Andreu Vall

Philipp Seidl, Andreu Vall, Sepp Hochreiter et al.

Activity and property prediction models are the central workhorses in drug discovery and materials sciences, but currently they have to be trained or fine-tuned for new tasks. Without training or fine-tuning, scientific language models could be used for such low-data tasks through their announced zero- and few-shot capabilities. However, their predictive quality at activity prediction is lacking. In this work, we envision a novel type of activity prediction model that is able to adapt to new prediction tasks at inference time, via understanding textual information describing the task. To this end, we propose a new architecture with separate modules for chemical and natural language inputs, and a contrastive pre-training objective on data from large biochemical databases. In extensive experiments, we show that our method CLAMP yields improved predictive performance on few-shot learning benchmarks and zero-shot problems in drug discovery. We attribute the advances of our method to the modularized architecture and to our pre-training objective.

Markus Hofmarcher, Andreas Mayr, Elisabeth Rumetshofer et al.

Due to the current severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) pandemic, there is an urgent need for novel therapies and drugs. We conducted a large-scale virtual screening for small molecules that are potential CoV-2 inhibitors. To this end, we utilized "ChemAI", a deep neural network trained on more than 220M data points across 3.6M molecules from three public drug-discovery databases. With ChemAI, we screened and ranked one billion molecules from the ZINC database for favourable effects against CoV-2. We then reduced the result to the 30,000 top-ranked compounds, which are readily accessible and purchasable via the ZINC database. Additionally, we screened the DrugBank using ChemAI to allow for drug repurposing, which would be a fast way towards a therapy. We provide these top-ranked compounds of ZINC and DrugBank as a library for further screening with bioassays at https://github.com/ml-jku/sars-cov-inhibitors-chemai.

Matthias Dorfer, Jan Schlüter, Andreu Vall et al.

Cross-modality retrieval encompasses retrieval tasks where the fetched items are of a different type than the search query, e.g., retrieving pictures relevant to a given text query. The state-of-the-art approach to cross-modality retrieval relies on learning a joint embedding space of the two modalities, where items from either modality are retrieved using nearest-neighbor search. In this work, we introduce a neural network layer based on Canonical Correlation Analysis (CCA) that learns better embedding spaces by analytically computing projections that maximize correlation. In contrast to previous approaches, the CCA Layer (CCAL) allows us to combine existing objectives for embedding space learning, such as pairwise ranking losses, with the optimal projections of CCA. We show the effectiveness of our approach for cross-modality retrieval on three different scenarios (text-to-image, audio-sheet-music and zero-shot retrieval), surpassing both Deep CCA and a multi-view network using freely learned projections optimized by a pairwise ranking loss, especially when little training data is available (the code for all three methods is released at: https://github.com/CPJKU/cca_layer).

Simon Deltadahl, Andreu Vall, Vijay Ivaturi et al.

We present a novel framework for synthesizing patient data with complex covariates (e.g., eye scans) paired with longitudinal observations (e.g., visual acuity over time), addressing privacy concerns in healthcare research. Our approach introduces controlled association in latent spaces generating each data modality, enabling the creation of complex covariate-longitudinal observation pairs. This framework facilitates the development of predictive models and provides openly available benchmarking datasets for healthcare research. We demonstrate our framework using optical coherence tomography (OCT) scans, though it is applicable across domains. Using 109,309 2D OCT scan slices, we trained an image generative model combining a variational autoencoder and a diffusion model. Longitudinal observations were simulated using a nonlinear mixed effect (NLME) model from a low-dimensional space of random effects. We generated 1.1M OCT scan slices paired with five sets of longitudinal observations at controlled association levels (100%, 50%, 10%, 5.26%, and 2% of between-subject variability). To assess the framework, we modeled synthetic longitudinal observations with another NLME model, computed empirical Bayes estimates of random effects, and trained a ResNet to predict these estimates from synthetic OCT scans. We then incorporated ResNet predictions into the NLME model for patient-individualized predictions. Prediction accuracy on withheld data declined as intended with reduced association between images and longitudinal measurements. Notably, in all but the 2% case, we achieved within 50% of the theoretical best possible prediction on withheld data, demonstrating our ability to detect even weak signals. This confirms the effectiveness of our framework in generating synthetic data with controlled levels of association, providing a valuable tool for healthcare research.

Shreyan Chowdhury, Andreu Vall, Verena Haunschmid et al.

