Mara Graziani

Trial-and-error approaches in chemistry generate abundant unsuccessful experiments, yet the potential of these so-called negative results remains largely underutilized. Here, we demonstrate that information from negative chemical reactions can be leveraged to improve reactivity-prediction models, offering advantages in scenarios with a limited volume of successful data. We extend the tuning of language models with reinforcement learning to the chemistry domain, training a transformer model for chemical reaction prediction. Our approach is evaluated using both a rigorously controlled dataset and a realistic high-throughput dataset comprising extensive reaction screenings across diverse catalysts sets and experimental conditions. The model achieves state-of-the-art performance by leveraging information from as few as 20 positive data points in the controlled dataset, supported by a negative dataset at least 40 times larger. Consistent results on both datasets demonstrate that, with an appropriate optimization strategy and the inclusion of unsuccessful experimental data, models can be effectively trained even when successful reactions are underrepresented.

Summary: The application of machine learning methods to biomedical applications has seen many successes. However, working with transcriptomic data on supervised learning tasks is challenging due to its high dimensionality, low patient numbers, and class imbalances. Machine learning models tend to overfit these data and do not generalize well on out-of-distribution samples. Data augmentation strategies help alleviate this by introducing synthetic data points and acting as regularizers. However, existing approaches are either computationally intensive, require population parametric estimates, or generate insufficiently diverse samples. To address these challenges, we introduce two classes of phenotype-driven data augmentation approaches—signature-dependent and signature-independent. The signature-dependent methods assume the existence of distinct gene signatures describing some phenotype and are simple, non-parametric, and novel data augmentation methods. The signature-independent methods are a modification of the established Gamma-Poisson and Poisson sampling methods for gene expression data. As case studies, we apply our augmentation methods to transcriptomic data of colorectal and breast cancer. Through discriminative and generative experiments with external validation, we show that our methods improve patient stratification by 5−15% over other augmentation methods in their respective cases. The study additionally provides insights into the limited benefits of over-augmenting data. Availability and implementation: Code for reproducibility is available on GitHub.

Deep Learning (DL) can predict biomarkers from cancer histopathology. Several clinically approved applications use this technology. Most approaches, however, predict categorical labels, whereas biomarkers are often continuous measurements. We hypothesize that regression-based DL outperforms classification-based DL. Therefore, we develop and evaluate a self-supervised attention-based weakly supervised regression method that predicts continuous biomarkers directly from 11,671 images of patients across nine cancer types. We test our method for multiple clinically and biologically relevant biomarkers: homologous recombination deficiency score, a clinically used pan-cancer biomarker, as well as markers of key biological processes in the tumor microenvironment. Using regression significantly enhances the accuracy of biomarker prediction, while also improving the predictions’ correspondence to regions of known clinical relevance over classification. In a large cohort of colorectal cancer patients, regression-based prediction scores provide a higher prognostic value than classification-based scores. Our open-source regression approach offers a promising alternative for continuous biomarker analysis in computational pathology.

With machine learning taking over biomedical applications, working with transcriptomic data on supervised learning tasks is challenging due to high dimensionality, low patient numbers and class imbalances. Machine learning models tend to overfit these data and do not generalise well on out-of-distribution samples. Data augmentation strategies help alleviate this by introducing synthetic data points and acting as regularisers. However, existing approaches are either computationally intensive, require population parametric estimates or generate insufficiently diverse samples. To address these challenges, we introduce two classes of phenotype driven data augmentation approaches – signature-dependent and signature-independent. The signature-dependent methods assume the existence of distinct gene signatures describing some phenotype and are simple, non-parametric, and novel data augmentation methods. The signature-independent methods are a modification of the established Gamma-Poisson and Poisson sampling methods for gene expression data. As case studies, we apply our augmentation methods to transcriptomic data of colorectal and breast cancer. Through discriminative and generative experiments with external validation, we show that our methods improve patient stratification by 5 − 15% over other augmentation methods in different cases. The study additionally provides insights into the limited benefits of over-augmenting data. The code is hosted on GitHub , and includes a link to the augmented datasets.

