The anti-HER2 monoclonal antibody drug trastuzumab (Herceptin®) was approved for treatment of HER2 positive breast cancer by the FDA in 1998. Concurrently, the HER2 IHC assay HercepTest™ was approved as an aid in the assessment of breast cancer patients where trastuzumab treatment was considered (review by Jørgensen et al. 2021). Recently, FDA approved antibody drug conjugate (ADC) trastuzumab deruxtecan (T-DXd; Enhertu®) for the treatment of metastatic HER2-positive and HER2-low breast cancer. Results from the Destiny Breast-06 clinical trial showed that T-Dxd provides statistically significant and clinically meaningful benefits to HER2-ultralow breast cancer patients as well. However, sensitive, accurate and quantitative evaluation of HER2 expression based on currently approved IHC assays remains challenging, especially in low and ultra-low ranges of HER2. Advanced computational approaches could improve the interpretation of such IHC assays and could be of high benefit for identifying treatment options for current and future HER2 targeted therapies. More importantly, it could provide a basis for objectively assessing the lower level of HER2 protein expression that maximizes therapeutic benefits, while minimizing drug exposure risks for patients. Here we demonstrate a novel AI-based method that quantifies HER2 protein expression following staining using the HercepTest™ mAB pharmDx (Dako Omnis, GE001) assay (HercepTest™ (mAB)) for breast cancer cases characterized as HER2 IHC 0 or 1+. We inferred HER2 expression spatially across the entire breast cancer tissue section and quantified its heterogeneity. We identified low levels of HER2 expression that are challenging to detect by the human eye and provided a graphical overlay of stained tissues for visualization and quantification of the heterogeneity of HER2 expression. We used the quantitative IHC (qIHC) method described by Jensen et al. (2017) to quantitively measure the HER2 expression in invasive breast carcinomas. These measurements were used to train an AI model to predict the expression based on the HercepTest™ (mAb) stain which was previously demonstrated to detect HER2 expression with higher sensitivity in the lower ranges of HER2 compared to PATHWAY 4B5 (Ruschoff et al, 2022). Note that quantitative analysis of the qIHC assay could also be used to directly quantify HER2 expression. 82 formalin fixed, paraffin embedded (FFPE) tissue blocks of invasive breast carcinoma with HER2 IHC scores 0 or 1+ and with areas of solid tumor tissue were selected for this study. Serial sections from each tissue block were stained with H&E, HercepTest™ (mAB), qIHC and p63 (GA662, Dako Omnis). Stained tissues were scanned on the Philips Ultra Fast Scanner and digitally aligned. Tumor areas were manually selected and reviewed by expert pathologists in HercepTest™ (mAB) using aligned H&E and p63 IHC staining information on consecutive tissue sections to help identify tumor areas and separate invasive tumor areas from carcinoma in situ. HER2 expression was evaluated based on the qIHC assay in each 128µm2 area within tumor regions. The differences in average qIHC expressions across tumor areas were found to be statistically significant between IHC 0, IHC>0 and IHC<1+, and IHC 1+ groups. We observed high level of spatial heterogeneity of the HER2 expression levels within the same tissue, up to five-fold in some cases. Using these qIHC based estimates of HER2 expression as ground truth, we trained an AI based interpretation of HercepTest™ (mAB) assay. K-fold cross validation scheme was used to separate train and test slides, with test folds aggregated for model evaluation. We demonstrated high slide-level agreement of the AI-based estimate of HER2 expression in tumor regions and the ground-truth with Pearson correlation of 0.94, and R2 of 0.87. In the future we expect AI-assisted quantification and visualization of HER2 expression to enable fast and safe treatment decisions. Citation Format: Anya Tsalenko, Frederik Aidt, Elad Arbel, Itay Remer, Oded Ben-David, Amir Ben-Dor, Daniela Rabkin, Kirsten Hoff, Karin Salomon, Sarit Aviel-Ronen, Jens Mollerup, Lars Jacobsen. AI Based Quantitative Estimation of HER2 Protein Expression in Low and Ultra Low Ranges [abstract]. In: Proceedings of the San Antonio Breast Cancer Symposium 2024; 2024 Dec 10-13; San Antonio, TX. Philadelphia (PA): AACR; Clin Cancer Res 2025;31(12 Suppl):Abstract nr P5-03-05.

