:4007-4021 (1975), herein incorporated by reference in its entirety for its teaching regarding two-dimensional electrophoresis methods. Other examples include but are not limited to, those found in Anderson, L and Anderson, NG, High resolution two-dimensional electrophoresis of human plasma proteins. Proc. Natl, Acad, Sci. 74:5421-5425 (1977), Omstein, I... Disc electrophoresis, L. Ann. N. Y. Acad. Sci. 121:321349 (1964), each of which is herein incorporated by reference in its entirety for teachings regarding electrophoresis methods. Laemmli, U. K., Cleavage of structural proteins during the assembly of the head of bacteriophage T4, Nature 227:680 (1970), which is herein incorporated by reference in its entirety for teachings regarding electrophoresis methods, discloses a discontinuous system for resolving proteins denatured with SDS. The leading ion in the Laemmli buffer system is chloride, and the trailing ion is glycine. Accordingly, the resolving gel and the stacking gel comprise Tris-HCl buffers (of different concentrations and pH levels), while the tank buffer is Tris-glycine. All buffers contain 0.1% SDS.

[0056] Blue Native PAGE. Blue Native PAGE (BN-PAGE) is used to separate intact protein complexes. Clear Native PAGE (CN-PAGE) is used to separate proteins under native conditions without Coomassie Blue dye.Attorney Docket No. 11258-048WO125018-01

[0057] Isoelectric Focusing. Isoelectric Focusing (IEF) may be used as part of Blue Native PAGE to separate proteins based on their isoelectric point (pl). Proteins are amphoteric compounds; their net charge, therefore, is determined by the pH of the medium in which they are suspended. In a solution with a pH above its isoelectric point, a protein has a net negative charge and migrates towards the anode in an electrical field. Below its isoelectric point, the protein is positively charged and migrates towards the cathode. The net charge carried by a protein is, in addition, independent of its size, i.e., the charge carried per unit mass (or length, given that proteins and nucleic acids are linear macromolecules) of a molecule differs from protein to protein. At a given pH, therefore, and under non-denaturing conditions, the electrophoretic separation of proteins is determined by both size and charge of the molecules, IEF combines electrophoresis with antigen-antibody reactions,

[0058] Immunoelectrophoresis. Immunoelectrophoresis is used to identify and quantify specific proteins, e.g., monoclonal immunoglobulins, among others described or referenced herein. Examples of proteins of interest include hemoglobin, von Willebrand factor, and lipoproteins. For immunofixation, proteins of interest can include IgG, IgM, IgA heavy chains, kappa, and lambda light chains, which are the major components of antibodies typically detected in IFE, in addition to IgD (less common) and IgE (less common) heavy chains.

[0059] Rocket Immunoelectrophoresis. Rocket Immunoelectrophoresis is used to quantify specific proteins using antibodies, e.g., ([Ab–Ag]n), among others described or referenced herein.

[0060] Western Blotting. Western Blotting (immunoblotting) is used to detect specific proteins in a sample. Western blotting or immunoblotting allows the determination of the molecular mass of a protein and the measurement of relative amounts of the protein present in different samples. Detection methods include chemiluminescence and chromogenic detection. Standard methods for Western blot analysis can be found in, for example, D M. Bollag et al., Protein Methods (2d edition 1996) and E. Harlow & D. Lane, Antibodies, a Laboratory Manual (1988), U. S. Patent 4,452,901, each of which is herein incorporated by reference in their entirety for teachings regarding Western blot methods. Generally, proteins are separated by gel electrophoresis, usually SDS-PAGE. The proteins are transferred to a sheet of special blotting paper, e.g., nitrocellulose, though other types of paper, or membranes, can be used. The proteins retain the same pattern of separation they had on the gel. The blot is incubated with a generic protein (such as milk proteins) to bind to anyAttorney Docket No. 11258-048WO125018-01 remaining sticky places on the nitrocellulose. An antibody is then added to the solution, which can bind to its specific protein.

[0061] The attachment of specific antibodies to specific immobilized antigens can be readily visualized by indirect enzyme immunoassay techniques, usually using a chromogenic substrate (e.g., alkaline phosphatase or horseradish peroxidase) or chemiluminescent substrates. Other possibilities for probing include the use of fluorescent or radioisotope labels (e.g., fluorescein,125I). Probes for the detection of antibody binding can be conjugated to anti-immunoglobulins, conjugated to Staphylococcal Protein A (which binds IgG), or probes to biotinylated primary antibodies (e.g., conjugated avidin / streptavidin).

[0062] The technique's power lies in the simultaneous detection of a specific protein by¬ means of its antigenicity and molecular mass. Proteins are first separated by mass in the SDS-PAGE step and then specifically detected in the immunoassay step. Thus, protein standards (ladders) can be run simultaneously to approximate the molecular mass of the protein of interest in a heterogeneous sample.

[0063] Some exemplary nucleic acid-separation gel electrophoresis methods are Agarose Gel Electrophoresis (AGE), Polyacrylamide Gel Electrophoresis (PAGE), Pulsed-Field Gel Electrophoresis (PFGE), Capillary Gel Electrophoresis (CGE), Denaturing Gradient Gel Electrophoresis (DGGE), Temperature Gradient Gel Electrophoresis (TGGE). Single-Strand Conformation Polymorphism (SSCP) Electrophoresis, gel shift assay or electrophoretic mobility shift assay (EMSA), and Microchip Gel Electrophoresis.

[0064] Agarose Gel Electrophoresis (AGE) is used for DNA and RNA separation based on size. Polyacrylamide Gel Electrophoresis (PAGE) for Nucleic Acids is used for high- resolution separation of small DNA and RNA fragments.

[0065] Pulsed-Field Gel Electrophoresis (PFGE) is used to separate large DNA fragments (50-10,000 kb) using alternating electric fields to separate large DNA molecules.

[0066] Capillary Gel Electrophoresis (CGE) uses capillary tubes and high voltage to separate nucleic acids and proteins at high resolution.

[0067] Denaturing Gradient Gel Electrophoresis (DGGE) is used to detect mutations and genetic polymorphisms. During DGGE, DNA fragments are separated in a chemical gradient due to sequence differences. Temperature Gradient Gel Electrophoresis (TGGE) is used similarly to DGGE but uses a temperature gradient instead of a chemical gradient.

[0068] Single-Strand Conformation Polymorphism (SSCP) Electrophoresis is used to detect single nucleotide polymorphisms (SNPs) in DNA, wherein the single-stranded DNA adopts different conformations based on sequence.Attorney Docket No. 11258-048WO125018-01

[0069] Microchip Gel Electrophoresis is used for rapid and automated DNA separation using microfluidic channels.

[0070] The gel shift assay or electrophoretic mobility shift assay (EMSA) can be used to detect the interactions between DNA binding proteins and their cognate DNA recognition sequences, both qualitatively and quantitatively. Exemplary techniques are described in Ornstein L., Disc electrophoresis - 1: Background and theory, Ann. NY Acad. Sci. 121:321- 349 (1964), and Matsudaira, PT and DR Burgess, SDS microslab linear gradient polyacrylamide gel electrophoresis, Anal. Biochem. 87:386-396 (1987), each of which is herein incorporated by reference in its entirety for teachings regarding gel-shift assays.

[0071] In a general gel-shift assay, purified proteins or crude cell extracts can be incubated with a labeled (e.g.,32P -radiolabeled) DNA or RNA probe, followed by separation of the complexes from the free probe through a nondenaturing polyacrylamide gel. The complexes migrate more slowly through the gel than unbound probe. Depending on the activity of the binding protein, a labeled probe can be either double-stranded or singlestranded. To detect DNA binding proteins such as transcription factors, either purified or partially purified proteins, or nuclear cell extracts can be used. To detect RNA binding proteins, either purified or partially purified proteins, or nuclear or cytoplasmic cell extracts can be used. The specificity of the DNA or RNA binding protein for the putative binding site is established by competition experiments using DNA or RNA fragments, oligonucleotides containing a binding site for the protein of interest, or other unrelated sequences. The differences in the nature and intensity of the complex formed in the presence of specific and nonspecific competitor allow identification of specific interactions. Refer to Promega, Gel Shift Assay FAQ, available at <http: / / www.promega.com / faq / gelshfaq.html> (last visited March 25, 2005), which is herein incorporated by reference in its entirety for teachings regarding gel shift methods.

[0072] Gel shift methods can include using, for example, colloidal forms of COOMASSIE (Imperial Chemicals Industries, Ltd) blue stain to detect proteins in gels such as polyacry lamide electrophoresis gels. Such methods are described, for example, in Neuhoff et al., Electrophoresis 6:427-448 (1985), and Neuhoff et al., Electrophoresis 9:255-262 (1988), each of which is herein incorporated by reference in its entirety for teachings regarding gel shift methods. In addition to the conventional protein assay methods referenced above, a combination cleaning and protein staining composition is described in U.S. Patent 5,424,000, herein incorporated by reference in its entirety for its teaching regarding gel shift methods. The solutions can include phosphoric, sulfuric, nitric acids, and acid violet dye.Attorney Docket No. 11258-048WO125018-01

[0073] Other gel electrophoresis methods include Field Inversion Gel Electrophoresis (FIGE), Transverse Alternating Field Electrophoresis (TAFE), and Microchip Electrophoresis. Field Inversion Gel Electrophoresis (FIGE) is used to separate large DNA molecules, wherein the field direction is alternated to improve the resolution of large fragments. Transverse Alternating Field Electrophoresis (TAFE) is used for separating very large DNA fragments, and Microchip Electrophoresis is used for high-speed, small-volume electrophoresis using a lab-on-a-chip system,

[0074] Serum protein electrophoresis (SPE or SPEP) is a laboratory technique used to separate and analyze proteins in the blood serum based on their size and electrical charge. It is commonly used to detect abnormalities in protein levels. It can help diagnose various conditions, such as multiple myeloma, monoclonal gammopathies, immune disorders, and liver or kidney diseases.

[0075] Electrophoresis separates a sample into protein fractions. During electrophoresis, samples from a subject such as serum proteins are separated into five main fractions: 1) albumin, the most abundant protein, is responsible for maintaining oncotic pressure and transporting substances; 2) alpha-1 globulins--- including alpha- 1 antitrypsin --help protect tissues from inflammation; 3) alpha-2 globulins include haptoglobin (binds free hemoglobin) and ceruloplasmin (carries copper), 4) beta globulins - Includes transferrin (iron transport) and complement proteins (immune response), and 5) gamma globulins -- Contains immunoglobulins (IgG, IgA, IgM, etc.), which are essential for immune function.

[0076] SPE is often followed by immunofixation electrophoresis (IFE) to further characterize abnormal protein bands. IFE is a laboratory technique for identifying and characterizing specific proteins in a sample from a subject, particularly monoclonal immunoglobulins (M proteins). During IFE, separated proteins are exposed to specific antibodies against immunoglobulin heavy chains (IgG, IgA, IgM) and light chains (kappa and lambda). The antibody-bound proteins form visible precipitates, allowing for the identification of monoclonal immunoglobulins.

[0077] After separation, molecules in a gel need to be visualized for analysis. A further step after electrophoresis can include methods for detecting or quantifying the amount of a molecule of interest (such as the disclosed biomarkers or their antibodies) in a sample, which methods generally involve the detection or quantitation of any immune complexes formed during the binding process. In general, the detection of immunocomplex formation is well known in the art and can be achieved by applying numerous approaches. These methods are generally based upon detecting a label or marker, such as any radioactive, fluorescent,Attorney Docket No. 11258-048WO125018-01 biological or enzymatic tags or any other known label. Depending on whether DNA, RNA, or proteins are being analyzed, various staining and detection methods are used. Some of the most common detection methods of proteins include Coomassie Brilliant Blue (CBB) Staining, Silver Staining, Fluorescent Staining, Radiolabeling, and Acid Violet Stain. IFE is most commonly done with Acid Violet Stain. Serum protein electrophoresis (SPE) is most commonly done with Acid Blue Stain.

[0078] In Coomassie Brilliant Blue (CBB) Staining, CBB binds to proteins, causing a color change. This simple and non-toxic technique has good sensitivity for proteins at about 100 ng in concentration.

[0079] During Silver Staining, Silver binds to proteins and is reduced to metallic silver. This staining is highly sensitive (~0.1 ng).

[0080] Fluorescent Staining (SYPRO Ruby, Oriole, Flamingo) uses fluorescent dyes that bind to proteins and are detected using UV or blue light. Fluorescent signals are then captured with imaging systems. This method is highly sensitive and non-toxic.

[0081] During Radiolabeling (35S, 14C, 1251) proteins are labeled with radioactive isotopes and detected by autoradiography. Detection utilizes X-ray film. This method is ultra¬ sensitive.

[0082] During Western blotting, proteins are transferred to a membrane and detected using antibodies, chemiluminescence, fluorescence, or colorimetric detection. This method has high specificity for target proteins. The principle of chemiluminescence (Luminol-Based) is based on enzyme-linked detection, which produces light emission and is then captured by imaging systems. This method uses chemiluminescent reagents and has high sensitivity. Further densitometry & Image Analysis uses digital imaging software that quantifies band intensity. Proteins can also be excised from gels and analyzed by Mass Spectrometry (MS) for identification using the mass-to-charge ratio of peptides, which leads to the identification of protein sequences.

[0083] Some DNA and RNA detection methods include Ethidium Bromide (EtBr) Staining, SYBR Green / SYBR Gold / GelRed Staining, and Silver Staining.

[0084] Ethidium Bromide (EtBr) Staining, wherein ethidium bromide intercalates between DNA / RNA bases and fluoresces under UV light.

[0085] SYBR Green / SYBR Gold / GelRed can bind to nucleic acids and fluoresce under UV or blue light. Silver Staining can also be used wherein Silver ions bind to nucleic acids and are reduced to metallic silver for visualization. This technique is extremely sensitive (~10Attorney Docket No. 11258-048WO125018-01 pg detection). Methylene Blue can be used as a stain too. Methylene blue binds to DNA and RNA but requires extended staining times.

[0086] As used herein, a label can include a fluorescent dye, a member of a binding pair, such as biotin / streptavidin, a metal (e.g., gold), or an epitope tag that can specifically interact with a molecule that can be detected, such as by producing a colored substrate or fluorescence. Substances suitable for detectably labeling proteins include fluorescent dyes (also known herein as fluorochromes and fluorophores) and enzymes that react with colorimetric substrates (e.g., horseradish peroxidase). The use of fluorescent dyes is generally preferred in the practice of the invention as they can be detected at very low amounts.Furthermore, when multiple antigens are reacted with a single array, each antigen can be labeled with a distinct fluorescent compound for simultaneous detection. Labeled spots on the array are detected using a fluorimeter, indicating a signal indicating an antigen bound to a specific antibody.

