We explore the potential of Large Language Models (LLMs) to assist and potentially correct physicians in medical decision-making tasks. We evaluate several LLMs, including Meditron, Llama2, and Mistral, to analyze the ability of these models to interact effectively with physicians across different scenarios. We consider questions from PubMedQA and several tasks, ranging from binary (yes/no) responses to long answer generation, where the answer of the model is produced after an interaction with a physician. Our findings suggest that prompt design significantly influences the downstream accuracy of LLMs and that LLMs can provide valuable feedback to physicians, challenging incorrect diagnoses and contributing to more accurate decision-making. For example, when the physician is accurate 38{\%} of the time, Mistral can produce the correct answer, improving accuracy up to 74{\%} depending on the prompt being used, while Llama2 and Meditron models exhibit greater sensitivity to prompt choice. Our analysis also uncovers the challenges of ensuring that LLM-generated suggestions are pertinent and useful, emphasizing the need for further research in this area.

The advent of powerful neural classifiers has increased interest in problems that require both learning and reasoning. These problems are critical for understanding important properties of models, such as trustworthiness, generalization, interpretability, and compliance to safety and structural constraints. However, recent research observed that tasks requiring both learning and reasoning on background knowledge often suffer from reasoning shortcuts (RSs): predictors can solve the downstream reasoning task without associating the correct concepts to the high-dimensional data. To address this issue, we introduce rsbench, a comprehensive benchmark suite designed to systematically evaluate the impact of RSs on models by providing easy access to highly customizable tasks affected by RSs. Furthermore, rsbench implements common metrics for evaluating concept quality and introduces novel formal verification procedures for assessing the presence of RSs in learning tasks. Using rsbench, we highlight that obtaining high quality concepts in both purely neural and neuro-symbolic models is a far-from-solved problem. rsbench is available at: https://unitn-sml.github.io/rsbench.

Neuro-Symbolic (NeSy) predictors that conform to symbolic knowledge – encoding, e.g., safety constraints – can be affected by Reasoning Shortcuts (RSs): They learn concepts consistent with the symbolic knowledge by exploiting unintended semantics. RSs compromise reliability and generalization and, as we show in this paper, they are linked to NeSy models being overconfident about the predicted concepts. Unfortunately, the only trustworthy mitigation strategy requires collecting costly dense supervision over the concepts. Rather than attempting to avoid RSs altogether, we propose to ensure NeSy models are aware of the semantic ambiguity of the concepts they learn, thus enabling their users to identify and distrust low-quality concepts. Starting from three simple desiderata, we derive bears (BE Aware of Reasoning Shortcuts), an ensembling technique that calibrates the model’s concept-level confidence without compromising prediction accuracy, thus encouraging NeSy architectures to be uncertain about concepts affected by RSs. We show empirically that bears improves RS-awareness of several state-of-the-art NeSy models, and also facilitates acquiring informative dense annotations for mitigation purposes.

Large Language Models (LLMs) demonstrate an impressive capacity to recall a vast range of factual knowledge. However, understanding their underlying reasoning and internal mechanisms in exploiting this knowledge remains a key research area. This work unveils the factual information an LLM represents internally for sentence-level claim verification. We propose an end-to-end framework to decode factual knowledge embedded in token representations from a vector space to a set of ground predicates, showing its layer-wise evolution using a dynamic knowledge graph. Our framework employs activation patching, a vector-level technique that alters a token representation during inference, to extract encoded knowledge. Accordingly, we neither rely on training nor external models. Using factual and common-sense claims from two claim verification datasets, we showcase interpretability analyses at local and global levels. The local analysis highlights entity centrality in LLM reasoning, from claim-related information and multi-hop reasoning to representation errors causing erroneous evaluation. On the other hand, the global reveals trends in the underlying evolution, such as word-based knowledge evolving into claim-related facts. By interpreting semantics from LLM latent representations and enabling graph-related analyses, this work enhances the understanding of the factual knowledge resolution process.

