Abstract Protein structure prediction via AlphaFold2 has revolutionized drug discovery, yet its end-to-end execution remains computationally intensive. While GPUs are traditionally favored for deep learning, the AlphaFold2 algorithm consists of heterogeneous phases — preprocessing with sparse database searches and model inference with low-arithmetic-intensity attention modules — that present unique architectural challenges. In this work, we address these bottlenecks by introducing Open-Omics-AlphaFold2, a highly optimized implementation for Intel ® Xeon ® CPU. By leveraging the CPU’s versatility in handling both sparse preprocessing algorithms and dense matrix operations via Intel Advanced Matrix Extensions (AMX), we accelerate the entire pipeline end-to-end. Our optimization strategy employs multi-level parallelism — spanning multiprocessing, multi-threading, and vectorization — alongside cacheaware tiling and operator fusion. Our results demonstrate that, on a Xeon CPU, Open-Omics-AlphaFold2 achieves 2 7.58 speedup for preprocessing and 19.8 29.2 speedup for model inference over baseline Deepmind-AlphaFold2. Moreover, for a proteome of 391 proteins, Open-Omics-AlphaFold2 running on a dual-socket Intel Xeon 6980P system achieves a remarkable 76% higher through-put over the state-of-the-art GPU-accelerated solution, FastFold, running on a single-socket Intel Xeon 6980P CPU with an NVIDIA H100 offioad. Code availability Baremetal: https://github.com/IntelLabs/open-omics-alphafold Containerized: https://github.com/IntelLabs/Open-Omics-Accelera tion-Framework/tree/main/pipelines/alphafold2-based-protein-folding

Protein-protein interactions (PPIs) are mediated at the residue level. Most sequence-based PPI models consider residue-residue interactions across two proteins, which can yield accurate interaction scores but are too slow to scale. At proteome scale, identifying candidate PPIs requires evaluating nearly *all possible protein pairs*. For $N$ proteins of average length $L$, exhaustive all-against-all search requires $\mathcal{O}(N^2L^2)$ computation, rendering conventional approaches computationally impractical. We introduce RaftPPI, a scalable framework that approximates residue-level PPI modeling while enabling efficient large-scale retrieval. RaftPPI represents residue interactions with a Gaussian kernel, approximated efficiently via structured random Fourier features, and applies a low-rank factorized attention mechanism that admits pooling into a compact embedding per protein. Each protein is encoded once into an indexable embedding, allowing approximate nearest-neighbor search to replace exhaustive pairwise scoring, reducing proteome-wide retrieval from *months* to *minutes* on a single GPU. On the human proteome with the D-SCRIPT dataset, RaftPPI retrieves the top 20\% candidate pairs ($\sim$200M) in 6 GPU minutes, covering 75.1\% of the true interacting pairs, compared to 4.9 GPU months for the best prior method (61.2\%). Across seven benchmarks with sequence- and degree-controlled splits, RaftPPI achieves state-of-the-art PPI classification and retrieval performance, while enabling residue-aware, retrieval-friendly screening at proteome scale.

As super-resolution techniques continue to evolve, there is a growing requirement for more advanced methods to capture finer details, particularly when dealing with the smaller pixels within an image. In remote sensing, enhanced spatial details can find utility in diverse applications, such as disaster management, urban planning, and environmental change detection. Many existing image super-resolution algorithms are there to improve image resolution. However, they are not explicitly crafted to accommodate the distinctive attributes of remote-sensing images, rendering them less effective in restoring the details of the images. Therefore, we proposed a convolutional block attention residual network with joint adversarial mechanisms (CRNJAM) to capture finer details in remote sensing images. We first designed a generator based on the residual network and attention mechanism. This has the ability to produce high-quality images with superior resolution, even when the input is of low quality. Then, we train the super-resolved images with high-resolution images with the help of two types of discriminators to generate more realistic images. The first discriminator evaluates an input sample’s local regions or patches. On the other hand, the second discriminator evaluates the entire input sample as a whole. The result shows that the proposed model can significantly reduce the noise in the generated super-resolved image; also, the SR image generated using the proposed method provides competitive advantages over the images generated using other models.

With the continuous expansion of image processing applications, there is a growing need for enhanced image resolution to capture finer details, particularly in images containing small objects. Nonetheless, several existing approaches to deep-based algorithm reconstruction of finer images still encounter difficulties in effectively harnessing hierarchical features present in the ground images. Henceforth, their performance remains relatively suboptimal. This paper proposes a Deep Enhanced Residual Generative Network with Adversarial Attentive Mechanism (DERAM) for single image super-resolution. The generator incorporates a sequential addition of multiple residual blocks featuring skip connections, while the discriminator employs a combination of numerous convolutional blocks with dense layers. Our experimentation with three extensively employed natural image datasets illustrates that our model exhibits notably higher metrics than the existing noble approaches. As evidenced by the outcomes, our model effectively diminishes noise and undesired pixels in the super-resolved (SR) image.