[0087] Fluorophores are compounds or molecules that luminesce. Typically, fluorophores absorb electromagnetic energy at one wavelength and emit electromagnetic energy at a second wavelength. Representative fluorophores include, but are not limited to, 1,5 IAEDANS; 1,8-ANS; 4- Methyl umbelliferone; 5-carboxy-2,7-dichlorofluorescein; 5-Carboxyfl uorescein (5-F AM); 5-Carboxynapthofluorescein; 5-Carboxytetramethylrhodamine (5-TAMRA); 5-Hydroxy Tryptamine (5-HAT); 5-ROX (carboxy-X-rhodamine); 6-Carboxyrhodamine 6G; 6-CR6G; 6-JOE; 7-Amino-4-methylcoumarin; 7 -Aminoactinomycin D (7-AAD); 7-Hydroxy-4-methylcoumarin; 9-Amino-6-chloro-2-methoxyacridine (ACMA); ABQ; Acid Fuchsin; Acridine Orange; Acridine Red; Acridine Yellow; Acriflavin; Acriflavin Feulgen SITSA; Aequorin (Photoprotein); AFPs - AutoFluorescent Protein - (Quantum Biotechnologies) see sgGFP, sgBFP; Alexa Fluor 350™; Alexa Fluor 430™; Alexa Fluor 488™; Alexa Fluor 532™; Alexa Fluor 546™; Alexa Fluor 568™; Alexa Fluor 594™; Alexa Fluor 633™; Alexa Fluor 647™; Alexa Fluor 660™; Alexa Fluor 680™;Alizarin Complexon; Alizarin Red; Allophycocyanin (APC); AMC, AMCA-S;Aminomethylcoumarin (AMCA); AMCA-X; Aminoactinomycin D; Aminocoumarin; Anilin Blue; Anthrocyl stearate; APC-Cy7; APTRA-BTC; APTS; Astrazon Brilliant Red 4G;Astrazon Orange R; Astrazon Red 6B; Astrazon Yellow 7GLL; Atabrine; ATTO-TAG™ CBQCA; ATTO-TAG™ FQ; Auramine; Aurophosphine G; Aurophosphine; BAO 9 (Bisaminophenyloxadiazole); BCECF (high pH); BCECF (low pH); Berberine Sulphate; Beta Lactamase; BFP blue shifted GFP (Y 66H); Blue Fluorescent Protein; BFP / GFP FRET;Attorney Docket No. 11258-048WO125018-01 Bimane; Bisbenzemide; Bisbenzimide (Hoechst); bis- BTC; Blancophor FFG; Blancophor SV; BOBO™ -1; BOBO™-3; Bodipy492 / 515; Bodipy493 / 503; Bodipy 500 / 510; Bodipy; 505 / 515; Bodipy 530 / 550; Bodipy 542 / 563; Bodipy 558 / 568; Bodipy 564 / 570; Bodipy 576 / 589; Bodipy 581 / 591; Bodipy 630 / 650-X; Bodipy 650 / 665-X; Bodipy 665 / 676; Bodipy FI; Bodipy FL ATP; Bodipy Fl-Ceramide; Bodipy R6G SE; Bodipy TMR; Bodipy TMR-X conjugate; Bodipy TMR-X, SE; Bodipy TR; Bodipy TR ATP; Bodipy TR-X SE; BO-PRO™ -1; BO-PRO™ -3; Brilliant Sulphoflavin FF; BTC; BTC-5N; Calcein; Calcein Blue; Calcium Crimson™; Calcium Green; Calcium Green- 1 Ca2+Dye; Calcium Green-2 Ca2+; Calcium Green-5N Ca2+; Calcium Green-C18 Ca2+; Calcium Orange; Calcofluor White; Carboxy -X-rhodamine (5-ROX); Cascade Blue™; Cascade Yellow; Catecholamine; CCF2 (GeneBlazer); CFDA; CFP (Cyan Fluorescent Protein); CFP / YFP FRET; Chlorophyll; Chromomycin A; Chromomycin A; CL-NERF; CMFDA; Coelenterazine; Coelenterazine cp; Coelenterazine f; Coelenterazine fcp; Coelenterazine h; Coelenterazine hep; Coelenterazine ip; Coelenterazine n; Coelenterazine O; Coumarin Phalloidin; C-phycocyanine; CPM Methylcoumarin; CTC; CTC Formazan; Cy2™; Cy3.18; Cy3.5™; Cy3™; Cy5.18; Cy5.5™; Cy5™; Cy7™; Cyan GFP; cyclic AMP Fluorosensor (FiCRhR); Dabcyl; Dansyl; Dansyl Amine; Dansyl Cadaverine; Dansyl Chloride; Dansyl DHPE; Dansyl fluoride; DAPI; Dapoxyl; Dapoxyl 2; Dapoxyl 3’DCFDA; DCFH (Dichlorodihydrofluorescein Diacetate); DDAO; DHR (Dihydorhodamine 123); Di-4-ANEPPS; Di-8-ANEPPS (non-ratio); DiA (4-Di 16-ASP); Dichlorodihydrofluorescein Diacetate (DCFH); DiD- Lipophilic Tracer; DiD (DiIC18(5)); DIDS; Dihydorhodamine 123 (DHR); DiI (DiIC18(3)); Dinitrophenol; DiO (DiOC18(3)); DiR; DiR (DiIC18(7)); DM-NERF (high pH); DNP; Dopamine; DsRed; DTAF; DY-630- NHS; DY-635-NHS; EBFP; ECFP; EGFP; ELF 97; Eosin; Erythrosin; Erythrosin ITC; Ethidium Bromide; Ethidium homodimer-1 (EthD-1); Euchrysin; EukoLight; Europium (111) chloride; EYFP; Fast Blue; FDA; Feulgen (Pararosaniline); FIF (Formaldehyde Induced Fluorescence); FITC; Flazo Orange; Fluo-3; Fluo-4; Fluorescein (FITC); Fluorescein Diacetate; Fluoro-Emerald; Fluoro-Gold (Hydroxystilbamidine); Fluor-Ruby; FluorX; FM 1-43™; FM 4-46; Fura Red™ (high pH); Fura Red™ / Fluo-3; Fura-2; Fura-2 / BCECF; Genacryl Brilliant Red B; Genacryl Brilliant Yellow 10GF; Genacryl Pink 3G; Genacryl Yellow 5GF; GeneBlazer; (CCF2); GFP (S65T); GFP red shifted (rsGFP); GFP wild ty pe’ non-UV excitation (wtGFP); GFP wild type, UV excitation (wtGFP); GFPuv; Gloxalic Acid; Granular blue; Haematoporphyrin; Hoechst 33258; Hoechst 33342; Hoechst 34580; HPTS;Hydroxy coumarin; Hydroxystilbamidine (FluoroGold); Hydroxytryptamine; Indo-1, highAttorney Docket No. 11258-048WO125018-01 calcium; Indo-1 low calcium; Indodicarbocyanme (DiD); Indotricarbocyanine (DiR);Intrawhite Cf; JC-1; JOJO-1; JO-PRO-1; LaserPro; Laurodan; LDS 751 (DNA); LDS 751 (RNA); Leucophor PAF; Leucophor SF; Leucophor WS; Lissamine Rhodamine; Lissamine Rhodamine B; Calcein / Ethidium homodimer; LOLO-1; LO-PRO-1; Lucifer Yellow; Lyso Tracker Blue; Lyso Tracker Blue-White; Lyso Tracker Green; Lyso Tracker Red; Lyso Tracker Yellow; LysoSensor Blue; LysoSensor Green; LysoSensor Yellow / Blue; Mag Green; Magdala Red (Phloxin B); Mag-Fura Red; Mag-Fura-2; Mag-Fura-5; Mag-Indo-1;Magnesium Green; Magnesium Orange; Malachite Green; Marina Blue; Maxilon Brilliant Flavin 10 GFF; Maxilon Brilliant Flavin 8 GFF; Merocyanin; Methoxycoumarin; Mitotracker Green FM; Mitotracker Orange; Mitotracker Red; Mitramycin; Monobromobimane;Monobromobimane (mBBr-GSH); Monochlorobimane; MPS (Methyl Green Pyronine Stilbene); NBD; NBD Amine; Nile Red; Nitrobenzoxedidole; Noradrenaline; Nuclear Fast Red; Nuclear Yellow; Nylosan Brilliant Flavin E8G; Oregon Green™; Oregon Green™ 488; Oregon Green™ 500; Oregon Green™ 514; Pacific Blue; Pararosaniline (Feulgen); PBFI; PE-Cy5; PE-Cy7; PerCP; PerCP-Cy5.5; PE-TexasRed (Red 613); Phloxin B (Magdala Red); Phorwite AR; Phorwite BKL; Phorwite Rev; Phorwite RPA; Phosphine 3R; PhotoResist; Phycoerythrin B [PE]; Phycoerythrin R [PE]; PKH26 (Sigma); PKH67; PMIA; Pontochrome Blue Black; POPO-1; POPO-3; PO-PRO-1; PO-PRO-3; Primuline; Procion Yellow;Propidium iodide (PI); PyMPO; Pyrene; Pyronine; Pyronine B; Pyrozal Brilliant Flavin 7GF; QSY 7; Quinacrine Mustard; Resorufin; RH 414; Rhod-2; Rhodamine; Rhodamine 110; Rhodamine 123; Rhodamine 5 GLD; Rhodamine 6G; Rhodamine B; Rhodamine B 200; Rhodamine B extra; Rhodamine BB; Rhodamine BG; Rhodamine Green; Rhodamine Phallicidine; Rhodamine Phalloidine; Rhodamine Red; Rhodamine WT; Rose Bengal; R-phycocyanine; R-phycoerythrin (PE); rsGFP; S65A; S65C; S65L; S65T; Sapphire GFP; SBFI; Serotonin; Sevron Brilliant Red 2B; Sevron Brilliant Red 4G; Sevron Brilliant Red B; Sevron Orange; Sevron Y ellow L; sgBFP™ (super glow BFP); sgGFP™ (super glow GFP); SITS (Primuline; Stilbene Isothiosulphonic Acid); SNAFL calcein; SNAFL-1;SNAFL-2; SNARF calcein; SNARF1; Sodium Green; SpectrumAqua; SpectrumGreen; SpectrumOrange; Spectrum Red; SPQ (6-methoxy- N-(3 sulfopropyl) quinolinium); Stilbene; Sulphorhodamine B and C; Sulphorhodamine Extra; SYTO 11; SYTO 12; SYTO 13; SYTO 14; SYTO 15; SYTO 16; SYTO 17; SYTO 18; SYTO 20; SYTO 21; SYTO 22; SYTO 23; SYTO 24; SYTO 25; SYTO 40; SYTO 41; SYTO 42; SYTO 43; SYTO 44; SYTO 45;SYTO 59; SYTO 60; SYTO 61; SYTO 62; SYTO 63; SYTO 64; SYTO 80; SYTO 81;Attorney Docket No. 11258-048WO125018-01 SYTO 82; SYTO 83; SYTO 84; SYTO 85; SYTOX Blue; SYTOX Green; SYTOX Orange; Tetracycline; Tetramethylrhodamine (TRITC); Texas Red™; Texas Red-X™ conjugate; Thiadicarbocyanine (DiSC3); Thiazine Red R; Thiazole Orange; Thioflavin 5; Thioflavin S; Thioflavin TON; Thiolyte; Thiozole Orange; Tinopol CBS (Calcofluor White); TIER; TO-PRO-1; TO-PRO-3; TO-PRO-5; TOTO-1; TOTO-3; TriColor (PE-Cy5); TRITC TetramethylRhodamineIsoThioCyanate; True Blue; Tru Red; Ultralite; Uranine B; Uvitex SFC; wt GFP; WW 781; X-Rhodamine; XRITC; Xylene Orange; Y66F; Y66H; Y66W; Yellow GFP; YFP; YO-PRO-1; YO-PRO-3; YOYO-1; YOYO-3; Sybr Green; Thiazole orange (interchelating dyes); semiconductor nanoparticles such as quantum dots; or caged fluorophore (which can be activated with light or other electromagnetic energy source), or a combination thereof.

[0088] A modifier unit such as a radionuclide can be incorporated into or attached directly to any compounds described herein by halogenation. Examples of radionuclides useful in this embodiment include, but are not limited to, tritium, iodine- 125, iodine-131, iodine-123, iodine-124, astatine-210, carbon-11, carbon-14, nitrogen-13, fluorine-18. In another aspect, the radionuclide can be attached to a linking group or bound by a chelating group, which is then attached to the compound directly or by means of a linker. Examples of radionuclides useful in the aspect include, but are not limited to, Tc-99m, Re-186, Ga-68, Re- 188, Y-90, Sm-153, Bi-212, Cu-67, Cu-64, and Cu-62. Radiolabeling techniques such as these are routinely used in the radiopharmaceutical industry.

[0089] The detecting antibody (the antibody for the molecule of interest) or detecting molecule (the molecule that can be bound by an antibody to the molecule of interest) includes a label. Detection of the label indicates the presence of the detecting antibody or detecting molecule, which in turn indicates the presence of the molecule of interest or an antibody to the molecule of interest, respectively. Other labeling modes include the detection of primary immune complexes by a two-step approach. For example, a molecule (referred to as a first binding agent), such as an antibody, that has binding affinity for the molecule of interest or corresponding antibody can be used to form secondary immune complexes, as described above. After washing, the secondary immune complexes can be contacted with another molecule (which can be referred to as a second binding agent) that has binding affinity for the first binding agent, again under conditions effective and for a period of time sufficient to allow the formation of immune complexes (thus forming tertiary immune complexes). The second binding agent can be linked to a detectable label or signal-generating molecule orAttorney Docket No. 11258-048WO125018-01 moiety, allowing the detection of the tertiary immune complexes that are thus formed. This system can provide for signal amplification. Assays involving this detection of a substance, such as a protein or an antibody to a specific protein, include protein separation methods (i.e., electrophoresis). Protein separation methods are also useful for evaluating the protein's physical properties, such as size or net charge.

[0090] Provided that the concentrations are sufficient, the molecular complexes ([Ab- Ag]n) generated by antibody-antigen interaction are visible to the naked eye, but smaller amounts may also be detected and measured due to their ability to scatter a beam of light. The formation of complexes indicates that both reactants are present. In immunoprecipitation assays, a constant concentration of a reagent antibody is used to measure specific antigen ([Ab-Ag]n), and reagent antigens are used to detect specific antibody ([Ab-Ag]n). If the reagent species is previously coated onto cells (as in hemagglutination assay) or very small particles (as in latex agglutination assay), “clumping” of the coated particles is visible at much lower concentrations. Various assays based on these elementary principles are commonly used, including Ouchterlony immunodiffusion assay, rocket immunoelectrophoresis, and immunoturbidometric and nephelometric assays. The main limitations of such assays are restricted sensitivity (lower detection limits) in comparison to assays employing labels and, in some cases, the fact that very high concentrations of analyte can inhibit complex formation, necessitating safeguards that make the procedures more complex. Some of these Group 1 assays date back to the discovery of antibodies, and none have an actual “label” (e.g, Ag-enz). Other kinds of immunoassays that are label free depend on immunosensors, and a variety of instruments that can directly detect antibody-antigen interactions are now commercially available. Most rely on generating an evanescent wave on a sensor surface with immobilized ligand, allowing continuous binding monitoring to the ligand, Immunosensors allow the easy investigation of kinetic interactions and, with the advent of lower-cost specialized instruments, may find wide application in immunoanalysis in the future.

[0091] Radioimmune Precipitation Assay (RIP A) is a sensitive assay that uses radiolabeled antigens to detect specific antibodies in the serum. The antigens are allowed to react with the serum and then precipitated using a special reagent, such as, for example, protein A Sepharose beads. The bound radiolabeled immunoprecipitate is then commonly analyzed by gel electrophoresis. Radioimmunoprecipitation assay (RIP A) is often used as a confirmatory test for diagnosing the presence of HIV antibodies. RIPA is also referred to in the art as Farr Assay, Precipitin Assay, Radioimmune Precipitin Assay;Attorney Docket No. 11258-048WO125018-01 Radioimmunoprecipitation Analysis; Radioimmunoprecipitation Analysis, and Radioimmunoprecipitation Analysis.

[0092] Example System

[0093] Figs. 1A, IB, 1C, and ID each show an example automated computer vision tool system 100 (shown as 100a, 100b, 100c) for the automated identification and interpretation of a diagnostic analysis gel, such as serum protein electrophoresis (SPE), based on machine learning and artificial intelligence (Al), in accordance with an illustrative embodiment. The exemplary' system 100a may be used for the automated interpretation of Immunofixation Electrophoresis (IFE), among other electrophoresis methods described herein, e.g., for clinical laboratories for the diagnosis of myeloma, other disease, or blood cell cancer. The exemplary system and method can identify the presence of diagnostic proteins, antibodies, DNA, or RN A in a gel, which can require significant time, effort, and expertise from skilled medical technologists, pathologists or clinical chemists, to triage patients so that interpreters have more time to spend on difficult cases that are most pertinent to patient care. The automated computer vision tool system 100 may be used for prescreening or automated screening of patients in a clinical setting, research settings, or as a diagnostic-as-a-service solution.

[0094] Example #1. In the example shown in Fig, 1A, the system 100a includes a plurality of trained Al classifiers module 102 configured to generate a probability, score, or indication value 104 for the presence or non-presence of a gel band from an acquired gel image 106. In some embodiments, an individually trained Al classifier is employed for the classification / identification of each respective gel lane 108 in the gel image 106 or for a particular gel lane portion (e.g., protein). The trained Al classifier module 102 (shown as trained Al model 102’ in the training operation) may be trained using an Al training system 110 trained by a training data set 112 (shown as “Training Gel Images” 112) comprising, e.g., similar gel images as image 106 in format and data type, or different. The input gel image provided to the Al tram classifier may be processed using a saliency mask / occlusion computer vision analysis (shown as “Saliency Mask Processing” 114) to provide clear interpretability and transparency of the Al output.

[0095] As used herein, the terms “probability value,” “score value,” “indication value” are used interchangeably herein and refers to a direct or indirect output of the trained Al classifier. In some embodiments, the output of the trained Al classifier, as a probability value (e.g., between 0 and 1), score (e.g., between -1 and 1 or 0 and 1 or some other scale), orAttorney Docket No. 11258-048WO125018-01 indication value (e.g., 0 or 1, and such equivalence, including tiers), can be modified or derived through additional processing (e.g., normalization, comparison / thresholding, etc.).

[0096] Saliency Mask / Occlusion. To improve the interpretability and transparency of the trained Al classifier, the visualization of the gels may be modified via computer vision analysis to saliently emphasize the pertinent, relevant gel bands of interest by masking / occluding one or more bands in one or more gel lanes. The mask can be tuned (i) for a particular protein, antibody, RNA, or DNA, or (ii) for a set of bands for multiple proteins, antibodies, RNA, or DNA to provide an image frame of interest to the trained Al classifier 102 or the Al training system 110. In some other embodiments, rather than masking / occluding, the saliency operation may involve segmenting parts of the gel image of interest (e.g., one or more gel lanes or a portion of a gel lane) to provide segmented image portion of interest to the trained Al classifier 102 or Al training system 106. Occlusion and masking may be preferred as providing an image of the same size and dimension may be straightforward to implement,

[0097] Other alternative methods include SmoothGrad, Integrated Gradients, Shapiev Additive Explanations (SHAP), and GradCAM++ (e.g., which may work with CNNs).Occlusion sensitivity is a post-hoc analysis that may be applied to other Al model output.

[0098] Pre-processing. The trained Al classifier 102 and / or training system 110 may employ image pre-processing 116 (shown as “Quality Control Pre-s creen / Image processing” 116a and 116b). The pre-processing (e.g., 116a, 116b) may be the same for images as used for the trained Al classifier 102 (for run-time evaluation) and the training system 110 (for training). In other embodiments, the pre-processing (e.g., 116a, 116b) may be different for images used for the trained Al classifier 102 for run-time evaluation and for the training system 110 for training. In some embodiments, the Al model (e.g., 102’) may be trained by one group (e.g., a research / development group), and the trained model 102 is then provided to another group (e.g., electrophoresis manufacturer or software vendor) that incorporates the model 102, or a re-trained / adapted or modified version thereof, into an automated gel analysis software (e.g,, configured with densitometry, electrophoretograms, and image processing features), or other image analysis software.

[0099] In some embodiments, the pre-processing may include a quality' control evaluation, e.g., to alert and / or omit images of insufficient quality (e.g., poor scan, incorrect scan). In some embodiments, the pre-processing module (e.g,, 116), or a separate module (not shown), may perform image rotation and / or orientation determination / adjustment to align / center an input gel image prior to the subsequent analysis.Attorney Docket No. 11258-048WO125018-01

[0100] Post-processing. In the example shown in Fig. 1 A, the system 100a optionally includes an aggregation module 118 (shown as “Aggregation” 118).