Over the last decade, several regulatory bodies have started requiring the disclosure of non-financial information from publicly listed companies, in light of the investors' increasing attention to Environmental, Social, and Governance (ESG) issues. Publicly released information on sustainability practices is often disclosed in diverse, unstructured, and multi-modal documentation. This poses a challenge in efficiently gathering and aligning the data into a unified framework to derive insights related to Corporate Social Responsibility (CSR). Thus, using Information Extraction (IE) methods becomes an intuitive choice for delivering insightful and actionable data to stakeholders. In this study, we employ Large Language Models (LLMs), In-Context Learning, and the Retrieval-Augmented Generation (RAG) paradigm to extract structured insights related to ESG aspects from companies' sustainability reports. We then leverage graph-based representations to conduct statistical analyses concerning the extracted insights. These analyses revealed that ESG criteria cover a wide range of topics, exceeding 500, often beyond those considered in existing categorizations, and are addressed by companies through a variety of initiatives. Moreover, disclosure similarities emerged among companies from the same region or sector, validating ongoing hypotheses in the ESG literature. Lastly, by incorporating additional company attributes into our analyses, we investigated which factors impact the most on companies' ESG ratings, showing that ESG disclosure affects the obtained ratings more than other financial or company data.

The development of efficient exact and approximate algorithms for probabilistic inference is a long-standing goal of artificial intelligence research. Whereas substantial progress has been made in dealing with purely discrete or purely continuous domains, adapting the developed solutions to tackle hybrid domains, characterized by discrete and continuous variables and their relationships, is highly non-trivial. Weighted Model Integration (WMI) recently emerged as a unifying formalism for probabilistic inference in hybrid domains. Despite a considerable amount of recent work, allowing WMI algorithms to scale with the complexity of the hybrid problem is still a challenge. In this paper we highlight some substantial limitations of existing state-of-the-art solutions, and develop an algorithm that combines SMT-based enumeration, an efficient technique in formal verification, with an effective encoding of the problem structure. This allows our algorithm to avoid generating redundant models, resulting in drastic computational savings. Additionally, we show how SMT-based approaches can seamlessly deal with different integration techniques, both exact and approximate, significantly expanding the set of problems that can be tackled by WMI technology. An extensive experimental evaluation on both synthetic and real-world datasets confirms the substantial advantage of the proposed solution over existing alternatives. The application potential of this technology is further showcased on a prototypical task aimed at verifying the fairness of probabilistic programs.

Being able to provide counterfactual interventions—sequences of actions we would have had to take for a desirable outcome to happen—is essential to explain how to change an unfavourable decision by a black-box machine learning model (e.g., being denied a loan request). Existing solutions have mainly focused on generating feasible interventions without providing explanations of their rationale. Moreover, they need to solve a separate optimization problem for each user. In this paper, we take a different approach and learn a program that outputs a sequence of explainable counterfactual actions given a user description and a causal graph. We leverage program synthesis techniques, reinforcement learning coupled with Monte Carlo Tree Search for efficient exploration, and rule learning to extract explanations for each recommended action. An experimental evaluation on synthetic and real-world datasets shows how our approach, FARE (eFficient counterfActual REcourse), generates effective interventions by making orders of magnitude fewer queries to the black-box classifier with respect to existing solutions, with the additional benefit of complementing them with interpretable explanations.

Although the efficacies of vaccines against SARS-CoV-2, i.e., the virus that causes Covid-19, have been publicized and praised, and although they are assumed to encourage vaccine compliance, little is known about how well these figures are understood by the general public. Our study aims to fill this gap by investigating whether laypeople have an adequate grasp of what vaccine efficacy means and, if not, which misconceptions and consequences are the most common. To this end, we carried out three online behavioral experiments involving 1800 participants overall. The first, exploratory experiment, with a sample of 600 UK participants, allowed us to document, by means of both an open-ended question and a multiple-choice question, a common misinterpretation of the efficacy of SARS-CoV-2 vaccines as the non-incidence rate among the vaccinated. We formally demonstrated that this error leads to a systematic overestimation of the probability of individuals who are vaccinated developing Covid-19. The second experiment confirmed the prevalence of this misinterpretation in a new sample of 600 UK and Italian participants, by means of a slightly different multiple-choice question that included more response options. Finally, in a third experiment, involving another 600 UK and Italian participants, we investigated the behavioral implications of the documented error and showed that it might undermine the general positive attitude toward vaccines as well as the intention to get vaccinated. On the whole, the results of this study reveal a general misunderstanding of vaccine efficacy that may have serious consequences for the perceived benefits of SARS-CoV-2 vaccines and, thus, the willingness to be vaccinated.