Structure-based protein design has attracted increasing interest, with numerous methods being introduced in recent years. However, a universally accepted method for evaluation has not been established, since the wet-lab validation can be overly time-consuming for the development of new algorithms, and the in silico validation with recovery and perplexity metrics is efficient but may not precisely reflect true foldability. To address this gap, we introduce two novel metrics: refoldability-based metric, which leverages high-accuracy protein structure prediction models as a proxy for wet lab experiments, and stability-based metric, which assesses whether models can assign high likelihoods to experimentally stable proteins. We curate datasets from high-quality CATH protein data, high-throughput de novo designed proteins, and mega-scale experimental mutagenesis experiments, and in doing so, present the PDB-Struct benchmark that evaluates both recent and previously uncompared protein design methods. Experimental results indicate that ByProt, ProteinMPNN, and ESM-IF perform exceptionally well on our benchmark, while ESM-Design and AF-Design fall short on the refoldability metric. We also show that while some methods exhibit high sequence recovery, they do not perform as well on our new benchmark. Our proposed benchmark paves the way for a fair and comprehensive evaluation of protein design methods in the future. Code is available at https://github.com/WANG-CR/PDB-Struct.

t-SNE remains one of the most popular embedding techniques for visualizing high-dimensional data. Most standard packages of t-SNE, such as scikit-learn, use the Barnes-Hut t-SNE (BH t-SNE) algorithm for large datasets. However, existing CPU implementations of this algorithm are inefficient. In this work, we accelerate the BH t-SNE on CPUs via cache optimizations, SIMD, parallelizing sequential steps, and improving parallelization of multithreaded steps. Our implementation (Acc-t-SNE) is up to 261x and 4x faster than scikit-learn and the state-of-the-art BH t-SNE implementation from daal4py, respectively, on a 32-core Intel(R) Icelake cloud instance.

Identifying accessible chromatin regions is a fundamental problem in epigenomics with ATAC-seq being a commonly used assay. Exponential rise in ATAC-seq experiments has made it critical to accelerate processing of ATAC-seq data that can have a low signal-to-noise ratio for various reasons including low coverage or low cell count. To denoise and identify accessible chromatin regions from noisy ATAC-seq data, use of deep learning on 1D data - using large filter sizes, long tensor widths, and/or dilation - has recently been proposed. Convolutions over 1D data consume a majority of the runtime in these methods. However, existing implementations of the 1D convolution layer for CPUs and GPUs fail to efficiently use the underlying architecture especially in the case of large filter sizes, long tensor widths, and dilation. Here, we present ways to accelerate the end-to-end training performance of these deep learning based methods. We evaluate our approach on the recently released AtacWorks toolkit using modern CPUs. Compared to AtacWorks running on an Nvidia DGX-1 box with 8 V100 GPUs, we get up to 2.27× speedup using just 16 CPU sockets. To achieve this, we build an efficient 1D dilated convolution layer and demonstrate reduced precision (BFloat16) training and nearly linear scaling from 1 to 16 sockets. Code Availability: https://github.com/IntelLabs/Trans-Omics-Acceleration-Library/tree/ATAC-Seq/applications/ATAC-Seq

Convolutional neural networks (CNNs) have found many applications in tasks involving two-dimensional (2D) data, such as image classification and image processing. Therefore, 2D convolution layers have been heavily optimized on CPUs and GPUs. However, in many applications - for example genomics and speech recognition, the data can be one-dimensional (1D). Such applications can benefit from optimized 1D convolution layers. In this work, we introduce our efficient implementation of a generic 1D convolution layer covering a wide range of parameters. It is optimized for x86 CPU architectures, in particular, for architectures containing Intel AVX-512 and AVX-512 BFloat16 instructions. We use the LIBXSMM library's batch-reduce General Matrix Multiplication (BRGEMM) kernel for FP32 and BFloat16 precision. We demonstrate that our implementation can achieve up to 80% efficiency on Intel Xeon Cascade Lake and Cooper Lake CPUs. Additionally, we show the generalization capability of our BRGEMM based approach by achieving high efficiency across a range of parameters. We consistently achieve higher efficiency than the 1D convolution layer with Intel oneDNN library backend for varying input tensor widths, filter widths, number of channels, filters, and dilation parameters. Finally, we demonstrate the performance of our optimized 1D convolution layer by utilizing it in the end-to-end neural network training with real genomics datasets and achieve up to 6.86x speedup over the oneDNN library-based implementation on Cascade Lake CPUs. We also demonstrate the scaling with 16 sockets of Cascade/Cooper Lake CPUs and achieve significant speedup over eight V100 GPUs using a similar power envelop. In the end-to-end training, we get a speedup of 1.41x on Cascade Lake with FP32, 1.57x on Cooper Lake with FP32, and 2.27x on Cooper Lake with BFloat16 over eight V100 GPUs with FP32.