[0101] Aggregation module 118 is configured to combine the output 104 of the trained Al classifier 10, which may include multiple classifier outputs 104 (shown as “P(Mi).. P(MN)”) for the number of trained Al classifiers employed for the classification / identification of each respective gel lane 108 and an output 104 (P(MGEL)) for a trained Al classifier employed for the classification / identification of the entire gel 107. Aggregation module 118 may employ rule-based logic for the combination. In some embodiments, the aggregation module 118 may compare one or more classifier outputs 104 (e.g., “P(Mi)... P(MN)”) against a threshold value to assess for the presence or non-presence of a disease or condition state that is then outputted 122 (shown as 122a) on a graphic user interface or in a report.

[0102] The visualization enhancement module (e.g., 114) is configured to (i) generate an enhanced image portion (adjust contrast, apply image correction) for a gel band of interest or (ii) generate a rendered visualization element (highlighting, text, color, etc.) to annotate or direct attention to a gel band. The annotation or rendered visualization element may be generated during the probability values, e.g., as compared to a threshold value. The rendered output 122 (shown as 122b) may also be output on the graphic user interface or in the report.

[0103] The output of the trained Al classifier, the aggregation module 118, and / or visualization enhancement module (e.g., 114) may be used for clinical or research purposes (e.g., to assist with the interpretation of the gel (e.g., 107), e.g., for the diagnosis of a disease or condition of interest. In some embodiments, the output can be used to trigger additional analyses, manual, automated, or otherwise, e.g., for additional processing of proteins, RNA, DNA, or molecules at the classified / identified gel bands.

[0104] Example #2. Fig. IB shows an example automated computer vision tool system 100 (shown as 100b) for an automated identification and interpretation of a diagnostic analysis gel with additional features and configurations that may be implemented in the training operation. The discussed features of Fig. IB may be implemented for a system show n in Fig. IB or implemented in combination with the features shown in systems of Fig.1A and / or Fig. 1C, among other figures described and shown herein.

[0105] In the example shown in Fig. lb, the system 100b includes an Al training system 110 (shown as 110’) comprising a set of fine-tuned Al classifier models (102a, 102b,..., 102n) and a coarse Al classifier model 102x. Of course, more than one coarse Al classifier model (e.g., 102x) (not shown) may be implemented, e.g., one for each pathology, disease, or condition to which a gel is being evaluated.Attorney Docket No. 11258-048WO125018-01

[0106] The output of the Al training system 110’ thus includes a corresponding number of fine-tuned Al classifier models (102a, 102b,..., 102n) and coarse Al classifier model 102x, as model 102’, which can be implemented in the trained Al classifier module 102 for the runtime application system 102b (or 102a, 102c, etc,).

[0107] Training Data Augmentation module. In some embodiments, the training system may employ a training data set augmentation module 124 configured to add noise or variations into the training data set 112 to produce additional training data set 125 (shown as 125a, 125b) for the Al training system 110’. In some embodiments, the noise and variations may be added as shapes (126a) of different randomly selected sizes and / or contrasts added to random positions in the gel images and / or gel lanes. In some embodiments, the noise and variations may be added as speckles (126b) of different or fixed sizes and / or contrasts added to random positions in the gel images and / or gel lanes. In some embodiments, the noise and variations may be added as duplicates or ghosts (126c) of an identified band in the gel images and / or gel lanes. In the example shown in Fig. IB, the augmentation module 124 may provide its outputs 125a to the saliency mask pre-processing module 114 (e.g., implemented as described in relation to Fig. 1A, 1C, etc.) or directly (shown as outputs 125b) to the Al training system 110’.

[0108] Example #3. Fig. 1C shows an example automated computer vision tool system 100 (shown as 100c) for an automated identification and interpretation of a diagnostic analysis gel with additional features and configurations that may be implemented in the runtime classification operation. The discussed features of Fig. 1 C may be implemented for a system shown in Fig. 1C or implemented in combination with the features shown in systems of Fig. 1A and / or Fig. IB, among other figures described and shown herein. The salient mask processing may be performed prior to the Al classification.

[0109] Additional image pre-processing / normalization. To provide consistent image features to the trained Al classifier 102 (or the Al training system 110, 110’; not shown), the image pre-processing module 116 (shown as 116a’) may provide a number of image preprocessing operations. In the example shown in Fig. 1C, the image pre-processing operations may include one or more of: a greyscale conversion operation 128a, a resize operation 128b, a denoise operation 128c, a contrast enhancement operation 128d, a tensor conversion 128e, a normalization operation 128f, among others. Additional example descriptions of these submodules are described herein, e.g., in relation to Fig. 4B, Other like operators may be used.Attorney Docket No. H 258-048WO125018-01

[0110] Example Saliency Operation / Shared. Fig. 1C shows an example of the saliency operation by generating / occluding a gel image having an occluded region 130b and a non¬ occluded region 130a. As shown, the same gel image (e.g., corresponding to images 106, 112) may be reproduced (shown as 132a, 132b,... 132n), having a respective gel lane occluded / masked for a particular gel band. In some embodiments, a portion of the gel lane of interest may also be occluded / masked (not shown, see Fig. 4G). The terms “occluded” and “masked” are used interchangeably herein, where occluded refers to having been obstructed or having an obstruction, and masked similarly refers to being covered or obstructed.

[0111] The occluded images (e.g., 132a, 132b,..., 132n) may be used to generate the saliency map (e.g., 122b). That is, the processing operation may generate an output that may¬ be later used in a subsequent process, e.g., for the output visualization. In some embodiments, the same occlusion / masking operation can be invoked by the visualization module and the pre-processing module.

[0112] In some embodiments, the saliency operation may also include an image processing operation to apply a filter or adjust portions of the non-occluded image 130a, e.g., for contrast, brightness, etc. The filtered or modified non-occluded portion of the portion (shown as “enhanced” gel bands 132 in Fig. 1C) may be used for the subsequent processing or training,

[0113] Example #4. Fig. ID shows an example automated computer vision tool system 100 (shown as lOOd) for automated identification and interpretation of a diagnostic analysis gel, with a validation module 140 configured to prompt user validations (also referred to as user interpretations) of the results (e.g., probability values 104) generated by the trained Al classifiers 102. The user validations (shown as validation results) are then used as ground truth to train the Al classifiers, improving their performance in generating probability values 104.

[0114] In Fig. ID, the validation module 140 is configured to receive (i) prescreened gel slide images from the image preprocessing operation 116a, and / or (ii) probability7values 104 from the trained Al classifiers 102, The validation module 140 is then configured to prompt, via a user interface 142 (see Figs. 3A - 3G), the users (e.g., pathologists, technologists, chemists, etc.) for their inputs (e.g., markings, annotations) on the received gel images and / or probability values 104. The validation module 140 is then configured to transmit the user validation inputs (shown as validation result) to the trained Al classifiers 102, or the training system 110, as ground truth to tram the trained Al classifiers 102, improving theirAttorney Docket No. 11258-048WO125018-01 performance in detecting bands in the gel image 106 and determining probability values 104 for the detected gel bands.

[0115] Example Trained Al Classifier Module

[0116] Figs. 2A, 2B, and 2C each show an example configuration 200 (shown as 200a, 200b, 200c) for a trained Al classifier module (e.g., 102, etc.) or a model in an Al train system (e.g., 110, 110’, etc.), e.g., as described in relation to Figs. IB. In Figs. 2A and 2B, the Al module 200a, 200b (as the trained Al classifier module (e.g., 102, etc.) or the Al training system 110 (e.g., 110, 110’)) comprises a set of fine-tuned Al classifier models (102a, 102b,..., 102n) and a coarse Al classifier model 102x.

[0117] In Fig. 2A, at least one Al model (e.g., 102a, 102b,..., 102n) is implemented, as a fine-tuning model, for each of the gel lanes 108 (shown as 108a, 108b,..., 108n), and at least one Al model (102x) is implemented, as a coarse model for the gel 107. The outputs 202, 204 (shown as 202a, 202b,... 202n, 204) are each shown as a probability value (e.g., P(Mi), P(M2),... P(MN)) for a presence and / or non-presence of a gel band of interest having a correspondence to (i) a protein, antibody, nucleic acid (D A, RNA), or other molecules of interest described or referenced herein or (ii) for a set of protein, antibody, nucleic acid (DNA, RNA), or molecule for an evaluation of the entire gel 107. The output may be provided to the aggregation module 118, e.g., to combine or modify the output for final output to the user, or directly provided to a visualizatiorn'reporting module that presents the output in a graphical user interface or a report.

[0118] Fig. 2B shows at least one Al model (e.g., 102a,..., I02n) being implemented, as a fine-tuning model, for some of the gel lanes 108 (shown as 108a,..., 108n), and at least one Al model (102x) is implemented, as a coarse model for the gel 107. The output is shown to output a probability value for the presence and / or non-presence of a gel band of interest having a correspondence to a protein, antibody, nucleic acid (DNA, RNA), or other molecule of interest described or referenced herein. Indeed, the example automated computer vision tool system does not need to operate on each and every gel lane m a given gel. The system may have two or more application modules, each configured to evaluate for a given disease or condition.

[0119] Fig. 2C shows another configuration for the trained Al classifier module (e.g., 102, etc.) in which output 204 of the coarse model 102x (shown as “Al model for gel” 102x) is employed in a decision module 206 that evaluates a gel image for a particular pathology, disease, or condition, e.g., by comparing the output score, probability, or indication value to a threshold value. The decision module 206 may compare the output 204 of the coarse modelAttorney Docket No. 11258-048WO125018-01 102x to a pre-defined threshold that indicates the presence (e.g., above the threshold) or non¬ presence (e.g., below the threshold) of protein, DNA, RNA, antibody, disease, or condition to which finer models 102a, 102b, 102n may then be employed to identify the specific protein, DNA, RNA, antibody or molecule in the gel image.

[0120] Example Validation Graphical User Interface (GUI) Module

[0121] To support the safe clinical deployment of the exemplary system, a validation graphical user interface (GUI) module (e.g., 140) is developed as an application that operationalizes both model outputs and human review. Rather than functioning as a retrospective validation environment, the validation module (e.g., 140) is configured as an integrated, front-end validation and verification layer embedded within the exemplary system. In some embodiments, the validation module (e.g., 140) facilitates structured postimplementation verification, discrepancy review, and longitudinal performance monitoring in routine laboratory use, embedding validation within real-world clinical operation. The role of the validation module is to facilitate initial clinical validation and to provide a continuous, integrated mechanism for automated verification and surveillance of model performance after implementation, consistent with lifecycle expectations for Al-enabled medical devices and supporting progression toward reduction-to-practice implementation.

[0122] In some embodiments, the validation module (e.g., 140) is configured as a digital pathology-style review environment in which a laboratory professional (e.g., a technologist, clinical chemist, or pathologist) reviews the scanned immunofixation electrophoresis (IFE) image. The interface (e.g., 142) of the module displays the full gel image and provides visualization adjustments, including zoom and image enhancement functions, to improve human review while preserving the original image data. In addition to displaying the raw image, the interface (e.g., 142) presents model-derived outputs, such as saliency map overlays and predicted band localizations, allowing direct comparison between human interpretation and Al output within the same operational environment. Users can assign a structured interpretation for each case, including negative, positive, or cannot interpret. For uninterpretable cases, the user can record the reason, such as poor scan quality, artifact, incomplete image capture, or other technical limitations. For positive cases, the user can annotate the precise locations of monoclonal bands and assign clonal relationships, lane identity, and confidence qualifiers such as confident, faint, or equivocal. These structured inputs generate a user-derived reference record that can be compared with the corresponding model output in real time.Attorney Docket No. 11258-048WO125018-01

[0123] The validation module (e.g., 140) can serve as a controlled, fully integrated review layer through which laboratories continuously assess the concordance between human interpretation and the deployed Al model as part of a routine workflow. In some embodiments, the validation module (e.g., 140) includes user authentication, role-based permissions, timestamps, software version logging, model version logging, scanner or acquisition metadata capture, and storage of user actions and adjudications. These elements can support (i) traceability, audit readiness, discrepancy investigation, and change control across the total product lifecycle, and (li) the incorporation of the validation module into the exemplary system as a core component rather than an auxiliary tool.

[0124] Clinically Operational Implementation. In some embodiments, a monitoring process of the validation module (e.g., 140) includes three review' streams as follows. First, a routine surveillance sample of cases can be selected at defined intervals (e.g., weekly, monthly) for human review' within the GUI (e.g., 142). These cases should be sampled to capture the clinical patient mix for those confidently predicted by the model. Enrichment of rare or complex but clinically important classes may also be appropriate, so long as the enriched sample is analyzed separately from unbiased surveillance samples.

[0125] Second, Al model-human discrepancies identified during routine clinical use may¬ be routed into the GUI (e.g,, 142) for secondary' review', a process overseen by director-level personnel of the clinical laboratory', including (i) cases w'here the model and laboratory interpretation disagree at the case level, lane level, or clone level, and (ii) cases with low-confidence or equivocal outputs. This discrepancy-focused approach is a high-yield component of monitoring because it identifies potential failure modes and supports targeted verification m the deployed environment.

[0126] Third, flagged reviews can occur after any relevant operational changes, such as (i) modifications / upgrades / repairs / replacements to gel electrophoresis equipment, scanner equipment, or software, (ii) image preprocessing changes, (lii) laboratory' information system (LIS) or middleware integration changes, (iv) staining or gel-material changes, or (v) shifts in patient population. This is important for image-based Al systems because analytical drift may arise from the model itself and from changes in image acquisition, formatting, or workflow. Integration of the review' triggers within the GUI (e.g., 142) can further reinforce the validation module’s role as a front-end validation and monitoring mechanism.

[0127] Comparison Process. In some embodiments, the user inputs received by the GUI (e.g., 142) can be compared against model outputs at several levels of granularity'. At the case level, agreement can be assessed for overall interpretation categories, such as positive,Attorney Docket No. 11258-048WO125018-01 negative, and cannot interpret. At the analytical level, the validation module (e.g., 140) can compare the identified heavy- and light-chain lanes, the number of clones detected, and the presence of faint or equivocal bands. At the spatial level, annotated pixel coordinates or user- defined band centers can be compared with model-localized band predictions, including saliency-derived regions, to assess localization concordance. A tiered comparison process is preferable because a case may be clinically concordant with the overall disease state while revealing a localization or subclassification error that is important for drift detection and longitudinal disease monitoring.[0128| In a clinical workflow, the comparison outputs can feed an analytics module that computes predefined performance indicators over time. Core indicators can include sensitivity, specificity, area under the receiver operating characteristic curve (AUROC), positive predictive value, negative predictive value, calibration with Brier score, Matthews Correlation Coefficient (MCC), and overall accuracy for case-level classification, along with lane-level agreement, clone-level agreement, and uninterpretable-case frequency. Cohen’s kappa or related agreement statistics may be useful for assessing Al model-human concordance, while trend charts can track stability overtime. Laboratory directors and / or vendors should define in advance which metrics are safety-critical and which thresholds should trigger investigation, escalation, or formal corrective action.

[0129] Monitoring Schedule. An implementation of the validation module (e.g., 140) can involve enhanced review during the early post-go-live phase, followed by steady-state surveillance. For example, an institution can review a large fraction of cases during the first several weeks after implementation, then transition to periodic sampling, in addition to reviewing flagged discrepancies and technically problematic cases. Additional intensified monitoring can be reintroduced after major changes. The exact schedule can depend on case volume, risk classification, and staffing, but the key principle is that monitoring should be ongoing, predefined, and integrated with the deployed exemplary system rather than performed only as a retrospective validation exercise. When discrepancies occur, the investigation should classify the cause as model error, image-quality degradation, workflow or accessioning mismatch, ambiguous biological pattern, human oversight, or another operational issue. Formal corrective actions to be taken after discrepancies occur can be aligned either with regulatory' standards or at the discretion of the laboratory director.

[0130] Graphical User Interface. Figs. 3A - 3G each show an example graphical user interface (e.g., 142) of the validation module (e.g., 140), in accordance with an illustrative embodiment. The validation module (e.g., 140) and associated interface (e.g., 142) areAttorney Docket No. 11258-048WO125018-01 configured for interpretations of electrophoresis gel scans by technologists, pathologists, clinical chemists, or other users. Outputs of the validation module (e.g., 140) can include case classification (e.g., positive, negative, faint, equivocal) and the pixel coordinates of monoclonal protein bands for positive cases, aligned by case identification (ID). These data can be reviewed by a laboratory director or other personnel for either model validation / verification or quality control and performance monitoring of the Al models in the exemplary' system. In some embodiments, error checking is built into the validation tool to help prevent accidental selections when reviewing each case.[01311 In Fig. 3 A, an example graphical user interface is shown that includes indication of the probability, score, or indication value 104 for the presence or non-presence of a gel band from an acquired gel image 106. The output may be presented with a confidence value of the probability, score, or indication value 104. The graphical user interface may be presented in combination with an interface for a validation operation. The output may indicate a result of positive, negative, or undetermined. In some embodiments, the output may indicate the lane that is indicating the presence of the monoclonal protein. In the example shown in Fig. 3A, the outputs of the Al classifiers (e.g., 102) on the user interface are “Final Interpretation: Positive’’ and “This case (Sample _1. png) contains a single clone of an IgA lambda monoclonal protein”.

[0132] In some embodiments, the GUI may include input for the manual annotation by a technician (e.g., interpretation 303, confidential level 301, notes 305).

[0133] In some embodiments, after the validation module starts, a window may appear on the user interface (e.g., 142) that prompts the user for their assigned ID (e.g., username) and password. In Fig. 3B, the user interface (e.g., 142) displays a full IFE slide image 302 of a clone, e.g., after the validation module (e.g., 140) verifies the user's ID and password. Users can review all lanes for the presence or absence of bands to interpret the slide image 302. A control panel 304 is configured to adjust the visualization of the image 302, including contrast, brightness, gamma, tone curve, and zoomness. The visualization adjustments do not alter the content of the original image 302 or affect the results of the electrophoresis gel scans. The verification allows user profiles to be associated with a validation operation. In some embodiments, the log in is performed for each of the analyses.