Recent results of clinical trials in antibody drug conjugate (ADC) therapies have significantly broadened treatment options for the HER2 low and ultra-low breast cancer patients. However, sensitive, accurate and quantitative evaluation of HER2 expression based on current immunohistochemistry (IHC) assays remains challenging, especially in low and ultra-low HER2 expression ranges. We developed a novel methodology for quantifying HER2 protein expression, targeting breast cancer cases in the HER2 IHC 0 and 1+ categories. We measured HER2 expression using quantitative IHC (qIHC) that enables precise and tunable HER2 detection across different expression levels as demonstrated in formalin-fixed paraffin-embedded cell lines. Additionally, we developed an AI-based interpretation of HercepTest™ mAb pharmDx (Dako Omnis) (HercepTest™ mAb) using qIHC measurements as the ground truth. Both methodologies allowed spatial resolution and visualization of low and ultra-low levels of HER2 expression across entire tissue sections to demonstrate and enable quantification of heterogeneity of HER2 expression. Serial sections of 82 formalin-fixed paraffin-embedded tissue blocks of invasive breast carcinoma with HER2 IHC scores 0 or 1+ were stained with H&E, HercepTest™ (mAb), qIHC and p63, then scanned and digitally aligned. Tumor areas were manually selected and reviewed by expert pathologists. HER2 expression was quantitatively evaluated based on the qIHC assay in each 128x128μm2 area within tumor regions. We observed statistically significant differences in HER2 expression between IHC 0, 0 < IHC < 1+, and IHC 1+ groups, and a high degree of spatial heterogeneity of the HER2 expression levels within the same tissue, up to five-fold in some cases. We demonstrated high slide-level tumor region agreement of estimates of HER2 expression between the AI-based interpretation of HercepTest™ mAb and the qIHC ground truth with a Pearson correlation of 0.94, and R2 of 0.87. The developed methodologies can be used to stratify HER2 low-expression patient groups, potentially improving the interpretation of IHC assays and maximizing therapeutic benefits. This method can be implemented in histology labs without requiring a specialized workflow.

e13603 Background: Recent advancements in antibody drug conjugate therapies have significantly broadened treatment options for a substantial subgroup of breast cancer patients. However, sensitive, accurate and quantitative evaluation of HER2 expression based on current immunohistochemistry (IHC) assays remains challenging, especially in low and ultra-low HER2 expression ranges. Advanced computational approaches can improve the interpretation of such IHC assays and could be of high benefit for identifying the best treatment options for current and future HER2 targeted therapies. More importantly, it may provide a basis for objectively assessing the lower level of HER2 protein expression that maximizes therapeutic benefits, while minimizing drug exposure risks for patients. Methods: We developed a novel methodology for quantifying HER2 protein expression, targeting breast cancer cases in the HER2 IHC 0 and 1+ categories. We measured HER2 expression using a quantitative IHC (qIHC) assay (Jensen et al, Mod Path 2017) that enables precise and tunable HER2 detection across different expression levels as demonstrated in formalin-fixed paraffin-embedded (FFPE) cell lines . Additionally, we developed an AI-based interpretation of HercepTest™ mAb pharmDx (Dako Omnis) (HercepTest™ mAb) using qIHC measurements as the ground truth. Both methodologies allowed spatial resolution and visualization of low and ultra-low levels of HER2 expression across entire tissue sections to demonstrate and enable quantification of heterogeneity of HER2 expression. Serial sections of 82 FFPE patient tissue blocks of invasive breast carcinoma with HER2 IHC scores 0 or 1+ were stained with H&E, HercepTest™ mAb, qIHC and p63, then scanned and digitally aligned. Tumor areas were manually selected and reviewed by expert pathologists. HER2 expression was quantitatively evaluated based on the qIHC assay in each 128x128µm 2 area within tumor regions. Results: We observed statistically significant differences in HER2 expression between IHC 0, 0 < IHC < 1+, and IHC 1+ groups, and a high level of spatial heterogeneity of the HER2 expression levels within the same tissue, up to five-fold in some cases. We demonstrated high slide-level tumor region agreement of estimates of HER2 expression between the AI-based interpretation of HercepTest™ mAb and the qIHC ground truth with a Pearson correlation of 0.94, and R 2 of 0.87. Conclusions: The developed methodologies can be used to stratify HER2 low-expression patient groups, potentially improving the interpretation of IHC assays and maximizing therapeutic benefits. This method can be implemented in histology labs without requiring a specialized workflow. Disclaimer: This study is for proof of concept only. This study does not imply any clinical functionality, nor off-label use for any products mentioned.

In this paper, we introduce a novel deep-learning based method for virtual stain multiplexing of immunohistochemistry (IHC) stains. Traditional IHC techniques generally involve a single stain that highlights a single target protein, but this can be enriched with stain multiplexing. Our proposed method leverages sequential staining to train a model to virtually stain multiplex additional IHC on top of a digitally scanned whole slide image (WSI), without requiring a complex setup or any additional tissue sections and stains. To this end, we designed a novel model architecture, guided by the physical sequential staining process which provides superior performance. The model was optimized using a custom loss function that combines mean squared error (MSE) with semantic information, allowing the model to focus on learning the relevant differences between the input and ground truth. As an example application, we consider the problem of detecting macro-phages on PD-L1 IHC 22C3 pharmDx NSCLC WSIs. We demonstrated virtual stain multiplexing CD68 on top of PD-L1 22C3 pharmDx stained slides, which helps to detect macrophages and distinguish them from PD-L1+ tumor cells, which are often visually similar. Our pilot-study results showed significant improvement in a pathologist's ability to distinguish macrophages when using the virtually stain multiplexed CD68 decision supporting layer.