Interpreting the inner workings of deep learning models is crucial for establishing trust and ensuring model safety. Concept-based explanations have emerged as a superior approach that is more interpretable than feature attribution estimates such as pixel saliency. However, defining the concepts for the interpretability analysis biases the explanations by the user's expectations on the concepts. To address this, we propose a novel post-hoc unsupervised method that automatically uncovers the concepts learned by deep models during training. By decomposing the latent space of a layer in singular vectors and refining them by unsupervised clustering, we uncover concept vectors aligned with directions of high variance that are relevant to the model prediction, and that point to semantically distinct concepts. Our extensive experiments reveal that the majority of our concepts are readily understandable to humans, exhibit coherency, and bear relevance to the task at hand. Moreover, we showcase the practical utility of our method in dataset exploration, where our concept vectors successfully identify outlier training samples affected by various confounding factors. This novel exploration technique has remarkable versatility to data types and model architectures and it will facilitate the identification of biases and the discovery of sources of error within training data.

Adopting Convolutional Neural Networks (CNNs) in the daily routine of primary diagnosis requires not only near-perfect precision, but also a sufficient degree of generalization to data acquisition shifts and transparency. Existing CNN models act as black boxes, not ensuring to the physicians that important diagnostic features are used by the model. Building on top of successfully existing techniques such as multi-task learning, domain adversarial training and concept-based interpretability, this paper addresses the challenge of introducing diagnostic factors in the training objectives. Here we show that our architecture, by learning end-to-end an uncertainty-based weighting combination of multi-task and adversarial losses, is encouraged to focus on pathology features such as density and pleomorphism of nuclei, e.g. variations in size and appearance, while discarding misleading features such as staining differences. Our results on breast lymph node tissue show significantly improved generalization in the detection of tumorous tissue, with best average AUC 0.89 (0.01) against the baseline AUC 0.86 (0.005). By applying the interpretability technique of linearly probing intermediate representations, we also demonstrate that interpretable pathology features such as nuclei density are learned by the proposed CNN architecture, confirming the increased transparency of this model. This result is a starting point towards building interpretable multi-task architectures that are robust to data heterogeneity. Our code is available at https://github.com/maragraziani/multitask_adversarial

Mechanistic interpretability aims to understand how models store representations by breaking down neural networks into interpretable units. However, the occurrence of polysemantic neurons, or neurons that respond to multiple unrelated features, makes interpreting individual neurons challenging. This has led to the search for meaningful vectors, known as concept vectors, in activation space instead of individual neurons. The main contribution of this paper is a method to disentangle polysemantic neurons into concept vectors encapsulating distinct features. Our method can search for fine-grained concepts according to the user's desired level of concept separation. The analysis shows that polysemantic neurons can be disentangled into directions consisting of linear combinations of neurons. Our evaluations show that the concept vectors found encode coherent, human-understandable features.

Since its emergence in the 1960s, Artificial Intelligence (AI) has grown to conquer many technology products and their fields of application. Machine learning, as a major part of the current AI solutions, can learn from the data and through experience to reach high performance on various tasks. This growing success of AI algorithms has led to a need for interpretability to understand opaque models such as deep neural networks. Various requirements have been raised from different domains, together with numerous tools to debug, justify outcomes, and establish the safety, fairness and reliability of the models. This variety of tasks has led to inconsistencies in the terminology with, for instance, terms such as interpretable, explainable and transparent being often used interchangeably in methodology papers. These words, however, convey different meanings and are “weighted" differently across domains, for example in the technical and social sciences. In this paper, we propose an overarching terminology of interpretability of AI systems that can be referred to by the technical developers as much as by the social sciences community to pursue clarity and efficiency in the definition of regulations for ethical and reliable AI development. We show how our taxonomy and definition of interpretable AI differ from the ones in previous research and how they apply with high versatility to several domains and use cases, proposing a—highly needed—standard for the communication among interdisciplinary areas of AI.