[0134] In Fig. 3C, an interpretation panel 306 on the user interface (e.g., 134) allows the users to (i) choose one of the interpretation options: “negative”, “positive”, or “cannot interpret”, and (ii) add user notes if needed. When the option “cannot interpret” is selected, the interpretation panel 306 provides a dropdown list, shown as “Reason”, for the users toAttorney Docket No. 11258-048WO125018-01 select a reason for the inability to interpret the image 302, including ‘-artifacts”, “poor quality gel”, “poor-quality scan”, “cropping issue”, or “other”.

[0135] In Fig. 3D, the user interface (e.g., 134) allows the users to mark (e.g., 308a, 308b) one or more bands within one or more lanes. A mark should be placed at the center of each marked band. In some embodiments, the coordinates of the marked band are shown in a coordinate panel 310. A band confidence level, such as “Confident, “Faint’, or “Equivocal”, can be selected for each band marking via panel 312, Using the panel 312, the users can also choose to review a new clone if an additional, distinct pattern is present on the same slide 302, or remove previous markings (or annotations).

[0136] In Fig. 3E, the “Save & Next” button 314 can be clicked to submit the interpretation and proceed to the next slide image, the “Previous” button 316 can be clicked to return to a previous slide image, the “Skip” button 318 can be clicked to skip the slide image if it cannot be reviewed at the time, the “Go to Case” button 320 can be clicked to review- or re-review a specific case number, and the “Jump to Next Unreviewed” 322 can be clicked to proceed to the next unreviewed slide image.

[0137] In Fig. 3F, after clicking the “Save & Next” button 314, a window 324 appears that summarizes the user selections for interpretation. The users can then review the selection and click either the “Confirm and continue” button 326 to save the selections for interpretation or the “Edit current scan” button 328 to edit the selections for interpretation. The users can then continue reviewing slide images 132 until all are reviewed. In some embodiments, ail interpretations and markings are saved to a .csv file upon submission and stored in a secure location accessible to the users (e.g., project manager).

[0138] In Fig. 3G, to enable faster review of the slide images 132, the user interface (e.g., 142) provides keyboard shortcuts, such as (i) up and down arrows to set overall interpretation, (ii) left and right arrows to cycle through the lanes, (iii) Ctrl and Up / Down arrow to select cycle band confidence, (iv) Shift and Enter to start a new clone, (v) Ctrl and Enter to save the interpretation and go to a next slide image (e.g., “Save & Next”), and (vi) Ctrl and Z to undo the last markings (e.g., annotations).

[0139] Example Pipeline / Operation Flow

[0140] Figs. 4A - 4G, respectively, show components of an example pipeline / operation flow-’ for an automated computer vision tool system, e.g., for detecting the presence of monoclonal proteins, and other molecules of interest as discussed herein. Specifically, Figs.4A - 4G show example operation flow, image segmentation operation, fine model training process, and saliency map generation process of the exemplary system, where Fig. 4A show sAttorney Docket No. 11258-048WO125018-01 an example AI model, Fig. 4B shows pre-processing operations, and Figs. 4C shows neural network architecture improvements to the Al model that may be implemented, Fig. 4D shows an example training operation with improved training using combined loss and optimization. Fig. 4E show's example data augmentation operations that may be performed, and Figs. 4F and 4G show the saliency map pre-processing operation to be used for the training. While the example is described in relation to an IFE gel scan, other gels and molecules of interest may be identified / classified using a trained Al model generated using a similar training and analysis operation having the saliency mask pre-processing.

[0141] Example Trained Al Classifier Module. Fig. 4A shows an example trained Al classifier module (e.g., 102) having fine and coarse Al models (e.g., 102a,... 102n, 102x) that may be used, e.g., to detect the presence of monoclonal proteins in IFE gel images.

[0142] In Fig. 4A, the operation flow employs a number of independent models (e.g., six shown in the example): a single ‘coarse’ model 102x (shown as 402) trained to detect the presence of monoclonal protein bands across the entire image and five additional ‘fine’ models 102a... 102n (shown as 404a, 404b, 404c, 404d, 404e) that are tuned to detect the presence of monoclonal protein bands within each of the five respective gel lanes (IgG, IgM, IgA, K, X). Additional or fewer numbers of models may be employed depending on the number of protein bands being evaluated, e.g., as described in relation to Figs. 2A and 2B.

[0143] In Fig. 4A, an input image (for training or evaluation) may be pre-processed via an image pre-processing (e.g., 116a, e.g., as described in relation to Fig. 1A), then evaluated by the coarse model 402. The output of the coarse model 402 is then used to invoke the fine models 404a... 404e, e.g., as described in relation to Figs. 2B. The operation could alternatively be performed as described in relation to Fig. 2A.

[0144] Example Pre-Processing Module. Fig. 4B shows an example image pre¬ processing module 116a (shown as 410) configured to transform and normalize before undergoing the subsequent analysis / classification. In Fig. 4B, each image (e.g., 106) can be read (412) from disk (e.g., data store), converted (414) to grayscale, and transformed (415) before the model training (e.g., 110) or evaluation (e.g., 102),

[0145] Resize (416). In the example shown in Fig. 4B, the transformation operation 415 includes a resize operation 416 that resizes the image according to a defined size (e.g., 400x400, 600x600, etc.).

[0146] Denoise (418). The transformation operation 415 includes a bilateral denoising operation 418 configured to smooth the image pixels and remove digital noise inAttorney Docket No. 11258-048WO125018-01 homogenous regions while preserving edges. An example of the denoising operation 418 includes a cv2.bilateralFilter function (cv2 version 4.7.0).

[0147] Contrast Enhancement (420). The transformation operation 415 includes a contrast enhancement operation 420. An example includes a Contrast-Limited Adaptive Histogram Equalization (CLAHE) that can improve local contrast in small patches while limiting overamplification of noise, e.g., using the cv2.createCLAHE function (cv2 version 4.7.0).

[0148] Tensor Conversion. The transformation operation 415 includes a tensor conversion operation 422 configured to emphasize bands in the gel image that correspond to monoclonal proteins, particularly highlighting faint bands and helping them to stand out against polyclonal background staining. Tire final output of the image pre-processing module 410 is a gel image optimized for immunofixation electrophoresis that can operate on a ResNet18 neural network model, e.g., that can extract and focus on relevant patterns that signify the presence of a monoclonal band and that enhance the explainability of model outputs and allow for the direct assessment of human-machine agreement. The ResNet18 neural network and other convolutional neural networks may be used. In some embodiments, the neural network includes a Convolutional Block Attention Module (CBAM).

[0149] Example Improved Al Architecture. Figs. 4C shows neural network architecture 420 with architecture improvements having a CBAM that may be implemented, e.g., for a pre-trained model ResNet18. Other residual neural network and deep learning models may¬ be implemented, e.g., transformers (Vision Transformer (ViT), Swin Transformer) and hybrid models consisting of CNN and a transformer (MaxViT). Other specific models made of convolutional neural networks as alternatives to ResNet18 include EfficientNet, DenseNet, Inception, MobileNet, and SqueezeNet, as well as other Al architectures discussed or referenced herein.

[0150] It is possible to use the model architecture of Fig. 4C to output a probability ‘heatmap’, rather than a single value. Images can be annotated with ground truth by experts (draw a circle or a rectangle around the areas of interest that result in a "positive’ sample), and those annotations can be used to train the model with Dice loss (or similar). Rather than image annotation (this takes a lot more time and work), a simple binary classification can be used along with flexible geometric regularization in the loss function. While annotation may not be beneficial for binary classification labels, annotation and heatmap probability outputs may be beneficial for heatmap-type outputs.Attorney Docket No. 11258-048WO125018-01

[0151] In the example shown in Fig. 4C, improvements are provided m the initial blocks 422, the intermediate blocks 424, and the classification blocks 426. For example, in the initial blocks 422, after the input images are resized and transformed, e.g., as described in relation to Fig. 4B or other figures herein, the resized and transformed image is input into the Al model and first undergo initial convolution, batch normalization (BatchNorm), a rectified linear unit (ReLU), and max pooling (MaxPool) (see 424a) to generate a feature map 424b. The initial convolution may have an adjustable kernel size (sizekernel), stride size (sizestride), and dilation (optional), e.g., to provide flexibility in the receptive field to help handle the identification of monoclonal protein bands in specific gel lanes.

[0152] The intermediate blocks 424, e.g., located after each of the four main ResNet18 residual blocks, includes a Convolutional Block Attention Module, CBAM, to provide two forms of attention at each stage, including: channel attention (which weights each feature map channel by importance) and spatial attention (which highlights or suppresses different (x,y) locations in the feature maps). By stacking the CBAM modules after each standard residual block, the model can refine which channels to emphasize and which spatial regions to focus on, which can provide a more effective specialized operator in detecting faint bands in gel scans.

[0153] The classification block 426 may be configured to have each model of the intermediate block 424 terminate with a single fully connected layer that then produces a onedimensional output for binary classification, which is then passed through a sigmoid function to yield a probability of “positive” vs. “negative” for a band corresponding to the presence of a monoclonal protein.

[0154] Example Improved Training Method for the Trained Al Classifier. Fig. 4D shows an example training operation 430 with improved training using combined loss 432 and optimization 434. In the example shown in Fig, 4D, during model training, a loss is computed with a unique combined loss function, including binary cross entropy with logit loss (432a), dice loss (432b), and flexible geometric regularization (432c). Binary crossentropy with logit loss (432a) can provide a measure of how well the model’s final logit aligns with the true label and penalizes incorrect positives and negatives. When the dice loss is computed (432b), correct positive predictions are favored by acting as a stand-in for an Fl score in the context of a single binary classification.

[0155] Using the intermediate feature maps 434, e.g., from blocks 424, as described in relation to Fig. 4C, the flexible geometric regularization (432c) is then applied to act as a ‘soft’ constraint on the shapes of features the model highlights, thus facilitating that theAttorney Docket No. 11258-048WO125018-01 model’s learned activation maps correspond to a region whose aspect ratio and width / height (relative to the image) match the expected band geometry. Based on expert domain knowledge, the expected aspect ratio of monoclonal protein bands can be set to between ‘T” and “3,” the width ratio between 4% and 9% of the image, and the height ratio between 2% and 12% of the image.

[0156] The flexible geometric regularization penalty can encourage the model to learn that ‘positive’ images must contain objects within a defined range of shapes and sizes characteristic of monoclonal proteins in IFE gels. This can be accomplished by collapsing the channel dimension as a single average activation heatmap per sample after each forward pass. For each sample in a batch, a threshold based on the mean is applied to the activation heatmap, so a mask of ‘activated’ and ‘non-activated’ pixels is created. The coordinates of all activated pixels are used to create a bounding box, from which the aspect ratio, width, and height (relative to the entire image) are found. If these values are outside the specified range, the difference from the boundary is squared and scaled by a weighting factor (λaspect, size). Once the penalties for all out-of-range conditions are summed per image, they are accumulated into a single scalar termed ‘geometric regularization loss’. The entire geometric penalty may then be averaged over the batch and across layers.

[0157] During optimization 434, e.g., at step 434a, a higher learning rate may be assigned to the later layers, and a reduced learning rate (scaled by a decay factor) may be applied to earlier layers. This is based on the idea that earlier layers, which have already learned feature representations (due to pre-training on the ImageNet database), should be fine-tuned more gently, whereas the later layers, which are more task-specific, can learn faster and become sensitive to the unique characteristics of IFE gel images, such as the specific shapes, intensities, and spatial arrangements of the monoclonal bands. At step 434b, weight decay within the Adam optimizer may be employed to help regularize the model by discouraging large weights, preventing overfitting to the training data. At step 434c, a step-based learning rate scheduler may be employed to reduce the learning rate as training progresses, allowing the model to make large updates initially to move into a good region of the loss landscape and then fine-tune the network with smaller, more precise updates. This fine-tuned optimization procedure refines the model’s ability7to focus on the nuanced features of monoclonal protein bands.

[0158] Example Data Augmentation to Improve Training. Fig. 4E shows an example image segmentation operation 440 on the images in the training data based on expert domainAttorney Docket No. 11258-048WO125018-01 knowledge. As shown in Fig. 4E, for each original image 442, six additional augmented images 444 (shown as 444a, 444b, 444c, 444d, 444e, and 444f) can be created.

[0159] Added Shapes. In augmented image 444a, objects, including ‘circles’ and ‘squiggles’, are shown added with equal probability, where each shape is placed at a random center, uniform within the width and height of the image. A fraction of the image's height determined each shape's size. The density of the shapes in each augmented image was determined by a single value (n = 10). Since the placement of each shape was random, no collision checks or overlap constraints were employed. Circles were drawn, filled with no boundary line, and translucent (a = 0.4) using the symbols function (graphics version 4.4.1). Squiggles were drawn as a random walk (in (x,y)) for a specified number of steps, with each step within [-Sizeshape / 2.... Sizeshape / 2], while being more opaque than the circles (a = 0.8) using the lines function (graphics version 4.4.1).

[0160] Added speckles. In augmented image 444b, speckles are shown added by passing a size of 1 to the character expansion (cex) argument of the points function (graphics version 4.4.1), which may be drawn with partial translucence (a = 0.5). The density of the speckles in each augmented image may be determined by a single value (n = 300). Since the placement of each speckle may be random, no collision checks or overlap constraints may need to be employed. In augmented image 444c, both shapes (circles and squiggles) and speckles are added to create a third augmented image.

[0161] In each of the augmented images 444e and 444f, the number of shapes and / or speckles may be shown to be doubled, creating a second set of augmented images for six augmented images per original image. The augmentation process sextupled the number of images in the training set. Test data may be obtained by scanning additional routine IFE gel films obtained at the clinical laboratory, which were left out of the model training process.

[0162] Example Saliency Module. Figs. 4F and 4G show saliency map processing operations that may be performed to provide enhanced visualization to supplement the output of the Al classifier, e.g., for the coarse model (see 449) and / or the fine model (see 450). The training data for the coarse model (shown as 449) shows the entire gel image being used.

[0163] Fig. 4F also shows an example fine model training with transfer learning and masking configured to focus on specific gel lanes. As shown, in the fine (lane-specific) model training, “masking” can be used to focus the network’s attention on a specific vertical region of the input that represents a 1 / 6thvertical slice of the image (the gel lane of interest). This mask can then be applied to the feature maps generated by the convolutional layers.Attorney Docket No. 11258-048WO125018-01 That is, the mask is applied to each of the gel lanes 452 (shown as 452a, 452b, 452c, 452d, 452e) to occlude non-band regions (454).

[0164] By applying a forward hook, the model can filter its own feature maps so that only the features from the vertical slice (gel lane) of interest contribute to the subsequent layers. This means only the activations from the region of interest can pass through unchanged, while activations from other areas can be suppressed. The fine models can use transfer learning by initializing their weights from a fully trained whole-image model, taking advantage of rich feature representations learned from the whole image, such as textures, edges, and patterns. The lane-specific training can refine these features to emphasize subtle localized patterns specific to gel lanes. For example, based on domain knowledge, IgG heavy chain bands may appear within the context of dense polyclonal background staining and may be more challenging to detect in digitized gel scans. IgM monoclonal protein bands, on the other hand, may be more difficult to distinguish from normal polyclonal IgM bands, as they may appear with a dense morphology, similar to other types of monoclonal protein bands. Fine models can be trained using the same Modified ResNet18 with CBAM model architecture and combined loss / optimization implementation as in the coarse model.

[0165] Saliency Mask Map Generation. Fig. 4G shows a process 460 to produce saliency maps for the fine Al model using refined occlusion sensitivity. As shown, when a monoclonal protein band is predicted to be present within a specific gel lane by classification, an occlusion sensitivity-based saliency map can be generated for the respective image, outlining the area of interest. Occlusion sensitivity can operate by turning off (i.e., masking) portions of the image and quantifying how this impacts the model’s output. When areas near monoclonal protein bands are masked, a drop in the model’s predicted probability of a positive sample can occur.

[0166] Saliency pixels within the operation flow of the exemplary system can be computed in a refined and iterative process. In step A (shown as 462) or I, the unmasked region of the image is split into two vertical halves. In steps B - G (shown as 464, 466, 468, 470, 472) or J - O, the masked half, which causes a decrease in model probability, can be further refined into additional masked and unmasked segments until a specified number of iterations is reached. In step H or P (474), once the refined occlusion sensitivity method is computed, the saliency pixels can be overlayed with either the original image (for visual inspection and interpretability) or saved as an image with a black background (for heavy / light chain combination assessment). The saliency maps with a black background can then be used to determine the exact monoclonal protein identity by assessing the degree of horizontalAttorney Docket No. 11258-048WO125018-01 overlap across pixels in heavy chain lanes and light chain lanes while ensuring interpretability by simple visual inspection by domain experts. The grayscale image can first be pre-processed by dividing the width evenly among the five lanes, as shown in Equation 1.WLanewidth= W / 5; W = lane width in pixels(Eq. 1)

[0167] Then, lane-specific saliency pixel intensity profiles can be extracted and compressed in the horizontal direction, as shown in Equation 2.For lane I at index i (with i — 0,...,4): — i x Lanewidth; xend= (i + 1) x Lanewidth(Eq. 2)

[0168] Saliency pixels can then be summed over the vertical axis to obtain a one-dimensional (1D) vector (ty®), as shown in Equation 3.v(i)(y) = Σ s(y, x), y = 0,..., H − 1 where S(y, x) = raw pixel intensity at positionx=xstaTt(V, X)(Eq. 3)

[0169] The overlap between gel lanes can then be computed. For each potential heavy / light chain pair, the element-wise minimum can be found, involving first finding the minimum of the two intensity profiles, as shown in Equation 4,m(y) = min(v(H)(y), v(L)(y)) (Eq. 4)

[0170] Next, the overlap across heavy / light chains can be summed to find the total intensity of each lane, as shown in Equation Set 5.overlap(H, U) = Xy=o ™(y) Hsum= Lsum= va)(y) (Eq. Set 5)

[0171] A final score can then be derived, as shown in Equation 6.Score(H, L) = (1 / 2)(m(y) / Hsum+ m(y) / Lsum) + λ · (overlap(H,L) / Hsum+ overlap(H,L) / Lsum) (Eq. 6)2 Hsum ^sum

[0172] Other methods of occluding / masking the gel images may be performed, e.g., in a sequential order (rather than the binary method). In some embodiments, the gel images may be segmented to include only gel lanes of interest.

