Insights into the inner workings of transformer models for protein function prediction.

MOTIVATION: We explored how explainable artificial intelligence (XAI) can help to shed light into the inner workings of neural networks for protein function prediction, by extending the widely used XAI method of integrated gradients such that latent representations inside of transformer models, which were finetuned to Gene Ontology term and Enzyme Commission number prediction, can be inspected too. RESULTS: The approach enabled us to identify amino acids in the sequences that the transformers pay particular attention to, and to show that these relevant sequence parts reflect expectations from biology and chemistry, both in the embedding layer and inside of the model, where we identified transformer heads with a statistically significant correspondence of attribution maps with ground truth sequence annotations (e.g. transmembrane regions, active sites) across many proteins. AVAILABILITY AND IMPLEMENTATION: Source code can be accessed at https://github.com/markuswenzel/xai-proteins.

Fall risk stratification of community-living older people. Commentary on the world guidelines for fall prevention and management.

The Task Force on Global Guidelines for Falls in Older Adults has put forward a fall risk stratification tool for community-dwelling older adults. This tool takes the form of a flowchart and is based on expert opinion and evidence. It divides the population into three risk categories and recommends specific preventive interventions or treatments for each category. In this commentary, we share our insights on the design, validation, usability and potential impact of this fall risk stratification tool with the aim of guiding future research.

UDSMProt: universal deep sequence models for protein classification.

MOTIVATION: Inferring the properties of a protein from its amino acid sequence is one of the key problems in bioinformatics. Most state-of-the-art approaches for protein classification are tailored to single classification tasks and rely on handcrafted features, such as position-specific-scoring matrices from expensive database searches. We argue that this level of performance can be reached or even be surpassed by learning a task-agnostic representation once, using self-supervised language modeling, and transferring it to specific tasks by a simple fine-tuning step. RESULTS: We put forward a universal deep sequence model that is pre-trained on unlabeled protein sequences from Swiss-Prot and fine-tuned on protein classification tasks. We apply it to three prototypical tasks, namely enzyme class prediction, gene ontology prediction and remote homology and fold detection. The proposed method performs on par with state-of-the-art algorithms that were tailored to these specific tasks or, for two out of three tasks, even outperforms them. These results stress the possibility of inferring protein properties from the sequence alone and, on more general grounds, the prospects of modern natural language processing methods in omics. Moreover, we illustrate the prospects for explainable machine learning methods in this field by selected case studies. AVAILABILITY AND IMPLEMENTATION: Source code is available under https://github.com/nstrodt/UDSMProt. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

Machine Learning for Health: Algorithm Auditing & Quality Control.

Developers proposing new machine learning for health (ML4H) tools often pledge to match or even surpass the performance of existing tools, yet the reality is usually more complicated. Reliable deployment of ML4H to the real world is challenging as examples from diabetic retinopathy or Covid-19 screening show. We envision an integrated framework of algorithm auditing and quality control that provides a path towards the effective and reliable application of ML systems in healthcare. In this editorial, we give a summary of ongoing work towards that vision and announce a call for participation to the special issue  Machine Learning for Health: Algorithm Auditing & Quality Control in this journal to advance the practice of ML4H auditing.

USMPep: universal sequence models for major histocompatibility complex binding affinity prediction.

BACKGROUND: Immunotherapy is a promising route towards personalized cancer treatment. A key algorithmic challenge in this process is to decide if a given peptide (neoepitope) binds with the major histocompatibility complex (MHC). This is an active area of research and there are many MHC binding prediction algorithms that can predict the MHC binding affinity for a given peptide to a high degree of accuracy. However, most of the state-of-the-art approaches make use of complicated training and model selection procedures, are restricted to peptides of a certain length and/or rely on heuristics. RESULTS: We put forward USMPep, a simple recurrent neural network that reaches state-of-the-art approaches on MHC class I binding prediction with a single, generic architecture and even a single set of hyperparameters both on IEDB benchmark datasets and on the very recent HPV dataset. Moreover, the algorithm is competitive for a single model trained from scratch, while ensembling multiple regressors and language model pretraining can still slightly improve the performance. The direct application of the approach to MHC class II binding prediction shows a solid performance despite of limited training data. CONCLUSIONS: We demonstrate that competitive performance in MHC binding affinity prediction can be reached with a standard architecture and training procedure without relying on any heuristics.

Automatic classification of dopamine transporter SPECT: deep convolutional neural networks can be trained to be robust with respect to variable image characteristics.