A recent line of works studied wide deep neural networks (DNNs) by approximating them as Gaussian Processes (GPs). A DNN trained with gradient flow was shown to map to a GP governed by the Neural Tangent Kernel (NTK), whereas earlier works showed that a DNN with an i.i.d. prior over its parameters maps to the so-called Neural Network Gaussian Process (NNGP). Here we consider a DNN training protocol, involving noise, weight decay and finite width, whose outcome corresponds to a certain non-Gaussian stochastic process. An analytical framework is then introduced to analyze this non-Gaussian process, whose deviation from a GP is controlled by the finite width. Our contribution is three-fold: (i) In the infinite width limit, we establish a correspondence between DNNs trained with noisy gradients and the NNGP, not the NTK. (ii) We provide a general analytical form for the finite width correction (FWC) for DNNs with arbitrary activation functions and depth and use it to predict the outputs of empirical finite networks with high accuracy. Analyzing the FWC behavior as a function of $n$, the training set size, we find that it is negligible for both the very small $n$ regime, and, surprisingly, for the large $n$ regime (where the GP error scales as $O(1/n)$). (iii) We flesh-out algebraically how these FWCs can improve the performance of finite convolutional neural networks (CNNs) relative to their GP counterparts on image classification tasks.

A recent line of works studied wide deep neural networks (DNNs) by approximating them as Gaussian Processes (GPs). A DNN trained with gradient flow was shown to map to a GP governed by the Neural Tangent Kernel (NTK), whereas earlier works showed that a DNN with an i.i.d. prior over its weights maps to the so-called Neural Network Gaussian Process (NNGP). Here we consider a DNN training protocol, involving noise, weight decay and finite width, whose outcome corresponds to a certain non-Gaussian stochastic process. An analytical framework is then introduced to analyze this non-Gaussian process, whose deviation from a GP is controlled by the finite width. Our contribution is three-fold: (i) In the infinite width limit, we establish a correspondence between DNNs trained with noisy gradients and the NNGP, not the NTK. (ii) We provide a general analytical form for the finite width correction (FWC) for DNNs with arbitrary activation functions and depth and use it to predict the outputs of empirical finite networks with high accuracy. Analyzing the FWC behavior as a function of $n$, the training set size, we find that it is negligible for both the very small $n$ regime, and, surprisingly, for the large $n$ regime (where the GP error scales as $O(1/n)$). (iii) We flesh out algebraically how these FWCs can improve the performance of finite convolutional neural networks (CNNs) relative to their GP counterparts on image classification tasks.

A fundamental question in deep learning concerns the role played by individual layers in a deep neural network (DNN) and the transferable properties of the data representations which they learn. To the extent that layers have clear roles, one should be able to optimize them separately using layer-wise loss functions. Such loss functions would describe what is the set of good data representations at each depth of the network and provide a target for layer-wise greedy optimization (LEGO). Here we derive a novel correspondence between Gaussian Processes and SGD trained deep neural networks. Leveraging this correspondence, we derive the Deep Gaussian Layer-wise loss functions (DGLs) which, we believe, are the first supervised layer-wise loss functions which are both explicit and competitive in terms of accuracy. Being highly structured and symmetric, the DGLs provide a promising analytic route to understanding the internal representations generated by DNNs.

The Yubileinaya pipe is situated in the central part of the Daldyn-Alakit kimberlite field of the Yakutian diamondiferous province. The pipe age determined by the zircon-based U-Pb method is 358.1 Ma. (Davis et al., 1980). The new data on the inclusions of K-rich melts (Zedgenizov et al, 1998), Cr-Ca-rich majoritic garnet coexisting with a Cr-Ca-rich non-majoritic garnet and olivine (Sobolev et al, 2002), Fe-olivines (Sobolev et al, 2000) in Yubileinayan diamonds have been recently obtained. This fact was stimulus to investigate possible melt or fluid inclusions in Yubileinayan diamonds. Fluid or melt inclusions are valuable sources of direct information regarding the chemical composition of fluids in the upper mantle. It is known that fibrous diamonds contain numerous micro (<1μm) inclusions consisting of carbonates, water and incompatible elements. Such microinclusions are found in diamonds from Zaire, Botswana and South Africa (Navon et al, 1988; Akagi and Masuda, 1988, Schrauder and Navon, 1994, Izraeli et al., 2001). About 1% of the Yubileinayan diamonds have shapes which may be described as cubes (fibrous). Thus the present study focuses on the composition and origin of microinclusions in fibrous diamonds.