[0173] Experimental Results and Additional Examples

[0174] A study was conducted to develop and evaluate an experimental computer vision tool system, also referred to as PROBE (Protein Recognition by Optical Band Extraction), for the automation of immunofixation electrophoresis (IFE) gel interpretations.Attorney Docket No. 11258-048WO125018-01

[0175] Experiment # 1[0176| PROBE was developed using archived pathology reports from routine serum protein electrophoresis at a clinical laboratory. Training data was augmented with expert domain knowledge to make the model robust against skewed predictions caused by visual artifacts commonly encountered in gel electrophoresis. PROBE was built on the ResNet18 backbone (pre-trained on ImageNet), with several modifications and enhancements tailored to the clinical IFE domain. PROBE was a machine learning pipeline that included a data preparation step, an image quality control check, and six independent classification models for the detection of monoclonal protein bands: a ‘coarse’ model to detect the presence of a monoclonal protein anywhere in the image and five additional ‘fine’ models that were tuned to detect the presence of monoclonal protein bands within each of the five respective gel lanes (e.g., IgG, IgM, IgA, K, A), and saliency map generation via post-hoc model analysis. Models output a probability corresponding to the likelihood of monoclonal protein presence in the entire image or in a specific gel lane.

[0177] PROBE used a combination of specific image transformations, a Convolutional Block Attention Module (CBAM), flexible geometric regularization, and occlusion sensitivity to detect and highlight monoclonal protein bands. Saliency maps were configured to predict monoclonal protein identity based on the degree of horizontal alignment of saliency map pixels across the heavy- and light-chain lanes of an IFE gel, as derived from model outputs. PROBE could output an interpretive report configured for review by clinical pathologists or clinicians that contained statements on the probability of the presence of monoclonal proteins and highlighted the regions of interest corresponding to the locations of monoclonal proteins within a gel. PROBE showed good predictive performance in test data and given the relatively ‘light-weight’ nature of the models, implementing PROBE on modest computer hardware was practical.

[0178] The study obtained and processed archived pathology reports from a clinical laboratory to use as training and validation data. Each archived report was saved as a PDF file, containing a representative image of the IFE scan for each respective sample and interpretive comments from routine pathology visual inspection. After removing equivocal samples, 4,061 images remained available for training and validation. Text in each report was processed using the pdf text function (e.g., PDF Tools version 3.4.0), and a partial string match was used to search for keywords to extract interpretative comments. These interpretive comments were used to assign ground truth outcome labels to each image (IgG monoclonal protein present, lambda monoclonal protein present, IgA K monoclonal protein present, etc.).Attorney Docket No. 11258-048WO125018-01 The IFE scan for each respective sample was cut from the PDF file and saved as a separate image, with a deidentified sample ID to link to the interpretive comments from the same reports. 70% of this data was reserved for training; the remainder was validation data.Training data was augmented according to expert domain knowledge to make the model robust against random artifacts encountered in IFE gels.

[0179] Training Results. Fig. 5 A shows the combined loss over epochs for training a coarse model and five fine models for detecting monoclonal protein bands. The coarse model was trained for five epochs, and each fine model was trained for two epochs. In test data (N = 906), AUROC was 0.984. At a threshold of 0.5, Fl, Cohen’s kappa, Matthew’s Correlation Coefficient (MCC), sensitivity, and specificity were 0.97, 0.95, 0.95, 0.97, and 0.98, respectively, as shown in Fig. 5B.

[0180] Classification Performance Results. Figs. 5C - 5F show the performance of the final models used in the exemplary system. For the IgG lane model in test data (N = 906), AUROC was 0.966. At a threshold of 0.5, Fl, Cohen’s kappa, Matthew’s Correlation Coefficient (MCC), sensitivity, and specificity were 0.90, 0.86, 0.87, 0.84, and 0.99, respectively (shown in Fig. 6C). For the IgA lane model in test data, AUROC was 0.961. At a threshold of 0.5, Fl, Cohen’s kappa, Matthew’s Correlation Coefficient (MCC), sensitivity, and specificity were 0.89, 0.88, 0.88, 0.85, and 1,00, respectively (shown in Fig. 6D), For the IgM lane model in test data, AUROC was 0.975. At a threshold of 0.5, Fl, Cohen’s kappa, Matthew’s Correlation Coefficient (MCC), sensitivity, and specificity were 0.90, 0.90, 0.90, 0.86, and 1,00, respectively (shown in Fig, 6E). For the K lane model in test data, AUROC w'as 0.980. At a threshold of 0.5, Fl, Cohen’s kappa, Mattliew'’s Correlation Coefficient (MCC, sensitivity, and specificity were 0.94, 0.92, 0.92, 0.89, and 1.00, respectively (shown in Fig. 6F). For the k lane model in test data, AUROC was 0.987. At a threshold of 0.5, Fl, Cohen’s kappa, Matthew'’ s Correlation Coefficient (MCC), sensitivity, and specificity were 0.90, 0.89, 0.89, 0.88, and 0.99, respectively (shown in Fig. 5G).

[0181] Using refined occlusion sensitivity'- based saliency maps generated by the fine models, the exemplary system can have enhanced interpretability compared to other modeling systems for monoclonal protein band identification. Fig. 5H shows the performance of the final model for identifying heavy / light chain combinations by saliency maps in test data. As shown, the final score fell between 0 and 1, where a value closer to 1 indicated a high degree of horizontal overlap between the heavy and light chain saliency pixels, suggesting a strong pairing. The score was used to classify monoclonal protein identity (e.g., IgG lambda) and provided the computation of classification performance metrics (e.g.,Attorney Docket No. 11258-048WO125018-01 sensitivity) compared to ground truth labels. In test data, sensitivity for identifying heavy / light chain pairs in samples, which were predicted to be positive for monoclonal proteins (N = 357), was 1.0 for all six possible pairs at a score threshold of 0.3 and 0.5, and was 1.0 at a score threshold of 0.7 for all possible pairs except for IgA-lambda (sensitivity = 0.88). Area-under-the-precision-recall-curve (PRAUC) was 0.991 (N = 163), 0.787 (N = 69), 0.777 (N - 37), 0.694 (N - 17), 0.843 (N - 21), and 0.807 (N = 17) for the IgG-K, IgG-λ, IgA-κ, IgA-λ,, IgM-K, and IgM-λ heavy / light chain combinations, respectively, for monoclonal protein identification.

[0182] Before images were transformed and fed into the classification model pipeline to detect monoclonal protein bands, an image processing procedure took place for quality control of gel scans. A Modified ResNetl 8 with CBAM model was trained to detect the presence of a clear serum protein (SP) lane on the left-most side of the image, ensuring that gels were scanned correctly and did not contain significant artifacts before monoclonal band detection to reduce the likelihood of false positives and false negatives. Gel scan images containing SP lanes from training data were used as ‘positive’ controls. Other images used in the ‘negative’ set included positive control images rotated 90 degrees, 180 degrees, flipped horizontally, and rotated 180 degrees, images missing the SP lane, blank images (white background), and images of noise (speckles and shapes, generated the same as in the band detection model training data but on a white background). This quality control model (e.g., Modified ResNetl 8 with CBAM) aimed to issue a quality control warning that prohibited the classification models from being executed when the scanned gel image did not contain an SP lane, was oriented incorrectly, or contained a significant amount of noise / artifacts. If the image did contain a clear SP lane and was oriented correctly, the quality control model facilitated the pipeline to proceed with monoclonal band detection and image classification.

[0183] Validation Results. Fig. 51 shows the combined loss during quality control model training for 5 epochs. Once an image passed the quality control check, the SP lane was cropped before the monoclonal protein band detection.

[0184] Experiment #2

[0185] Data Collection and Augmentation. Fig. 6A shows a training and evaluation process of PROBE in experiment #2. As shown, PROBE was trained and evaluated on IFE gels produced on a Helena SPIFE Touch gel electrophoresis system. Archived routine pathology' reports from a clinical laboratory' were obtained and used as training / validation data. Each archived report was saved as a PDF containing a representative image of the IFE scan for each sample and interpretive comments from the visual inspection. Reports wereAttorney Docket No. 11258-048WO125018-01 analyzed using custom R code (e.g., version 4.4.1). 4,656 total images were available for training / validation. Test data were obtained by rescanning additional routine IFE gel films at 600 dpi resolution using an Epson Perfection V19 II photo scanner. The test data were obtained at a clinical laboratory', which was excluded from the model's training. Test samples were linked to interpretive comments through the electronic medical record after routine pathology review. Three clinical pathologists were responsible for signing interpretations of test data.

[0186] Model Pipeline and Architecture. PROBE was configured to analyze gel IFE scans for the automated detection of monoclonal proteins. The PROBE pipeline included (i) a quality control model, and (ii) six independent models for monoclonal protein band detection that were combined in a single ensemble: a single ‘coarse’ model trained to detect the presence of monoclonal protein bands across the entire image, and five additional ‘fine’ models tuned to detect the presence of monoclonal protein bands within each of the five respective gel lanes (e.g., IgG, IgM, IgA, K, X) (see Fig, 4A). PROBE was built on the backbone of the pre-trained ResNet18 convolutional neural network (CNN) with several modifications and enhancements [13’]. Specifically, the CNNs within PROBE applied a Convolutional Block Attention Module (CBAM) and a custom loss function that implemented geometric regularization and confidence regularization [14’], [15’]. The CNN was a Modified ResNet18 with CBAM (see Fig. 7 A) and was developed using custom Python code (e.g., 3.13.0). Model training, inference, and data pipelines used PyTorch version 2.9.1, including torch.utils.data and optimization via torch.optim. The confidence regularization could prevent the model from calling overconfident false negatives when interpreting IFE gel scan images and more closely mimic how pathologists interpret IFE gels.

[0187] Image loading and processing relied on OpenCV 4.7.0 and the Python Imaging Library' (e.g., PIL / Pillow) version 11.0.0, including operations used during preprocessing and saliency overlay generation. Data handling and numerical computation used pandas 2.2.3 and NumPy 2.1.1, with visualization via Matplotlib 3.9.2. Model development and evaluation were performed on a Windows 11 desktop (e.g., AMD Ryzen Threadripper PRO 7995WX at 4.5 GHz, 128 GB DDR5 at 4800 MHz) with an RTX 6000 Ada Generation GPU with 48 GB of ECC GDDR6 VRAM.

[0188] In the fine (lane-specific) model training, fixed binary masking was used to focus the network’s attention on a specific vertical region of the input, corresponding to a 1 / 5thvertical slice of the image (the gel lane of interest) (see Fig. 4F). The binary mask was then applied to the feature maps generated by the convolutional layers, zeroing activation outsideAttorney Docket No. 11258-048WO125018-01 of the respective lane region. By applying a forward hook, the model filtered its own feature maps so that only features from the vertical slice (gel lane) of interest contributed to subsequent layers. This means that only activations from the region of interest could pass through unchanged, while those from other areas were suppressed. The fine models used transfer learning by initializing their weights from a fully trained coarse model, leveraging feature representations learned from the entire image. The lane-specific training then further refined these features to emphasize subtle, localized patterns specific to gel lanes. Fine models were configured to emphasize subtle, localized patterns observed in monoclonal proteins specific to certain gel lanes and were trained using the same Modified ResNet18 with CBAM architecture and combined loss / optimization implementation as in the coarse model. For example, based on expert domain knowledge, IgG heavy chain bands often appear within the context of dense polyclonal background staining and may be more difficult to detect in digital gel scans. IgM monoclonal protein bands, on the other hand, are often more difficult to distinguish from normal polyclonal IgM bands, as they often appear with a dense morphology'.

[0189] PROBE was configured with confidence regularization in mind and for enhanced interpretability compared to other modeling attempts at monoclonal protein band identification [9’- 12’], with the ability to potentially detect and characterize complex monoclonal protein patterns. This was achieved through the lane-specific nature of the ensemble and the use of refined occlusion sensitivity -based saliency maps generated by the fine models. When a monoclonal protein band was predicted to be present in a specific gel lane by classification, an occlusion-sensitivity-based saliency map was generated for the corresponding image, outlining the area of interest. Occlusion sensitivity masked (e.g., “turning off’) portions of the image and quantified how this impacted the model’s output (see Fig. 4G).

[0190] Parameter Tuning and Selection of Final Model Ensemble. During model training, hyperparameters, such as batch size, stride size, image input size, learning rate, and several parameters related to the combined loss function, were tuned based on performance on the validation dataset, rhe final model ensemble w'as selected after tuning and optimizing the coarse model and each fine model based on classification performance in the validation dataset via Cohen’s kappa. Results in the study were derived from the optimal models (e.g., coarse, IgG, IgA, IgM, K, A) based on validation data, Average loss over epochs for the coarse and fine models during training is shown in Fig. 5A, and average loss over epochs for the quality' control model is shown in Fig. 51.Attorney Docket No. 11258-048WO125018-01

[0191] Several hyperparameters were tuned during model development to maximize performance on the validation dataset. Building on the pre-trained ResNet18 CNN backbone, the study determined the optimal image input size, stride size, and number of epochs during training of the coarse model and the lane models. For all six models utilized in the final model ensemble, input size was set to 600 x 600 pixels. The stride sizes for (i) the coarse and models were 13, (li) the IgG, IgA, and IgM models were 9, and (iii) the K model was 5. Other hyperparameters were set to default values early in the development process, including: batch size to 16, learning rate to 5x10‘6, learning rate decay factor to 0.5, dilation to True, loss threshold for early stopping to 0.2, patience for early stopping to 1, and minimum delta for early stopping to 0.01.[0192| Several parameters related to the combined loss function, including parameters for flexible geometric regularization, were also tuned. For example, aspect ratio lambda was tuned to 0.2 (e.g., weight for penalizing aspect ratio violations), size lambda was tuned to 0.2 (e.g., weight for penalizing width / height size violations), aspect ratio range was tuned to 1.5:4.0 (e.g., acceptable width as a fraction of input image width), width ratio range was tuned to 0.051:0.083 (e.g., acceptable width as a fraction of input image width), and height ratio range was tuned to 0.03:0.12 (e.g., acceptable height as a fraction of input image height). Confidence regularization was achieved with the following weights:= 0.5, ^faint=0.5, ^def pos=0 •, and ^def neg=0.8. Additionally, iterative occlusion sensitivity-based saliency maps could be fine-tuned to find a balance between computational efficiency and accuracy.

[0193] Quality Control Model. Fig. 6B shows the training process of a quality control model. The quality control model was configured to screen images before monoclonal protein band detection and to help prevent false positives or negatives. The quality control model was configured to detect unprocessed IFE gel scans from the Helena SPIFE Touch system, including a serum protein lane, which was then cropped from the image before monoclonal protein band detection. The quality control model could help prevent incorrectly scanned images from progressing further in the pipeline. The quality control model was trained on images from the training dataset, where the original images were ground truth positives, and six ground truth negative versions were generated from each original image, including (i) image with no serum protein lane, (ii) cropped images with only a portion of the IFE gel present, (iii) image of random noise (e.g., speckles, shapes), (iv) blank image, (v) rotated image, and (vi) flipped image (see Fig. 6B, subpanels (a) - (f)).Attorney Docket No. 11258-048WO125018-01

[0194] Training, Validation, and Test Data. A total of 4,656 IFE gel scan images were used during model development, including cases labeled as either equivocal, faint, or definitive negative / positive (see Table 1). 52.4% of images were positive for monoclonal proteins (including equivocal / faint samples), the remainder negative. The majority of positive cases were of single clones (1,998), with the most common clones being either IgG-K or IgG-A, Of the 4,656 images, 601 (12.9%) were labeled as either equivocal or faint. Of the 2,002 test IFE gel scan images, 52.2% were positive for monoclonal proteins (including equivocal / faint samples), the remainder negative. Similarly, the majority' of positive cases were single clones (760), with the most common clones being either IgG-K or IgG-, 424 (21.2%) of test data images were labeled as either equivocal or faint. Three clinical pathologists signed clinical interpretations of test data during routine clinical service: 790 interpretations from pathologist #1, 741 interpretations from pathologist #2, and 470 interpretations from pathologist #3. Test images were derived from 1,152 patients with a mean age of 64.1 years; patients who were positive for monoclonal proteins were, on average, older (68.8 years). Patients were nearly evenly distributed between males and females (526 vs. 620, respectively). 371 (32.2%) patients were diagnosed with a monoclonal gammopathy (e.g., multiple myeloma, Waldenstrom Macroglobulinemia, smoldering myeloma, other monoclonal gammopathy). 283 (24.6%) patients had another diagnosis, such as anemia, chronic kidney disease (CKD), or acute kidney injury' (AKI), yvhile the remaining 498 (43.2%) of patients were other or unknown.Table 1. Train ing / validation and test dataset characteristics. Training / Validation Data (N = 4,656)Negative (N) Positive (N)2,217 (47.6%) 2439 (52.4%)Heay y chain only Light chain(N) only (N)32 59IgG (N) IgA(N) IgM (N) K (N) A (N)2,002 386 199 1,747 939IgG-K (N) IgG-X (N) IgA-K (N) IgA-7. (N) IgM-K (N) IgM- / . (N) 1.452 732 230 88 118 52 Number of clones per sample0 clones 1 clone 2 clones 3 clones 4 clones2,217 1,998 430 8 3Number of equivocal / faint samplesAttorney Docket No. 11258-048WO125018-01 601 (12.9%)Test Data (N = 2,002)Negative (N) Positive (N)958 (47.8%) 1,044 (52.2%)Heavy chain only Light chain(N) only (N)42 16IgG (N) IgA(N) IgM (N) K (N) X (N )492 117 68 426 225IgG-K (N) IgG-X (N) IgA-K (N) IgA-X (N) IgM-ic (N) IgM-X (N) 331 158 45 36 37 28 Number of clones per sample0 clones 1 clone 2 clones 3 clones 4 clones958 760 269 9 6Number of interpretations per pathologistPathologist 1 Pathologist 2 Pathologist 3790 741 470Number of equivocal / faint samples424 (21.2%)