This paper introduces CLEO, a novel preference elicitation algorithm capable of recommending complex configurable objects characterized by both discrete and continuous attributes and constraints defined over them. While existing preference elicitation techniques focus on searching for the best instance in a database of candidates, CLEO takes a constructive approach to recommendation through interactive optimization in a space of feasible configurations. The algorithm assumes minimal initial information, i.e., a set of catalog attributes, and defines decisional features as logic formulae combining Boolean and algebraic constraints over the attributes. The (unknown) utility of the decision maker is modeled as a weighted combination of features. CLEO iteratively alternates a preference elicitation step, where pairs of candidate configurations are selected based on the current utility model, and a refinement step where the utility is refined by incorporating the feedback received. The elicitation step leverages a Max-SMT solver to return optimal configurations according to the current utility model. The refinement step is implemented as learning to rank, and a sparsifying norm is used to favor the selection of few informative features in the combinatorial space of candidate decisional features. A major feature of CLEO is that it can recommend optimal configurations in hybrid domains (i.e., including both Boolean and numeric attributes), thanks to the use of Max-SMT technology, while retaining uncertainty in the decision-maker's utility and noisy feedback. In so doing, it adapts the recently introduced learning modulo theory framework to the preference elicitation setting. The combinatorial formulation of the utility function coupled with the feature selection capabilities of 1-norm regularization allow to effectively deal with the uncertainty in the DM utility while retaining high expressiveness. Experimental results on complex recommendation tasks show the ability of CLEO to quickly identify optimal configurations, as well as its capacity to recover from suboptimal initial choices. Our empirical evaluation highlights how CLEO outperforms a state-of-the-art Bayesian preference elicitation algorithm when applied to a purely discrete task

Background The use of machine learning models to discriminate between patterns of neural activity has become in recent years a standard analysis approach in neuroimaging studies. Whenever these models are linear, the estimated parameters can be visualized in the form of brain maps which can aid in understanding how brain activity in space and time underlies a cognitive function. However, the recovered brain maps often suffer from lack of interpretability, especially in group analysis of multi-subject data. New method To facilitate the application of brain decoding in group-level analysis, we present an application of multi-task joint feature learning for group-level multivariate pattern recovery in single-trial magnetoencephalography (MEG) decoding. The proposed method allows for recovering sparse yet consistent patterns across different subjects, and therefore enhances the interpretability of the decoding model. Results Our experimental results demonstrate that the mutli-task joint feature learning framework is capable of recovering more meaningful patterns of varying spatio-temporally distributed brain activity across individuals while still maintaining excellent generalization performance. Comparison with existing methods We compare the performance of the multi-task joint feature learning in terms of generalization, reproducibility, and quality of pattern recovery against traditional single-subject and pooling approaches on both simulated and real \{MEG\} datasets. Conclusions These results can facilitate the usage of brain decoding for the characterization of fine-level distinctive patterns in group-level inference. Considering the importance of group-level analysis, the proposed approach can provide a methodological shift towards more interpretable brain decoding models.

Translation is increasingly recognized as a central control layer of gene expression in eukaryotic cells. The overall organization of mRNA and ribosomes within polysomes, as well as the possible role of this organization in translation are poorly understood. Here we show that polysomes are primarily formed by three distinct classes of ribosome assemblies. We observe that these assemblies can be connected by naked RNA regions of the transcript. We show that the relative proportions of the three classes of ribosome assemblies reflect, and probably dictate, the level of translational activity. These results reveal the existence of recurrent supra-ribosomal building blocks forming polysomes and suggest the presence of unexplored translational controls embedded in the polysome structure.