During the past decade, novel Deep Learning (DL) algorithms/workloads and hardware have been developed to tackle a wide range of problems. Despite the advances in workload/hardware ecosystems, the programming methodology of DL systems is stagnant. DL workloads leverage either highly-optimized, yet platform-specific and inflexible kernels from DL libraries, or in the case of novel operators, reference implementations are built via DL framework primitives with underwhelming performance. This work introduces the Tensor Processing Primitives (TPP), a programming abstraction striving for efficient, portable implementation of DL workloads with high-productivity. TPPs define a compact, yet versatile set of 2D-tensor operators (or a virtual Tensor ISA), which subsequently can be utilized as building-blocks to construct complex operators on high-dimensional tensors. The TPP specification is platform-agnostic, thus code expressed via TPPs is portable, whereas the TPP implementation is highly-optimized and platform-specific. We demonstrate the efficacy of our approach using standalone kernels and end-to-end DL workloads expressed entirely via TPPs that outperform state-of-the-art implementations on multiple platforms.

Identifying accessible chromatin regions is a fundamental problem in epigenomics with ATAC-seq being a commonly used assay. Exponential rise in single cell ATAC-seq experiments has made it critical to accelerate processing of ATAC-seq data. ATAC-seq data can have a low signal-to-noise ratio for various reasons including low coverage or low cell count. To denoise and identify accessible chromatin regions from noisy ATAC-seq data, use of deep learning on 1D data – using large filter sizes, long tensor widths, and/or dilation - has recently been proposed. Here, we present ways to accelerate the end-to-end training performance of these deep learning based methods using CPUs. We evaluate our approach on the recently released AtacWorks toolkit. Compared to an Nvidia DGX-1 box with 8 V100 GPUs, we get up to 2.27× speedup using just 16 CPU sockets. To achieve this, we build an efficient 1D dilated convolution layer and demonstrate reduced precision (BFloat16) training.

The estimation of line and contour geometries from real SEM images is a challenging problem due to the corruption of such images by Poisson noise, edge effects, and other SEM artifacts. We attempt simultaneous contour edge image prediction and SEM image denoising using a deep convolutional neural network LineNet2. To capture a range of edge effects in real SEM images, we simulate a training dataset of rough line SEM images with random edge effect parameters. We train the LineNet2 network on this training dataset and randomly rotate the images during the training phase. The retrained LineNet2 shows the ability to denoise real SEM images of line and contour geometries. We measure the line edge roughness (LER) parameter in isolated and dense regions of rough line images through multiple LER methods. Our experiments also demonstrate that the network can learn to recognize contour edges just by rotating rough line images.

Scanning electron microscopy images are an attractive option to estimate the roughness of nanostructures. Convolutional neural network (CNN) based algorithms have improved scanning electron microscope (SEM) image denoising and estimation of line roughness measurements. However, these algorithms need improvements to run at high speeds with a low memory footprint and without compromising accuracy. We introduce two approaches to reduce computation time and memory. We first propose deep CNNs LineNet1 and LineNet2 to perform simultaneous denoising and edge estimation on rough line SEM images. This multiple task formulation in LineNet1 and LineNet2 reduces training time, inference time and model sizes. LineNet2 also facilitates edge estimation in the multiple-line images and generalizes the approach for other geometries. Our training method uses supervised learning datasets of single-line SEM images and multiple-line SEM images together with edge positions information. We next consider multiple visualization tools to improve our understanding of the LineNet1 architecture and use the resulting insights from these visualizations to motivate a study of two variations of LineNet1 with fewer neural network layers. One of these visualization techniques is new to the visualization of denoising CNNs. Our results show that these approaches significantly reduce the memory and computation needed for edge estimation with a slight impact on accuracy.

Low dose scanning electron microscope (SEM) images are an attractive option to estimate the roughness of nanos- tructures. We recently proposed two deep convolutional neural network (CNN) architectures named “LineNet” to simultaneously perform denoising and edge estimation on rough line SEM images. In this paper we consider multiple visualization tools to improve our understanding of LineNet1; one of these techniques is new to the visualization of denoising CNNs. We use the resulting insights from these visualizations to motivate a study of two variations of LineNet1 with fewer neural network layers. Furthermore, although in classification CNNs edge detection is commonly believed to happen early in the network, the visualization techniques suggest that important aspects of edge detection in LineNet1 occur late in the network.