[0195] Detection of Monoclonal Proteins and Quality Control Model. In Fig. 6C, subpanels (a) - (f) show the final model results in the test data for detecting monoclonal proteins in the entire image, in the IgG lane, in the IgA lane, in the IgM lane, in the K lane, and in thek ' lane, respectively. Table 2 shows the classification performance of identifying monoclonal proteins in non-equivocal (i.e., definitive) predicted samples in the test dataset. Only samples predicted as being positive for monoclonal proteins and containing single clones were analyzed.Table 2Validation Data (Non-equivocal predicted samples only; N = 454)TP TN FP FN AUROC PRAUCCoarse model 208 248 0 I 1.00 1.00 IgG model 156 284 3 14 0.98 0.98 IgA model 40 416 1 0 1.00 1.00 IgM model 11 446 0 0 1.00 1.00 K model 153 300 0 4 0.99 0.99X model 66 387 1 3 1.00 0.99Attorney Docket No. 11258-048WO125018-01 Test Data (Non-equivocal predicted samples only; N = 1,178)TP TN FP FN AUROC PRAUCCoarse model 485 676 12 4 0.99 0.99 IgG model 378 785 6 8 0.99 0.99 IgA model 71 1098 4 4 0.98 0.97 IgM model 49 1122 3 3 0.99 0.96 K model 336 818 17 6 1.00 0.99 X model 153 1003 12 9 0.98 0.95Validation Data (Non-equivocal predicted samples only; N = 454)Cohen’s Sensitivity Specificity MCC Fl scoreKappa Coarse model 1.00 1.00 1.00 1.00 1.00 IgG model 0.92 0.99 0.92 0.95 0.92 IgA model 1.00 1.00 0.99 0.99 0.99 IgM model 1.00 1.00 1.00 1.00 1.00 K model 0.97 1.00 0.98 0.98 0.98 X model 0.97 1.00 0.97 0.98 0.97 Test Data (Non-equivocal predicted samples only; N = 1,178)Cohen’s Sensitivity Specificity MCC Fl scoreKappa Coarse model 0.99 0.98 0.97 0.98 0.97 IgG model 0.98 0.99 0.97 0.98 0.97 IgA model 0.93 1.00 0.94 0.94 0.94 IgM model 0.94 0.99 0.94 0.94 0.94 K model 0.98 0.98 0.95 0.97 0.95 X model 0.95 0.99 0.93 0.94 0.93

[0196] In Fig. 6C and Table 2, ROC curves and additional classification performance metrics for the detection of monoclonal proteins in the test dataset indicated good overall performance. AUROC values were > 0.965, and the Matthews Correlation Coefficient (MCC) was > 0.85 for the coarse and fine models on all non-equivocal labeled test images (N = 1,578).

[0197] In Fig. 6D, subpanels (a) - (f) show the final model results, in the non-equivocal labeled samples for the validation data, for detecting monoclonal proteins in the entire image, in the IgG lane, in the IgA lane, in the IgM lane, in the K lane, and in the lane, respectively.Attorney Docket No. 11258-048WO125018-01 In Fig. 6D and Table 2, performance was also good for non-equivocal labeled samples in the validation data (N = 589), with AUROC values > 0.952 and Matthews Correlation Coefficient (MCC) > 0.83 at a threshold probability of 0.5. All six models, on average, achieved a sensitivity of 92% on the test dataset images.

[0198] Figs. 6E and 6F each show probability’ distributions for identifying monoclonal proteins in validation and test dataset images, respectively. In Figs. 6E and 6F, the majority of samples landed either in the definitive negative zone (left of the left dashed line) or in the definitive positive zone (right of the right dashed line).

[0199] In Fig. 6G, subpanels (a) and (b) show the performance results and probability distribution results, in the validation dataset of quality control images, of the quality control model, respectively. In Fig. 6G, the quality control model achieved an AUROC of 1.000.

[0200] Identification of Monoclonal Proteins by Heavy and Light Chain Pairing. Fig. 6H shows saliency maps generated by PROBE. In Fig. 6H, once monoclonal protein bands were identified and interpretable saliency maps were generated, the position of predicted bands within the gel could be identified. The saliency maps were then used to identify the monoclonal proteins by assessing the horizontal alignment of heavy and light chains. Of non-equivocal predicted samples with single clones only in the test dataset (N = 377), the two most common clones, IgG-K and IgG-X, each had an AUROC of 0.99 and an MCC of 0.95 and 0.94, respectively. Other single clones, such as IgA-X and IgM-X, were less abundant and showed lower classification performance, with AUROC values of 0.83 and 0.91, respectively.

[0201] Model Calibration and Assignment of Definitive and Equivocal Cases. Fig. 61 shows the distributions of the images in the validation and test datasets classified as definitive negatives, definitive positives, and equivocal. To prevent overconfident false positives and false negatives, probability thresholds of 0.2 and 0.8 were selected to define the lower and upper bounds for making definitive predictions, with all other samples landing in an equivocal zone. In the validation dataset, 42.1% of images were definitive negatives, 35.0% were definitive positives, and the remaining 22.9% were equivocal (see Fig. 61, subpanel (a)). In the test dataset, 43.0% of images were definitive negatives, 31.4% were definitive positives, and the remaining 25,5% were equivocal (see Fig, 61, subpanel (b)). In samples labeled as “equivocal” during routine pathology review, twice the number of images landed in PROBE’s predicted equivocal zone (e.g., 50.9%), with 12.0% as definitive negative and 37.0% as definitive positive (see Fig. 61, subpanel (c)).

[0202] Model Performance under Triaging Logic. A use case for the PROBE was to triage samples by screening for initial binning into negative, equivocal, or positive prior toAttorney Docket No. 11258-048WO125018-01 pathology review. Fig. 6J shows an operational flow of the triaging logic for the PROBE. In Fig. 6J, the triaging logic could prioritize sensitivity and allow PROBE to automate and standardize the determination of which samples were definitively negative, while flagging all other samples for further pathology review. Under this logic, only 3 of 2,002 samples in the test dataset were predicted as definitive false negatives for the presence of disease, achieving 99.5% sensitivity.

[0203] Fig. 6K shows 3 samples in the test dataset predicted as definitive false negatives for the presence of disease. In Fig. 6K, subpanel (a), the ground truth interpretive comment for the false negative was: “Serum immunofixation electrophoresis (IFE) shows atypical banding in the IgM fraction. The interpretation of which is complicated by the application point of the gel. Interval testing is recommended for clarification”. In Fig. 6K, subpanel (b), the ground truth interpretive comment for the false negative was: “Serum immunofixation electrophoresis shows an IgA heavy chain monoclonal protein”. In Fig. 6K, subpanel (c), the ground truth interpretive comment for the false negative was: “Serum immunofixation electrophoresis shows an IgM-K monoclonal protein”.

[0204] Computational Speed. With PROBE evaluated using an NVIDIA RTX 6000 Ada Generation GPU, the average computation time for binary classification across the coarse and fine models for a single image was 0.03 seconds. Explainable saliency map generation took 0.74 seconds for each positive predicted gel lane in a single image. In a representative workload scenario, assuming that 50% of samples were predicted as positive for monoclonal proteins, where 73% of positive cases were single clones, and the remaining 27% had multiple clones, the study anticipated that PROBE could require 16 minutes to interpret 1,000 IFE gel scans, including explainable saliency map outputs, on the same desktop hardware. The study could run PROBE on a less expensive NVIDIA GPU if the recommended minimum specifications of CUDA 11.3 or later and 8GB of VRAM or more were met (e.g., NVIDIA GeForce RTX 3050).

[0205] Model Architecture with Temperature Scaling. Fig. 7 A shows neural network architecture improvements to the Al model that could be implemented using temperature scaling. In Fig. 7A, subpanel (a), after images were resized and transformed, they were input into the model and underwent initial convolution, batch normalization (BatchNorm), a rectified linear unit (ReLU), and max pooling (MaxPool). The initial convolution had an adjustable kernel size (denoted as sizekernel), a stride size (denoted as sizestride), and a dilation, allowing flexibility in the receptive field to help identify monoclonal protein bands m specific gel lanes.Attorney Docket No. 11258-048WO125018-01

[0206] In Fig. 7A, subpanel (b), after each of the four main ResNet18 residual blocks in the intermediate block, CBAM introduced two forms of attention at each stage: channel attention (which applied weights to each feature map channel by importance) and spatial attention (which highlighted or suppressed different (x,y) locations in the feature maps) [26’]. By stacking the CBAM modules after each residual block, the model adaptively refined which channels it emphasized and which spatial regions it focused on, making it more effective at detecting faint bands in gel scans.

[0207] In Fig. 7A, subpanel (c), each model then terminated with a single fully connected layer that produced a one-dimensional output for binary classification, passed through temperature scaling and a sigmoid function to yield a probability of “positive” vs. “negative” for a band corresponding to the presence of a monoclonal protein.

[0208] Combined Loss Function and Temperature Scaling. Fig. 7B shows a combined loss function and optimization implementation for PROBE. During model training, the study implemented a combined loss function configured to couple accurate binary classification with probability calibration to discourage overconfident negative predictions, especially for equivocal or faintly positive patterns, including binary cross-entropy with logit loss, Dice loss, flexible geometric regularization, and confidence regularization (see Fig. 7B, subpanel (a)). Binary' cross-entropv with logit loss first measured how well the model’s final logit aligned with the true label and penalized incorrect positives and negatives (see Fig. 7B, subpanel (a)). Each image and each lane in each image had a soft target t E {0.0, 0.5, 0.7, 1.0} representing definitive negative, equivocal, faint positive, and definitive positive, respectively. For logits and probabilities pt= tr(ty) in a batch of N samples, the total loss for a given batch could be defined per Equation 7, where a, ft,Brier, and y are weights.^total d" frL[)ice+ Brier L Brier d” Y^Geo d~ ^evid(Eq. 7)

[0209] A binary' cross-entropy with logits function could be used on all soft labels, as shown in Equation 8.1NLBCE = 77 y [-t; log(pj - (1 - t;)log (1 - Pi)] / V —*E = 1(Eq. 8)

[0210] A Brier score term could be used to promote calibrated probabilities, as shown in Equation 9,Attorney Docket No. 11258-048WO125018-01 N^Brier T,'vj / (. Pi1 = 1(Eq- 9)

[0211] To emphasize overlap for hard positives and negatives, a Dice loss term was added to act as a stand-in for the Fl score in the context of a single binary classification, operating internally on cr(z,:) and restricted only to definitive labelsE {0.0, 1.0} (see Fig.7B, subpanel (b)), as shown in Equation 10.DiCeLeWPt+LeW^ + E(Eq. 10)

[0212] In Equation 10, Idef denotes the index set of samples with definitive labels (e.g., 7dey = {J:ti £ {04}}) and E is a small positive constant added for numerical stability. An asy mmetric evidence penalty could be applied to perform confidence regularization by discouraging overconfident negative probabilities (see Fig, 7B, subpanel (c)), as shown in Equation 11.1Nj-‘evid ~N ^-O.sCO d” Afaint(mi ) lo.?(O "b ^-defpos (mi Y 11(0 i-1^-def neg(j^i ') lo(O (Eq. 11)

[0213] In Equation 11, mt+= max (0, 0.5 — p() when a prediction is negative,= max (0, 0.2 — pi) when a prediction is close to zero, and the indicator functions l0.s (i), lo.?(O> li(0> 1Q(0 are f°rselecting equivocal= 0.5), faint positive (tf= 0.7), definitive positive (t = 1), and definitive negative labels (tj = 0), respectively. Thus, for tj E {0.5, 0.7,1}, any pt< 0.5 incurred a quadratic penalty that pushed the prediction away from negative, and for— 0.0, small ptbelow the soft floor of 0.2 were nudged upward, preventing overconfident negative predictions while leaving reasonably negative predictions intact.

[0214] Geometric regularization was also applied to intermediate feature maps F to encourage activation consistent with band-like geometry of monoclonal protein bands (within aspect ratio and size constraints based on expert domain knowledge), ensuring that the model’s learned activation maps corresponded to a region whose aspect ratio andAttorney Docket No. 11258-048WO125018-01 width / height (relative to the image) matched the expected band geometry (see Fig. 7B, subpanel (d)), as shown in Equation 12.Lgeo= G(F)(Eq. 12)

[0215] Finally, temperature scaling wras applied to rescale logits with a learned scalar, calibrating the predicted probabilities without changing discrimination performance (see Fig.7 A, subpanel (c)). Overall, this combined loss function promotes discriminative performance by emphasizing reducing false negatives, proper handling of equivocal / faint labels, improved calibration, and a focus on plausible feature geometry informed by expert domain knowledge.

[0216] Saliency Map Analysis. Saliency pixels within PROBE were computed iteratively. Masking areas near monoclonal protein bands led to a substantial drop in the model’s predicted probability of a positive sample. A score was then derived to determine the degree of horizontal overlap between two bands in different gel lanes. This score naturally ranged from 0 to 1, with values closer to 1 indicating a high degree of horizontal overlap between the heavy- and light-chain saliency pixels, suggesting a strong pairing. The score was used to classify monoclonal protein identify (e.g., IgG-X; IgA-K, etc.) and to compute classification performance metrics (e.g., specificity) relative to ground truth labels.

[0217] Fig. 7C shows a process to produce saliency maps for the fine Al model in PROBE using refined occlusion sensitivity. First, the unmasked region of the image is split into two vertical halves (see Fig. 7C, subpanel (a) or (i)). The masked half, which caused a greater decrease in model probability, was then further refined into additional masked and unmasked segments, and so on, until a specified number of iterations had been reached (see Fig. 7C, subpanels (b) - (g) or (j) - (o)). Once the refined occlusion sensitivity method had been computed, the saliency pixels could be overlayed with either the original image (e.g., for visual inspection and interpretability) or saved as an image with a blank background (e.g., for heavy / light chain combination assessment) (see Fig. 7C, subpanel (h) or (p)). The saliency maps with a blank background w7ere then used to determine the exact monoclonal protein identity by assessing the degree of horizontal overlap across pixels in heavy- and light-chain lanes, while ensuring interpretability.

[0218] The image was first pre-processed by dividing the width evenly among the 5 lanes, as shown in Equation 13, where W is the lane width in pixels.Lanewi(ifil-p~(Eq. 13)Attorney Docket No. 11258-048WO125018-01

[0219] Then, lane-specific saliency pixel intensity profiles were extracted and compressed in the horizontal direction, as shown in Equation 14.For lane I at index i (with i — 0,...,4): x^art— i x Lanewidth; x^d(i d-1) x Lanew dh(Eq. 14)

[0220] Saliency pixels were then summed over the vertical axis to obtain a ID vector (v®(y)), as shown in Equation 15, where S(y, xt') = raw pixel intensity at position (y, x).xend1S(y, Xj), y — 0,..., H — 1z=x(0“start(Eq. 15)

[0221] The overlap between gel lanes was then computed. For each potential heavy / light chain pair, the element-w ise minimum w?as computed by taking the minimum of the two intensity profiles, as shown in Equation 16.m(y) v< L)(y))(Eq. 16)

[0222] Next, the overlap between heavy / light chains was summed to obtain the total intensity for each lane, as shown in Equation Set 17.H-loverlap(H, L~) = m(y),Hsum =y^sum(Eq. 17)

[0223] A final score w as then derived per Equation 18.Ty=om(y) Ey=om(y) \ 1 overlap(H, L') overlap(H, L) Score(H, L) ~ ~^y^H)(y)+£vvW(y) / 2(77^+(Eq. 18)

[0224] Discussion

[0225] Interpretation of IFE gels requires significant effort from highly specialized experts in clinical pathology and / or laboratory medicine. It was observed, for example, thatAttorney Docket No. 11258-048WO125018-01 among more than 20 pathologists in the Department of Pathology and Laboratory Medicine at one school of medicine, only 3 clinical pathologists with the necessary esoteric experience are available to share on-call duty and sign out IFE interpretation reports.

[0226] Simple algorithmic software, such as that currently deployed by clinical electrophoresis manufacturers, can perform densitometry, generate electrophoretograms, and perform basic image processing, but is not capable of reliably identifying diagnostic proteins in IFE gels. Many IFE gels are considered ‘equivocal’, meaning that the diagnostic pattern is difficult to interpret. The complexify of IFE gel interpretation warrants the use of advanced computer vision techniques based on deep neural networks. The development and validation of Al computer vision techniques for diagnostic applications in the clinical laboratory require skilled labor from domain experts, such as data analysts and software engineers, working in close collaboration with pathologists and clinical chemists with relevant experience.Additionally, the development of advanced Al computer vision techniques requires substantial datasets, often comprising at least several thousand representative samples.

[0227] IFE gel instrumentation manufacturers have provided automated IFE gel analysis software, which is commonly paired with the instrumentation. However, software suites can often perform only basic tasks, such as image processing. No Al-based automated interpretation software has been developed and brought to the market for the purpose of identifying diagnostic proteins in IFE gel samples.