PURPOSE: This study investigated the potential of deep convolutional neural networks (CNN) for automatic classification of FP-CIT SPECT in multi-site or multi-camera settings with variable image characteristics. METHODS: The study included FP-CIT SPECT of 645 subjects from the Parkinson's Progression Marker Initiative (PPMI), 207 healthy controls, and 438 Parkinson's disease patients. SPECT images were smoothed with an isotropic 18-mm Gaussian kernel resulting in 3 different PPMI settings: (i) original (unsmoothed), (ii) smoothed, and (iii) mixed setting comprising all original and all smoothed images. A deep CNN with 2,872,642 parameters was trained, validated, and tested separately for each setting using 10 random splits with 60/20/20% allocation to training/validation/test sample. The putaminal specific binding ratio (SBR) was computed using a standard anatomical ROI predefined in MNI space (AAL atlas) or using the hottest voxels (HV) analysis. Both SBR measures were trained (ROC analysis, Youden criterion) using the same random splits as for the CNN. CNN and SBR trained in the mixed PPMI setting were also tested in an independent sample from clinical routine patient care (149 with non-neurodegenerative and 149 with neurodegenerative parkinsonian syndrome). RESULTS: Both SBR measures performed worse in the mixed PPMI setting compared to the pure PPMI settings (e.g., AAL-SBR accuracy = 0.900 ± 0.029 in the mixed setting versus 0.957 ± 0.017 and 0.952 ± 0.015 in original and smoothed setting, both p < 0.01). In contrast, the CNN showed similar accuracy in all PPMI settings (0.967 ± 0.018, 0.972 ± 0.014, and 0.955 ± 0.009 in mixed, original, and smoothed setting). Similar results were obtained in the clinical sample. After training in the mixed PPMI setting, only the CNN provided acceptable performance in the clinical sample. CONCLUSIONS: These findings provide proof of concept that a deep CNN can be trained to be robust with respect to variable site-, camera-, or scan-specific image characteristics without a large loss of diagnostic accuracy compared with mono-site/mono-camera settings. We hypothesize that a single CNN can be used to support the interpretation of FP-CIT SPECT at many different sites using different acquisition hardware and/or reconstruction software with only minor harmonization of acquisition and reconstruction protocols.

Visuomotor functional network topology predicts upcoming tasks.

It is a vital ability of humans to flexibly adapt their behavior to different environmental situations. Constantly, the rules for our sensory-to-motor mappings need to be adapted to the current task demands. For example, the same sensory input might require two different motor responses depending on the actual situation. How does the brain prepare for such different responses? It has been suggested that the functional connections within cortex are biased according to the present rule to guide the flow of information in accordance with the required sensory-to-motor mapping. Here, we investigated with fMRI whether task settings might indeed change the functional connectivity structure in a large-scale brain network. Subjects performed a visuomotor response task that required an interaction between visual and motor cortex: either within each hemisphere or across the two hemispheres of the brain depending on the task condition. A multivariate analysis on the functional connectivity graph of a cortical visuomotor network revealed that the functional integration, i.e., the connectivity structure, is altered according to the task condition already during a preparatory period before the visual cue and the actual movement. Our results show that the topology of connection weights within a single network changes according to and thus predicts the upcoming task. This suggests that the human brain prepares to respond in different conditions by altering its large scale functional connectivity structure even before an action is required.

From modern CNNs to vision transformers: Assessing the performance, robustness, and classification strategies of deep learning models in histopathology.

While machine learning is currently transforming the field of histopathology, the domain lacks a comprehensive evaluation of state-of-the-art models based on essential but complementary quality requirements beyond a mere classification accuracy. In order to fill this gap, we developed a new methodology to extensively evaluate a wide range of classification models, including recent vision transformers, and convolutional neural networks such as: ConvNeXt, ResNet (BiT), Inception, ViT and Swin transformer, with and without supervised or self-supervised pretraining. We thoroughly tested the models on five widely used histopathology datasets containing whole slide images of breast, gastric, and colorectal cancer and developed a novel approach using an image-to-image translation model to assess the robustness of a cancer classification model against stain variations. Further, we extended existing interpretability methods to previously unstudied models and systematically reveal insights of the models' classification strategies that allow for plausibility checks and systematic comparisons. The study resulted in specific model recommendations for practitioners as well as putting forward a general methodology to quantify a model's quality according to complementary requirements that can be transferred to future model architectures.

Comparison of Anatomical Preformed Titanium Implants and Patient-Specific CAD/CAM Implants in the Primary Reconstruction of Isolated Orbital Fractures-A Retrospective Study.

BACKGROUND/AIM: Reconstruction of the fractured orbit remains a challenge. The aim of this study was to compare anatomical preformed titanium orbital implants with patient-specific CAD/CAM implants for precision and intraoperative applicability. MATERIAL AND METHODS: A total of 75 orbital reconstructions from 2012 to 2022 were retrospectively assessed for their precision of implant position and intra- and postoperative revision rates. For this purpose, the implant position after digital orbital reconstruction was checked for deviations by mirroring the healthy orbit at 5 defined points, and the medical records of the patients were checked for revisions. RESULTS: The evaluation of the 45 anatomical preformed orbital implant cases showed significantly higher deviations and an implant inaccuracy of 66.6% than the 30 CAD/CAM cases with only 10% inaccuracy. In particular, the CAD/CAM implants were significantly more precise in medial and posterior positioning. In addition, the intraoperative revision rates of 26.6% vs. 11% after 3D intraoperative imaging and the postoperative revision rates of 13% vs. 0 for the anatomical preformed implants were significantly higher than for patient-specific implants. CONCLUSION: We conclude that patient-specific CAD/CAM orbital implants are highly suitable for primary orbital reconstruction. These seem to be preferable to anatomical preformed implants in terms of precision and revision rates.