The static friction coefficient between two materials is considered to be a material constant. We present experiments demonstrating that the ratio of shear to normal force needed to move contacting bodies can, instead, vary systematically with controllable changes in the external loading configuration. Large variations in both the friction coefficient and consequent stress drop are tightly linked to changes in the rupture dynamics of the rough interface separating the two bodies.

We present an experimental study of the onset of local frictional motion along a long, spatially extended interface that separates two PMMA blocks in dry frictional contact. At applied shear forces significantly below the static friction threshold, rapid precursory detachment fronts are excited, which propagate at near sound speeds along the interface. These fronts initiate from the interface edge and arrest prior to traversing the entire sample length. Along the fronts’ path, we perform real-time measurements of the real contact area at every spatial point within the interface. In addition, the motion (slip) of the material adjacent to the interface is simultaneously measured at chosen locations. Upon their arrival at each spatial point along their path, these fronts instantaneously (within 4 μs) reduce the net contact area. Net slip is only initiated after this contact area reduction occurs. Slip is initially rapid and progresses at its initial velocity for a constant (60 μs) duration. Slip dynamics then undergo a sharp transition to velocities an order of magnitude slower, which remain nearly constant until slip arrest. We demonstrate that this scenario can be quantitatively explained by a model of interface weakening caused by instantaneous fracture-induced heating. Sustained rapid slip occurs in this weakened phase. Once the interface cools beneath its glass temperature the sharp transition to slow slip takes place. A similar fracture-induced weakening scenario might be expected in additional classes of materials.

The minutiae of friction The behaviour of systems as diverse as earthquakes and hard drives is influenced by frictional motion and its strength. What at first glance appears to be a continuous sliding process between touching surfaces is in fact a product of a series of 'slip' and 'stick' events on the microscopic scale. The mechanism of evolution of frictional strength at this level, though, is still unclear. Ben-David et al . have studied the evolution of the local contact area between two sliding bodies (PMMA plastic blocks) and the motion of their interface, and find that it involves four distinct phases. Within microseconds, all the contact area reduction has occurred. This is followed by a rapid slip phase, then a sharp transition to much slower slippage culminating in a 'stick' phase when motion is arrested. After several hundred microseconds the contact area begins to increase again. These results provide a basis for a better understanding of this kind of motion in many technologically important contexts.

The way in which a frictional interface fails is critical to our fundamental understanding of failure processes in fields ranging from engineering to the study of earthquakes. Frictional motion is initiated by rupture fronts that propagate within the thin interface that separates two sheared bodies. By measuring the shear and normal stresses along the interface, together with the subsequent rapid real-contact-area dynamics, we find that the ratio of shear stress to normal stress can locally far exceed the static-friction coefficient without precipitating slip. Moreover, different modes of rupture selected by the system correspond to distinct regimes of the local stress ratio. These results indicate the key role of nonuniformity to frictional stability and dynamics with implications for the prediction, selection, and arrest of different modes of earthquakes.

Experiments of pure tensile fracture in thin brittle gels reveal a new dynamic oscillatory instability whose onset occurs at a critical velocity, VC=0.87CS, where CS is the shear wave speed. Until VC, crack dynamics are well described by linear elastic fracture mechanics (LEFM). These extreme speeds are obtained by suppression of the microbranching instability, which occurs when sample thicknesses are made comparable to the minimum microbranch width. The wavelength of these sinusoidal oscillations is independent of the sample dimensions, thereby suggesting that these macroscopic effects are due to an intrinsic microscopic scale that is unrelated to LEFM.

The excitation of large amplitude nonlinear waves is achieved via parametric autoresonance of Faraday waves. We experimentally demonstrate that phase locking to low amplitude driving can generate persistent high-amplitude growth of nonlinear waves in a dissipative system. The experiments presented are in excellent agreement with theory.

We present an experimental study of the dynamics of rapid tensile fracture in brittle amorphous materials. We first compare the dynamic behavior of “standard” brittle materials (e.g. glass) with the corresponding features observed in “model” materials, polyacrylamide gels, in which the relevant sound speeds can be reduced by 2–3 orders of magnitude. The results of this comparison indicate universality in many aspects of dynamic fracture in which these highly different types of materials exhibit identical behavior. Observed characteristic features include the existence of a critical velocity beyond which frustrated crack branching occurs and the profile of the micro‐branches formed. We then go on to examine the behavior of the leading edge of the propagating crack, when this 1D “crack front” is locally perturbed by either an externally introduced inclusion or, dynamically, by the generation of a micro‐branch. Comparison of the behavior of the excited fronts in both gels and in soda‐lime glass reveals that,...