[0228] Additional Discussion. Serum protein electrophoresis (SPE) is a well-established routine laboratory' diagnostic technique used to physically separate and quantify' blood proteins, whose composition can be altered by various pathologies [1], IFE, a type of SPE, can help determine disease type, severity, and prognosis, and is useful for disease monitoring. In IFE, antisera are added to distinct gel lanes to identify antibody heavy chains (IgG, IgA, IgM, etc.) and light chains (K and / .). which correspond to specific monoclonal immunoglobulins and have clinical significance. After overlaying cellulose acetate membranes coated with anti-sera, the gel is incubated for approximately 30 minutes, allowing the anti-sera to bind with their specific target proteins. During this incubation, the formation of antigen-antibody complexes results in insoluble precipitates that later form visible bands [2], Monoclonal gammopathy, characterized by the presence of monoclonal immunoglobulins or light chains (paraproteins) in the blood or urine, arises from the clonal proliferation of mature B cells. It represents a spectrum of disorders ranging from benign conditions like monoclonal gammopathy of undetermined significance (MGUS) to malignant diseases such as multiple myeloma and Waldenstrom’s macrogio bulinemia. MGUS, the mostAttorney Docket No. 11258-048WO125018-01 common precursor state, is typically asymptomatic and requires ongoing surveillance due to its risk of progression to myeloma or related disorders [3, 4|. IFE is routinely used as a screening tool and is primarily indicated for patients suspected of having plasma cell dvscrasias / disorders or Waldenstrom Macroglobulinemia due to common comorbidities such as anemia, bone lesions, pathological fractures, hypercalcemia, renal dysfunction, or neuropathy [5],

[0229] Interpretation of IFE gels requires significant time and effort from highly skilled experts in the field of clinical pathology and / or laboratory medicine and is subject to inter¬ operator variability [6], Gel films (either digitized or physical) are visually inspected for the presence of bands that may represent monoclonal proteins and provide diagnostic evidence of malignant disease. Monoclonal proteins observed in IFE gels generally appear with a characteristic pattern but manifest with a variety of morphologies and intensities and are often obscured by artifacts that can either result in false positives or false negatives, making visual inspection non-trivial. The patient’s clinical history is also often taken into account during interpretation. Automation of IFE gel interpretation has significant potential benefits towards the reduction of turnaround time, labor costs, and the reduced rate of false negatives and false positives within the routine clinical laboratory environment. To date, no manufacturer of clinical protein electrophoresis equipment has brought automated interpretation of IFE gels to market. Although investigative efforts have been limited, Hu et al. have recently developed and published a qualitative model for predicting positive vs. negative (e.g., monoclonal protein present or absent) IFE gels [6], While their model had high accuracy, it lacked clear interpretability, which was showcased by the model’s inability to reliably highlight important regions for image classification. Our approach is unique in that the model architecture and hyperparameters have been tuned specifically to produce reliable saliency maps for the enhancement of clinical explainability, which will ensure that the model’s decisions are transparent and can be trusted by medical professionals. Our approach is also specifically designed to facilitate direct assessment of quantified human-machine agreement by visual inspection of model outputs by domain experts, which allows for ongoing monitoring of model performance and explainability of correct and incorrect predictions.

[0230] The exemplary system is configured for the automation of immunofixation electrophoresis (IFE) gel interpretations. The exemplary7system was developed using archived pathology' reports from routine serum protein electrophoresis at a clinical laboratory'. Training data was augmented based on expert domain knowledge to make the exemplary'Attorney Docket No. 11258-048WO125018-01 system robust against skewed predictions by visual artifacts encountered in gel electrophoresis. The exemplary system can be built upon the robust backbone of the ResNetI8 model (pre-trained on ImageNet) with modifications and enhancements configured for the specific domain of clinical IFE. The exemplar}7system is a machine learning pipeline / operation flow comprising a data preparation step, an image quality control check, and six independent classification models for the detection of monoclonal protein bands: a ‘coarse’ model to detect the presence of a monoclonal protein anywhere in the image and five additional ‘fine’ models that are tuned to detect the presence of monoclonal protein bands within each of the five respective gel lanes (e.g., IgG, IgM, IgA, K, ), and saliency map generation via post-hoc model analysis. Models output a probability corresponding to the likelihood of the presence of a monoclonal protein either in the entire image or relative to a specific gel lane. The exemplary system can use a combination of specific image transformations: a Convolutional Block Attention Module (CBAM), flexible geometric regularization, and occlusion sensitivity to detect and highlight monoclonal protein bands. Saliency maps are configured to predict the identity of monoclonal proteins via the degree of horizontal alignment of the saliency map pixels across the heavy chain lanes and light chain lanes of an IFE gel derived from model outputs. The exemplary' system can output an interpretive report that can be reviewed by clinical pathologists or clinicians, which contains statements about the probability of the presence of monoclonal proteins and highlights the regions of interest that may correspond to the locations of monoclonal proteins within a gel. The exemplary' system showed excellent predictive performance in test data, and given its lightweight nature, the exemplary system can be implemented on modest computer hardware.

[0231] Machine Learning. In addition to the machine learning operations described above, the trained Al classifier can be implemented using one or more artificial intelligence and machine learning operations. The term “artificial intelligence” can include any technique that enables one or more computing devices or computing systems (i.e., a machine) to mimic human intelligence. Artificial intelligence (Al) includes but is not limited to knowledge bases, machine learning, representation learning, and deep learning. The term “machine learning” is defined herein to be a subset of Al that enables a machine to acquire knowledge by extracting patterns from raw data. Machine learning techniques include, but are not limited to, logistic regression, support vector machines (SVMs), decision trees, Naive Bayes classifiers, and artificial neural networks. The term “representation learning” is defined herein to be a subset of machine learning that enables a machine to automatically discover representations needed for feature detection, prediction, or classification from raw data.Attorney Docket No. 11258-048WO125018-01 Representation learning techniques include, but are not limited to, autoencoders and embeddings. The term “deep learning” is defined herein to be a subset of machine learning that enables a machine to automatically discover representations needed for feature detection, prediction, classification, etc., using layers of processing. Deep learning techniques include, but are not limited to, artificial neural networks or multilayer perceptron (MLP).

[0232] Machine learning models include supervised, semi-supervised, and unsupervised learning models. In a supervised learning model, the model learns a function that maps an input (also known as a feature or features) to an output (also known as a target) during training with a labeled data set (or dataset). In an unsupervised learning model, the algorithm discovers patterns in the data. In a semi-supervised model, the model leams a function that maps an input (also known as a feature or features) to an output (also known as a target) during training with both labeled and unlabeled data.

[0233] Neural Networks. An artificial neural network (ANN) is a computing system including a plurality of interconnected neurons (e.g., also referred to as “nodes”). This disclosure contemplates that the nodes can be implemented using a computing device (e.g., a processing unit and memory as described herein). The nodes can be arranged in a plurality of layers, such as an input layer, an output layer, and optionally one or more hidden layers with different activation functions. An ANN having hidden layers can be referred to as a deep neural network or multilayer perceptron (MLP). Each node is connected to one or more other nodes in the ANN. For example, each layer is made of a plurality of nodes, where each node is connected to ail nodes in the previous layer. The nodes in a given layer are not interconnected with one another, i.e., the nodes in a given layer function independently of one another. As used herein, nodes m the input layer receive data from outside of the ANN, nodes in the hidden layer(s) modify the data between the input and output layers, and nodes in the output layer provide the results. Each node is configured to receive an input, implement an activation function (e.g., binary step, linear, sigmoid, tanh, or rectified linear unit (ReLU) function), and provide an output in accordance with the activation function. Additionally, each node is associated with a respective weight. ANNs are trained with a dataset to maximize or minimize an objective function. In some implementations, the objective function is a cost function, which is a measure of the ANN’S performance (e.g., error such as LI or L2 loss) during training, and the training algorithm tunes the node weights and / or bias to minimize the cost function. This disclosure contemplates that any algorithm that finds the maximum or minimum of the objective function can be used for training the ANN. Training algorithms for ANNs include, but are not limited to, backpropagation. It should beAttorney Docket No. 11258-048WO125018-01 understood that an artificial neural network is provided only as an example machine learning model. This disclosure contemplates that the machine learning model can be any supervised learning model, semi -supervised learning model, or unsupervised learning model. Optionally, the machine learning model is a deep learning model. Machine learning models are known in the art and are therefore not described in further detail herein.

[0234] A convolutional neural network (CNN) is a type of deep neural network that has been applied, for example, to image analysis applications. Unlike traditional neural networks, each layer in a CNN has a plurality' of nodes arranged in three dimensions (width, height, depth). CNNs can include different types of layers, e.g., convolutional, pooling, and fully-connected (also referred to herein as ‘‘dense’’) layers. A convolutional layer includes a set of filters and performs the bulk of the computations. A pooling layer is optionally inserted between convolutional layers to reduce the computational power and / or control overfitting (e.g., by downsampling). A fully-connected layer includes neurons, where each neuron is connected to all of the neurons in the previous layer. The layers are stacked similarly to traditional neural networks. GCNNs are CNNs that have been adapted to work on structured datasets such as graphs.

[0235] Other Supervised Learning Models. A logistic regression (LR) classifier is a supervised classification model that uses the logistic function to predict the probability of a target, which can be used for classification. LR classifiers are trained with a data set (also referred to herein as a “dataset”) to maximize or minimize an objective function, for example, a measure of the LR classifier’s performance (e.g., an error such as LI or L2 loss), during training. Ihis disclosure contemplates that any algorithm that finds the minimum of the cost function can be used. LR classifiers are known m the art and are therefore not described in further detail herein.

[0236] A Naive Bayes (NB) classifier is a supervised classification model that is based on Bayes’ Theorem, which assumes independence among features (i.e., the presence of one feature in a class is unrelated to the presence of any other features). NB classifiers are trained with a data set by computing the conditional probability distribution of each feature given a label and applying Bayes’ Theorem to compute the conditional probability distribution of a label given an observation. NB classifiers are known m the art and are therefore not described in further detail herein.

[0237] A k-NN classifier is an unsupervised classification model that classifies new data points based on similarity measures (e.g., distance functions). The k-NN classifiers are trained with a data set (also referred to herein as a “dataset”) to maximize or minimize aAttorney Docket No. 11258-048WO125018-01 measure of the k-NN classifier’s performance during training. This disclosure contemplates any algorithm that finds the maximum or minimum. The k-NN classifiers are known in the art and are therefore not described in further detail herein.

[0238] A majority voting ensemble is a meta-classifier that combines a plurality of machine learning classifiers for classification via majority voting. In other words, the majority voting ensemble’s final prediction (e.g., class label) is the one predicted most frequently by the member classification models. The majority voting ensembles are known in the art and are therefore not described in further detail herein.

[0239] Example Computing System

[0240] It should be appreciated that the logical operations described above for the automated computer vision tool system can be implemented (1) as a sequence of computer-implemented acts or program modules running on a computing system and / or (2) as interconnected machine logic circuits or circuit modules within the computing system. The implementation is a matter of choice, dependent on the performance and other requirements of the computing system. Accordingly, the logical operations described herein are referred to variously as state operations, acts, or modules. These operations, acts, and / or modules can be implemented in software, in firmware, in special purpose digital logic, in hardware, and any combination thereof. It should also be appreciated that more or fewer operations can be performed than shown in the figures and described herein. These operations can also be performed in a different order than those described herein.

[0241] The computer system is capable of executing the software components described herein for the exemplary method or systems. In an embodiment, the computing device may comprise two or more computers in communication with each other that collaborate to perform a task. For example, but not by way of limitation, an application may be partitioned in such a way as to permit concurrent and / or parallel processing of the instructions of the application. Alternatively, the data processed by the application may be partitioned in such a way as to permit concurrent and / or parallel processing of different portions of a data set by the two or more computers. In an embodiment, virtualization software may be employed by the computing device to provide the functionality of a number of servers that are not directly bound to the number of computers in the computing device. For example, virtualization software may provide twenty virtual servers on four physical computers. In an embodiment, the functionality disclosed above may be provided by executing the application and / or applications in a cloud computing environment. Cloud computing may comprise providing computing services via a network connection using dynamically scalable computingAttorney Docket No. 11258-048WO125018-01 resources. Cloud computing may be supported, at least in part, by virtualization software. A cloud computing environment may be established by an enterprise and / or can be hired on an as-needed basis from a third-party provider. Some cloud computing environments may comprise cloud computing resources owned and operated by the enterprise as well as cloud computing resources hired and / or leased from a third-party provider.

[0242] In its most basic configuration, a computing device includes at least one processing unit and system memory. Depending on the exact configuration and type of computing device, system memory may be volatile (such as random-access memory (RAM)), non-volatile (such as read-only memory (ROM), flash memory, etc.), or some combination of the two.

[0243] The processing unit may be a programmable processor that performs arithmetic and logic operations necessary for the operation of the computing device. While only one processing unit is shown, multiple processors may be present. As used herein, processing unit and processor refers to a physical hardware device that executes encoded instructions for performing functions on inputs and creating outputs, including, for example, but not limited to, microprocessors (MCUs), microcontrollers, graphical processing units (GPUs), and application-specific circuits (ASICs). Thus, while instructions may be discussed as executed by a processor, the instructions may be executed simultaneously, serially, or otherwise executed by one or multiple processors. The computing device may also include a bus or other communication mechanism for communicating information among various components of the computing device.

[0244] Computing devices may have additional features / functionality. For example, the computing device may include additional storage such as removable storage and non¬ removable storage including, but not limited to, magnetic or optical disks or tapes.Computing devices may also contain network connection(s) that allow the device to communicate with other devices, such as over the communication pathways described herein. The network connection(s) may take the form of modems, modem banks, Ethernet cards, universal serial bus (USB) interface cards, serial interfaces, token ring cards, fiber distributed data interface (FDDI) cards, wireless local area network (WLAN) cards, radio transceiver cards such as code division multiple access (CDMA), global system for mobile communications (GSM), long-term evolution (LTE), worldwide interoperability for microwave access (WiMAX), and / or other air interface protocol radio transceiver cards, and other well-known network devices. Computing devices may also have input device(s) such as keyboards, keypads, switches, dials, mice, trackballs, touch screens, voice recognizers, cardAttorney Docket No. 11258-048WO125018-01 readers, paper tape readers, or other well-known input devices. Output device(s) such as printers, video monitors, liquid crystal displays (LCDs), touch screen displays, displays, speakers, etc., may also be included. The additional devices may be connected to the bus in order to facilitate the communication of data among the components of the computing device.. Ml these devices are well known in the art and need not be discussed at length here.

[0245] The processing unit may be configured to execute program code encoded in tangible, computer-readable media. Tangible, computer-readable media refers to any media that is capable of providing data that causes the computing device (i.e., a machine) to operate in a particular fashion. Various computer-readable media may be utilized to provide instructions to the processing unit for execution. Example tangible, computer-readable media may include but is not limited to volatile media, non-volatile media, removable media, and non-removable media implemented in any method or technology for storage of information such as computer-readable instructions, data structures, program modules, or other data. System memory, removable storage, and non-removable storage are all examples of tangible computer storage media. Example tangible, computer-readable recording media include, but are not limited to, an integrated circuit (e.g., field-programmable gate array or applicationspecific IC), a hard disk, an optical disk, a magneto-optical disk, a floppy disk, a magnetic tape, a holographic storage medium, a solid-state device, RAM, ROM, electrically erasable program read-only memory (EEPROM), flash memory or other memory technology, CD-ROM, digital versatile disks (DVD) or other optical storage, magnetic cassettes, magnetic tape, magnetic disk storage or other magnetic storage devices.

[0246] In light of the above, it should be appreciated that many types of physical transformations take place m the computer architecture in order to store and execute the software components presented herein. It also should be appreciated that the computer architecture may include other types of computing devices, including hand-held computers, embedded computer systems, personal digital assistants, and other types of computing devices known to those skilled in the art.

[0247] In an example implementation, the processing unit may execute program code stored in the system memory'. For example, the bus may carry' data to the system memory, from which the processing unit receives and executes instructions. The data received by the system memory' may optionally be stored on the removable storage or the non-removable storage before or after execution by the processing unit,

[0248] ConclusionAttorney Docket No. 11258-048WO125018-01

[0249] The construction and arrangement of the systems and methods, as shown in the various implementations, are illustrative only. Although only a few implementations have been described in detail in this disclosure, many modifications are possible (e.g., variations in sizes, dimensions, structures, shapes, proportions of the various elements, values of parameters, mounting arrangements, use of materials, colors, orientations, etc.). For example, the position of elements may be reversed or otherwise varied, and the nature or number of discrete elements or positions may be altered or varied. Accordingly, all such modifications are intended to be included within the scope of the present disclosure. The order or sequence of any process or method steps may be varied or re-sequenced according to alternative implementations. Other substitutions, modifications, changes, and omissions may be made in the design, operating conditions, and arrangement of the implementations without departing from the scope of the present disclosure.

[0250] The present disclosure contemplates methods, systems, and program products on any machine-readable media for accomplishing various operations. The implementations of the present disclosure may be implemented using existing computer processors, or by a special purpose computer processor for an appropriate system, incorporated for this or another purpose, or by a hardwired system. Implementations within the scope of the present disclosure include program products, including machine-readable media for carrying or having machine-executable instructions or data structures stored thereon. Such machine-readable media can be any available media that can be accessed by a general-purpose or special-purpose computer or other machine with a processor. By way of example, such machine-readable media can comprise RAM, ROM, EPROM, EEPROM, CD-ROM or other optical disk storage, magnetic disk storage or other magnetic storage devices, or any other medium which can be used to carry or store desired program code in the form of machine¬ executable instructions or data structures, and which can be accessed by a general purpose or special purpose computer or other machine with a processor.

[0251] When information is transferred or provided over a network or another communications connection (either hardwired, wireless, or a combination of hardwired or wireless) to a machine, the machine properly views the connection as a machine-readable medium; thus, any such connection is properly termed a machine-readable medium.Combinations of the above are also included within the scope of machine-readable media. Machine-executable instructions include, for example, instructions and data that cause a general-purpose computer, special-purpose computer, or special-purpose processing machine to perform a certain function or group of functions.Attorney Docket No. 11258-048WO125018-01

[0252] Although the figures show a specific order of method steps, the order of the steps may differ from what is depicted. Also, two or more steps may be performed concurrently or with partial concurrence. Such variation will depend on the software and hardware systems chosen and on the designer's choice. All such variations are within the scope of the disclosure. Likewise, software implementations could be accomplished with programming techniques with rule-based logic and other logic to accomplish the various connection steps, processing steps, comparison steps, and decision steps.