Deriving models for lithographic masks based either on first principles or using an empirical model is becoming increasingly challenging as complex effects (once relegated to noise level) become more relevant. Deep Learning offers an alternative solution that can leapfrog the shortcomings of these previous approaches but requires a source of input data that contains enough diversity to allow an effective training of the neural networks.The solution for mask lithography modeling presented in this paper makes use of carefully calibrated SEM images to extract the information required to allow the training and testing of a deep convolutional neural network that achieves accuracy beyond what can be done in metrology-based methods. We demonstrate how the input data is calibrated to be consumed in this flow and present examples demonstrating its predicting power which can, for instance, detect the location and shape of hotspots in the layout. One significant additional advantage is the improvement in the ease and speed of building models compared to previous solutions which can dovetail well with regular production flows and can be adapted to dynamic changes in the mask process.

We propose deep convolutional neural networks LineNet1 and LineNet2 for simultaneous denoising and edge image prediction from low-dose scanning electron microscope images. Edge estimation of nanostructures from SEM images is needed for line edge roughness (LER) and line width roughness (LWR) estimation. Our method uses supervised learning datasets of single-line SEM images and multiple-line SEM images together with edge positions information for the training of LineNet1 and LineNet2. We simulate single-line and multiple-line SEM images with Poisson noise and other artifacts using the ARTIMAGEN library developed by the National Institute of Standards and Technology. The line edges were generated using the Thorsos method and the Palasantzas spectral model. The convolutional neural networks LineNet1 and LineNet2 each contain 17 convolutional layers, 16 batch-normalization layers and 16 dropout layers. Our results show that this approach (1) facilitates edge estimation in multiple-line images and (2) significantly reduces the memory needed for edge estimation in single-line images with a slight impact on accuracy.

We propose the use of deep supervised learning for the estimation of line edge roughness (LER) and line width roughness (LWR) in low-dose scanning electron microscope (SEM) images. We simulate a supervised learning dataset of 100,800 SEM rough line images constructed by means of the Thorsos method and the ARTIMAGEN library developed by the National Institute of Standards and Technology. We also devise two separate deep convolutional neural networks called SEMNet and EDGENet, each of which has 17 convolutional layers, 16 batch normalization layers, and 16 dropout layers. SEMNet performs the Poisson denoising of SEM images, and it is trained with a dataset of simulated noisy-original SEM image pairs. EDGENet directly estimates the edge geometries from noisy SEM images, and it is trained with a dataset of simulated noisy SEM image-edge array pairs. SEMNet achieved considerable improvements in peak signal-to-noise ratio as well as the best LER/LWR estimation accuracy compared with standard image denoisers. EDGENet offers excellent LER and LWR estimation as well as roughness spectrum estimation. (C) 2019 Society of Photo-Optical Instrumentation Engineers (SPIE)

We propose a deep convolutional neural network named EDGENet to estimate rough line edge positions in low-dose scanning electron microscope (SEM) images corrupted by Poisson noise, Gaussian blur, edge effects and other instrument errors and apply our approach to the estimation of line edge roughness (LER) and line width roughness (LWR). Our method uses a supervised learning dataset of 100800 input-output pairs of simulated noisy SEM rough line images with true edge positions. The edges were constructed by the Thorsos method and have an underlying Palasantzas spectral model. The simulated SEM images were created using the ARTIMAGEN library developed at the National Institute of Standards and Technology. The convolutional neural network EDGENet consists of 17 convolutional, 16 batch-normalization layers and 16 dropout layers and offers excellent LER and LWR estimation as well as roughness spectrum estimation.

We use deep supervised learning for the Poisson denoising of low-dose scanning electron microscope (SEM) images as a step in the estimation of line edge roughness (LER) and line width roughness (LWR). Our denoising algorithm applies a deep convolutional neural network called SEMNet with 17 convolutional, 16 batch-normalization and 16 dropout layers to noisy images. We trained and tested SEMNet with a dataset of 100800 simulated SEM rough line images constructed by means of the Thorsos method and the ARTIMAGEN library developed by the National Institute of Standards and Technology. SEMNet achieved considerable improvements in peak signal-to-noise ratio (PSNR) as well as the best LER/LWR estimation accuracy compared with standard image denoisers.

Multibeam electron beam systems will be used in the future for mask writing and for complimentary lithography. The major challenges of the multibeam systems are in meeting throughput requirements and in handling the large data volumes associated with writing grayscale data on the wafer. In terms of future communications and computational requirements Amdahl’s Law suggests that a simple increase of computation power and parallelism may not be a sustainable solution. We propose a parallel data compression algorithm to exploit the sparsity of mask data and a grayscale video-like representation of data. To improve the communication and computational efficiency of these systems at the write time we propose an alternate datapath architecture partly motivated by multibeam direct write lithography and partly motivated by the circuit testing literature, where parallel decompression reduces clock cycles. We explain a deflection plate architecture inspired by NuFlare Technology’s multibeam mask writing system and how our datapath architecture can be easily added to it to improve performance.