Emotional aspects play an important part in our interaction with music. However, modelling these aspects in MIR systems have been notoriously challenging since emotion is an inherently abstract and subjective experience, thus making it difficult to quantify or predict in the first place, and to make sense of the predictions in the next. In an attempt to create a model that can give a musically meaningful and intuitive explanation for its predictions, we propose a VGG-style deep neural network that learns to predict emotional characteristics of a musical piece together with (and based on) human-interpretable, mid-level perceptual features. We compare this to predicting emotion directly with an identical network that does not take into account the mid-level features and observe that the loss in predictive performance of going through the mid-level features is surprisingly low, on average. The design of our network allows us to visualize the effects of perceptual features on individual emotion predictions, and we argue that the small loss in performance in going through the mid-level features is justified by the gain in explainability of the predictions.

Andreu Vall, Gerhard Widmer

Music recommender systems have become a key technology supporting the access to increasingly larger music catalogs in on-line music streaming services, on-line music shops, and private collections. The interaction of users with large music catalogs is a complex phenomenon researched from different disciplines. We survey our works investigating the machine learning and data mining aspects of hybrid music recommender systems (i.e., systems that integrate different recommendation techniques). We proposed hybrid music recommender systems based solely on data and robust to the so-called "cold-start problem" for new music items, favoring the discovery of relevant but non-popular music. We thoroughly studied the specific task of music playlist continuation, by analyzing fundamental playlist characteristics, song feature representations, and the relationship between playlists and the songs therein.

Andreu Vall, Massimo Quadrana, Markus Schedl et al.

The automated generation of music playlists can be naturally regarded as a sequential task, where a recommender system suggests a stream of songs that constitute a listening session. In order to predict the next song in a playlist, some of the playlist models proposed so far consider the current and previous songs in the playlist (i.e., the song context) and possibly the order of the songs in the playlist. We investigate the impact of the song context and the song order on next-song recommendations by conducting dedicated off-line experiments on two datasets of hand-curated music playlists. Firstly, we compare three playlist models, each able to consider a different song context length: a popularity-based model, a song-based Collaborative Filtering (CF) model and a Recurrent-Neural-Network-based model (RNN). We also consider a model that predicts next songs at random as a reference. Secondly, we challenge the RNN model (the only model from the first experiment able to consider the song order) by manipulating the order of songs within playlists. Our results indicate that the song context has a positive impact on the quality of next-song recommendations, even though this effect can be masked by the bias towards very popular songs. Furthermore, in our experiments the song order does not appear as a crucial variable to predict better next-song recommendations.

Andreu Vall, Matthias Dorfer, Markus Schedl et al.

Automated music playlist continuation is a common task of music recommender systems, that generally consists in providing a fitting extension to a given playlist. Collaborative filtering models, that extract abstract patterns from curated music playlists, tend to provide better playlist continuations than content-based approaches. However, pure collaborative filtering models have at least one of the following limitations: (1) they can only extend playlists profiled at training time; (2) they misrepresent songs that occur in very few playlists. We introduce a novel hybrid playlist continuation model based on what we name "playlist-song membership", that is, whether a given playlist and a given song fit together. The proposed model regards any playlist-song pair exclusively in terms of feature vectors. In light of this information, and after having been trained on a collection of labeled playlist-song pairs, the proposed model decides whether a playlist-song pair fits together or not. Experimental results on two datasets of curated music playlists show that the proposed playlist continuation model compares to a state-of-the-art collaborative filtering model in the ideal situation of extending playlists profiled at training time and where songs occurred frequently in training playlists. In contrast to the collaborative filtering model, and as a result of its general understanding of the playlist-song pairs in terms of feature vectors, the proposed model is additionally able to (1) extend non-profiled playlists and (2) recommend songs that occurred seldom or never in training~playlists.

Andreu Vall, Hamid Eghbal-zadeh, Matthias Dorfer et al.

Automated music playlist generation is a specific form of music recommendation. Generally stated, the user receives a set of song suggestions defining a coherent listening session. We hypothesize that the best way to convey such playlist coherence to new recommendations is by learning it from actual curated examples, in contrast to imposing ad hoc constraints. Collaborative filtering methods can be used to capture underlying patterns in hand-curated playlists. However, the scarcity of thoroughly curated playlists and the bias towards popular songs result in the vast majority of songs occurring in very few playlists and thus being poorly recommended. To overcome this issue, we propose an alternative model based on a song-to-playlist classifier, which learns the underlying structure from actual playlists while leveraging song features derived from audio, social tags and independent listening logs. Experiments on two datasets of hand-curated playlists show competitive performance compared to collaborative filtering when sufficient training data is available and more robust performance when recommending rare and out-of-set songs. For example, both approaches achieve a recall@100 of roughly 35% for songs occurring in 5 or more training playists, whereas the proposed model achieves a recall@100 of roughly 15% for songs occurring in 4 or less training playlists, compared to the 3% achieved by collaborative filtering.