[0253] As used in the specification and the appended claims, the singular forms “a,” “an” and “the” include plural referents unless the context clearly dictates otherwise. Ranges may be expressed herein as from “about” one particular value, and / or to “about” another particular value. When such a range is expressed, another implementation includes from the one particular value and / or to the other particular value. Similarly, when values are expressed as approximations, by use of the antecedent “about,” it will be understood that the particular value forms another i plementation. It will be further understood that the endpoints of each of the ranges are significant both in relation to the other endpoint, and independently of the other endpoint.

[0254] “Optional” or “optionally” means that the subsequently described event or circumstance may or may not occur and that the description includes instances where said event or circumstance occurs and instances where it does not.

[0255] Throughout the description and claims of this specification, the w ord “comprise” and variations of the word, such as “comprising” and “comprises,” means “including but not limited to,” and is not intended to exclude, for example, other additives, components, integers or steps. “Exemplary” means “an example of’ and is not intended to convey an indication of a preferred or ideal implementation. “Such as” is not used in a restrictive sense but for explanatory' purposes.

[0256] Disclosed are components that can be used to perform the disclosed methods and systems. These and other components are disclosed herein, and it is understood that when combinations, subsets, interactions, groups, etc. of these components are disclosed while specific reference of each various individual and collective combinations and permutation of these may not be explicitly disclosed, each is specifically contemplated and described herein, for all methods and systems. This applies to all aspects of this application, including, but not limited to, steps in disclosed methods. Thus, if there are a variety of additional steps that can be performed it is understood that each of these additional steps can be performed with any specific implementation or combination of implementations of the disclosed methodsAttorney Docket No. 11258-048WO125018-01

[0257] The following patents, applications, and publications, as listed below and throughout this document, are hereby incorporated by reference in their entirety herein. Reference List #1[1] O'Connell, T. X., T. J, Horita, and B, Kasravi, Understanding and interpreting serum protein electrophoresis. Am Fam Physician, 2005. 71(1): p. 105-12.[2] Csako, G., Immunofixation Electrophoresis for Identification of Proteins and Specific Antibodies. Methods Mol Biol, 2019. 1855: p. 177-201,[3] Alaggio, R., et al., The 5th edition of the World Health Organization classification of haematolymphoid tumours: lymphoid neoplasms. Leukemia, 2022. 36(7): p. 1720- 1748.[4] Katzmann, J. A., et al.. Elimination of the need for urine studies in the screening algorithm for monoclonal gammopathies by using serum immunofixation and. free light chain assays. Mayo Clin Proc, 2006. 81(12): p. 1575-8.[5] Castillo, J. J., Plasma Cell Disorders. Prim Care, 2016, 43(4): p. 677-691,[6] Hu, H., et al., Expert-Level Immunofixation Electrophoresis Image Recognition based on Explainable and Generalizable Deep Learning. Clin Chem, 2023. 69(2): p. HOBO.Reference List #2[ 1 '] O'Connell TX, Horita TJ, Kasravi B Understanding and interpreting serum protein electrophoresis. Am Fam Physician 2005;71:l:105-12.[2'] Csako G. Immunofixation electrophoresis for identification of proteins and specific antibodies. Methods Mol Biol 2019;1855:177-201 doi: 10.1007 / 978-1-4939-8793- 1 17.[3'] Alaggio R, Amador C, Anagnostopoulos I, Attygalle AD, Araujo IBdO, Berti E, et al.The 5th edition of the World Health Organization classification of haematolymphoid tumours: Lymphoid neoplasms. Leukemia 2022:36:7:1720-48.[4'] Katzmann J A, Dispenzieri A, Kyle RA, Snyder MR, Plevak MF, Larson DR, et al.Elimination of the need for urine studies in the screening algorithm for monoclonal gammopathies by using serum immunofixation and free light chain assays, Mayo Clin Proc 2006;81: 12: 1575-8 doi: 10.4065 / 81.12.1575.[5!] Castillo JJ. Plasma cell disorders. Prim Care 2016;43:4:677-91. Epub 20161014 doi:10.1016 / j.pop.2016.07.002.Attorney Docket No. 11258-048WO125018-01 [6’] Hu H, Xu W, Jiang T, Cheng Y, Tao X, Liu W, et al. Expert-level immunofixation electrophoresis image recognition based on explainable and generalizable deep learning. Clin Chem 2023:69:2:130-9 doi: 10.1093 / clinchem / hvacl90.[7'] Kyle RA, Sequence of testing for monoclonal gammopathies, Arch Pathol Lab Med 1999:123:2: 114-8 doi: 10.5858 / 1999-123-0114-SOTFMG.[8'] Keren DF. Procedures for the evaluation of monoclonal immunoglobulins. Arch Pathol Lab Med 1999;123:2: 126-32 doi: 10.5858 / 1999-123-0126-PFTEOM.[9’] Hu H, Xu W, Jiang T, Cheng Y, Tao X, Liu W, et al. Expert-level immunofixation electrophoresis image recognition based on explainable and generalizable deep learning. Clinical Chemistry 2022:69:2:130-9 doi: 10.1093 / clinchem / hvacl90.[10'] Wei XY. Yang ZQ, Zhang XL, Liao G, Sheng AL, Zhou SK, et al. Deep collocative learning for immunofixation electrophoresis image analysis. IEEE Transactions on Medical Imaging 2021:40:7:1898-910 doi: 10.1109 / TMI.2021.3068404.[IF] Thiemann C, Klitzke B, Martinetz P, Griming P, Kaster T, Barth E, et al. Automated assessment of immunofixations with deep neural networks. Journal of Laboratory Medicine 2022;46:5:331-6 doi: doi:10.1515 / labmed-2022-0078.[12'] Oztas B, Kosesoy I. Immunofixation electrophoresis image interpretation using transfer learning method. Clinica Chimica Acta 2025: 120726,[13'] He K, Zhang X, Ren S, Sun J. Deep residual learning for image recognition.Proceedings of the IEEE conference on computer vision and pattern recognition. 2016. p. 770-8.[14'] Woo S, Park J, Lee J-Y, Kweon IS. Cbam: Convolutional block attention module.Proceedings of the European conference on computer vision (ECCV). 2018. p. 3-19.[15'] Pereyra G, Tucker G, Chorowski J, Kaiser L, Hinton G. Regularizing neural networks by penalizing confident output distributions, arXiv preprint arXiv: 1701065482017.[16'] Litwin C, Anderson S, Phipps G, Martins T, Jaskow’ski T, Hill H. Comparison of capillary zone and immunosubtraction with agarose gel and immunofixation electrophoresis for detecting and identifying monoclonal gammopathies. American journal of clinical pathology 1999;! 12:3:411-7.[ 17'] McCudden CR, Mathews SP, Hainsworth SA, Chapman JF, Hammett-Stabler CA, Willis MS, et al. Performance comparison of capillary and agarose gel electrophoresis for the identification and characterization of monoclonal immunoglobulins. American journal of clinical pathology' 2008;129:3:451-8.Attorney Docket No. 11258-048WO125018-01 [18’] Kohlhagen M, Dasari S, Willrich M, Hetrick M, Netzel B, Dispenzien A, et al.Automation and validation of a maldi-tof ms (mass-fix) replacement of immunofixation electrophoresis in the clinical lab. Clinical Chemistry and Laboratory Medicine (CCLM) 2021;59: 1: 155-63.[19'] Fatica EM, Martinez M, Ladwig PM, Murray JD, Kohlhagen MC, Kyle RA, et al.Maldi-tof mass spectrometry can distinguish im unofixation bands of the same isotype as monoclonal or bicional proteins. Clinical Biochemistry 2021;97:67-73.[20’] Ihoren KL, McCash SI, Murata K. Immunotyping provides equivalent results to immunofixation in a population with a high prevalence of monoclonal gammopathies. The journal of applied laboratory medicine 2021;6:6:1551-60.[21'] Wilhite D, Arfa A, Cotter T, Savage NM, Bollag RJ, Singh G. Multiple myeloma:Detection of free monoclonal light chains by modified immunofixation electrophoresis with antisera against free light chains. Practical Laboratory Medicine 2021;27:e00256,[22’] Labcorp. Immunofixation (ife), serum and protein electrophoresis, serum.https: / / www.labcorp.eom / tests / 001495 / imniunofixation-ife-serum-and-protem- electrophoresis-serum (Accessed March 12, 2025).[23'] Laboratories A, Immunofixation electrophoresis, serum..https: / / ltd.aruplab.com / Tests / Pub / 2012572 (Accessed March 12, 2025).[24'] Kirchhoff DC, Murata K, Thoren KL. Use of a daratumumab-specific immunofixation assay to assess possible immunotherapy interference at a major cancer center: Our experience and recommendations. The journal of applied laboratory medicine 2021;6:6: 1476-83.[25'] Zuiderveld, K., Contrast limited adaptive histogram equalization, in Graphics gems IV.1994. p. 474-485.[26'] Woo, S., et al. CBAM: Convolutional block attention module, in Proceedings of the European conference on computer vision (ECCV). 2018.

Claims

Attorney Docket No. 11258-048WO125018-01 What is claimed:

1. A system comprising:a processor; anda memory having instructions stored thereon, wherein execution of the instructions causes the processor to:recei ve, via the processor, an electrophoresis gel image having one or more gel lanes;determine, via one or more trained Al classifiers, via the processor, one or more probability, score, or indication values each for a presence or non-presence of a respective band in a respective gel lane of the one or more gel lanes, wherein the one or more trained Al classifiers were generated from a set of electrophoresis gel training images; andoutput, via the processor, the one or more probability, score, or indication values, wherein the outputted probability, score, or indication values are used in a subsequent clinical or research for detection of a protein, antibody, nucleic acid, or disease or condition associated with a protein, antibody, lipid, or nucleic acid.

2. The system of claim 1, wherein the one or more trained Al classifiers comprises: a coarse Al model configured to detect one or more bands in the electrophoresis gel image; anda set of fine Al model configured to detect a band in the respective gel lane in the electrophoresis gel image.

3. The system of any one of claims 1-2, wherein the one or more trained Al classifiers each comprises a neural network.

4. The system of any one of claims 1-3, wherein the one or more trained Al classifiers each comprise a convolutional block attention model (CBAM) in each neural network.5 The system of any one of claims 1-4, wherein the one or more trained Al classifiers were generated using training data formatted as a tensor, and wherein the one or more trained Al classifiers operate on the electrophoresis gel image converted to a tensor.Attorney Docket No. 11258-048WO125018-01 6. The system of any one of claims 1-4, wherein the one or more trained Al classifiers were generated using augmented training data derived from a set of electrophoresis gel training images, wherein the augmented training data each includes at least one of: image artifacts, ghost, speckle, shaped image elements, or a combination thereof.I. The system of any one of claims 1-6, wherein the electrophoresis gel image was acquired using immunofixation electrophoresis for a set of antibodies.

8. The system of any one of claims 1-7, wherein the instructions are performed for a batch of electrophoresis gel images to identify one or more proteins, antibodies, nucleic acids, lipids, or disease or condition, for each respective electrophoresis gel image of the batch.

9. The system of any one of claims 1 -8, wherein the one or more gel lanes in the gel image include an antibody selected from the group consisting of IgG, IgM, IgA, K, and A.

10. The system of any one of claims 1-9, wherein the electrophoresis gel image was acquired using at least one of SDS-PAGE (Sodium Dodecyl Sulfate-Polyacrylamide Gel Electrophoresis), Native PAGE, Two-Dimensional (2D) Gel Electrophoresis, Blue Native PAGE (BN-PAGE), Clear Native PAGE (CN-PAGE), Isoelectric Focusing (IEF), Immunoelectrophoresis, Rocket Immunoelectrophoresis, Western Blotting (Immunoblotting), or agarose gel electrophoresis.I I. The system of any one of claims 1-10, wherein the execution of the instructions further causes the processor to:in response to the one or more probability, score, or indication values exceeding a predefined threshold value, generate saliency maps, each having the respective band in the respective gel lane of the one or more gel lanes; andoutput the generated saliency maps, wherein the outputted generated saliency maps are subsequently used for clinical or research for detection of a protein, antibody, nucleic acid, disease, or condition associated with a protein, antibody, lipid, or nucleic acid.

12. The system of any one of claims 1-11, wherein the generation of the saliency maps is an iterative process configured for automated interpretation of patterns in electrophoresis gel scans containing multiple monoclonal protein bands.Attorney Docket No. 11258-048WO125018-0113. The system of any one of claims 1-12, wherein the execution of the instructions, prior to determining the one or more probability, score, or indication values, further causes the processor to:determine, via the processor, a presence or non-presence of the one or more gel lanes in the received electrophoresis gel image;determine, via the processor, a rotation or croppness of the one or more gel lanes in the received electrophoresis gel image; andin response to the one or more gel lanes having the non-presence, the rotation, or the croppness, generate, via the processor, an alert indicating the non-presence, the rotation, or the croppness of the one or more gel lanes.

14. The system of any one of claims 1-13, wherein the execution of the instructions further causes the processor to:display, via a user interface, the received electrophoresis gel image;adjust, via the user interface, contrast, brightness, tone cave, and scale of the displayed received electrophoresis gel image;receive, via the user interface, a first user-interpretation value indicating a presence or non-presence of the one or more bands in the electrophoresis gel image;receive, via the user interface, a second user-interpretation value indicating the presence or non-presence of the one or more bands in respective gel lanes;receive, via the user interface, a third user-interpretation value indicating positions of the one or more bands in the respective gel lanes; andoutput, via the user interface, the received first, second, and third user-interpretation values, wherein the outputted received first, second, and third user-interpretation values are used for validating the determining of the one or more probability, score, or indication values for the presence or non-presence of bands in respective gel lanes.

15. The system of any one of claims 1-14, wherein the respective band corresponds to a monoclonal protein in the respective gel lane.

16. The system of any one of claims 1-15, wherein the user interface is a controlled review layer configured for user interpretation of immunofixation electrophoresis gel images as a digital pathology operation.Attorney Docket No. 11258-048WO125018-0117. The system of any one of claims 1-16, wherein the controlled review layer is used for assessing concordance between the user interpretation and the one or more trained Al classifiers, and supporting traceability', audit readiness, discrepancy investigation, and change control across a lifecycle of electrophoresis gel imaging.18, The system of any one of claims 1-17, wherein the one or more trained Al classifiers are subsequently used for analysis of protein electrophoresis, immunofixation electrophoresis, and clinical data.

19. The system of any one of claims 1-18, wherein the clinical data is selected from the group consisting of protein electrophoresis gel images, densitometry, electrophoretograms, total protein concentration, heavy-chain immunoglobulin concentration, free light-chain immunoglobulin concentration, medication history’, comorbidities, bone marrow aspirate cytology reports, imaging reports indicating lytic lesions in the bone, serum enzyme concentrations, serum calcium concentrations, and renal function tests, andwherein the clinical data is used in combination with an immunofixation electrophoresis gel image classifier, via the one or more trained Al classifiers, in a diagnostic system.20, A method comprising:receiving, via a processor, an electrophoresis gel image having one or more gel lanes; determining, via one or more trained Al classifiers executed by the processor, one or more probability7, score, or indication values each for a presence or non-presence of a respective band in a respective gel lane of the one or more gel lanes, wherein the one or more trained Al classifiers were generated using training data derived from a set of electrophoresis gel training images; andoutputting, via the processor, the one or more probability7, score, or indication values, wherein the outputted probability, score, or indication values are used in a subsequent clinical or research for detection of a protein, antibody, nucleic acid, lipid, or disease or condition associated with a protein, antibody, nucleic acid.

21. The method of claim 20, wherein the one or more trained Al classifiers comprises:Attorney Docket No. 11258-048WO125018-01 a coarse Al model configured to detect one or more bands in the electrophoresis gel image; anda set of fine Al models configured to detect a band in the respective gel lane in the electrophoresis gel image.

22. The method of any one of claims 20-21, wherein the one or more trained Al classifiers each comprises a neural network.

23. The method of any one of claims 20-22, wherein the one or more trained Al classifiers were generated using training data formatted as a tensor, and wherein the method comprises:converting the electrophoresis gel image to a tensor as input to the one or more trained Al classifiers.

24. The method of any one of claims 20-22, wherein the one or more trained Al classifiers were generated using augmented training data derived from a set of electrophoresis gel training images, wherein the augmented training data each includes at least one of: image artifacts, ghost, speckle, shaped image elements, or a combination thereof.

25. The method of any one of claims 20-24, wherein the electrophoresis gel image was acquired using immunofixation electrophoresis for a set of antibodies.

26. The method of any one of claims 20-25, wherein the electrophoresis gel image was produced by scanning with LED-based scanner illumination using a contact image sensor (CIS), an LED-based charge-coupled device (CCD), optical density measurements by a photodiode densitometer system, a LTV or blue light transillumination, or a complementary metal oxide semiconductor (CMOS) device.

27. The method of any one of claims 20-26, wherein the method is performed for a batch of electrophoresis gel images to identify one or more proteins, antibodies, lipids, nucleic acids, or diseases or conditions, for each respective electrophoresis gel image of the batch.

28. The method of any one of claims 20-27, wherein the one or more gel lanes in the gel image include an antibody selected from the group consisting of IgG, IgM, IgA, K, and A.Attorney Docket No. H 258-048WO125018-0129. A non-transitory computer-readable medium having instructions stored thereon, wherein execution of the instructions by a processor causes the processor to:receive, via the processor, an electrophoresis gel image having one or more gel lanes; determine, via one or more trained Al classifiers, via the processor, one or more probability, score, or indication values each for a presence or non-presence of a respective band in a respecti ve gel lane of the one or more gel lanes, wherein the one or more trained Al classifiers were generated from a set of electrophoresis gel training images; andoutput, via the processor, the one or more probability, score, or indication values, wherein the outputted probability, score, or indication values are used in a subsequent clinical or research for detection of a protein, antibody, nucleic acid, or disease or condition associated with a protein, antibody, lipid, or nucleic acid.

Images