Viruses are typically characterized by high mutation rates, which allow them to quickly develop drug-resistant mutations. Mining relevant rules from mutation data can be extremely useful to understand the virus adaptation mechanism and to design drugs that effectively counter potentially resistant mutants.We propose a simple statistical relational learning approach for mutant prediction where the input consists of mutation data with drug-resistance information, either as sets of mutations conferring resistance to a certain drug, or as sets of mutants with information on their susceptibility to the drug. The algorithm learns a set of relational rules characterizing drug-resistance and uses them to generate a set of potentially resistant mutants. Learning a weighted combination of rules allows to attach generated mutants with a resistance score as predicted by the statistical relational model and select only the highest scoring ones.Promising results were obtained in generating resistant mutations for both nucleoside and non-nucleoside HIV reverse transcriptase inhibitors. The approach can be generalized quite easily to learning mutants characterized by more complex rules correlating multiple mutations.

Protein-protein interactions can be seen as a hierarchical process occurring at three related levels: proteins bind by means of specific domains, which in turn form interfaces through patches of residues. Detailed knowledge about which domains and residues are involved in a given interaction has extensive applications to biology, including better understanding of the binding process and more efficient drug/enzyme design. Alas, most current interaction prediction methods do not identify which parts of a protein actually instantiate an interaction. Furthermore, they also fail to leverage the hierarchical nature of the problem, ignoring otherwise useful information available at the lower levels; when they do, they do not generate predictions that are guaranteed to be consistent between levels.Inspired by earlier ideas of Yip et al. (BMC Bioinformatics 10:241, 2009), in the present paper we view the problem as a multi-level learning task, with one task per level (proteins, domains and residues), and propose a machine learning method that collectively infers the binding state of all object pairs. Our method is based on Semantic Based Regularization (SBR), a flexible and theoretically sound machine learning framework that uses First Order Logic constraints to tie the learning tasks together. We introduce a set of biologically motivated rules that enforce consistent predictions between the hierarchy levels.We study the empirical performance of our method using a standard validation procedure, and compare its performance against the only other existing multi-level prediction technique. We present results showing that our method substantially outperforms the competitor in several experimental settings, indicating that exploiting the hierarchical nature of the problem can lead to better predictions. In addition, our method is also guaranteed to produce interactions that are consistent with respect to the protein-domain-residue hierarchy.

The progress in mapping RNA-protein and RNA-RNA interactions at the transcriptome-wide level paves the way to decipher possible combinatorial patterns embedded in post-transcriptional regulation of gene expression.Here we propose an innovative computational tool to extract clusters of mRNA trans-acting co-regulators (RNA binding proteins and non-coding RNAs) from pairwise interaction annotations. In addition the tool allows to analyze the binding site similarity of co-regulators belonging to the same cluster, given their positional binding information. The tool has been tested on experimental collections of human and yeast interactions, identifying modules that coordinate functionally related messages.This tool is an original attempt to uncover combinatorial patterns using all the post-transcriptional interaction data available so far. PTRcombiner is available at http://disi.unitn.it/~passerini/software/PTRcombiner/.

High-throughput X-ray absorption spectroscopy was used to measure transition metal content based on quantitative detection of X-ray fluorescence signals for 3879 purified proteins from several hundred different protein families generated by the New York SGX Research Center for Structural Genomics. Approximately 9\% of the proteins analyzed showed the presence of transition metal atoms (Zn, Cu, Ni, Co, Fe, or Mn) in stoichiometric amounts. The method is highly automated and highly reliable based on comparison of the results to crystal structure data derived from the same protein set. To leverage the experimental metalloprotein annotations, we used a sequence-based de novo prediction method, MetalDetector, to identify Cys and His residues that bind to transition metals for the redundancy reduced subset of 2411 sequences sharing <70\% sequence identity and having at least one His or Cys. As the HT-XAS identifies metal type and protein binding, while the bioinformatics analysis identifies metal- binding residues, the results were combined to identify putative metal-binding sites in the proteins and their associated families. We explored the combination of this data with homology models to generate detailed structure models of metal-binding sites for representative proteins. Finally, we used extended X-ray absorption fine structure data from two of the purified Zn metalloproteins to validate predicted metalloprotein binding site structures. This combination of experimental and bioinformatics approaches provides comprehensive active site analysis on the genome scale for metalloproteins as a class, revealing new insights into metalloprotein structure and function.

MetalDetector identifies CYS and HIS involved in transition metal protein binding sites, starting from sequence alone. A major new feature of release 2.0 is the ability to predict which residues are jointly involved in the coordination of the same metal ion. The server is available at http://metaldetector.dsi.unifi.it/v2.0/.