In 2013, we showed how molecule-cell line datasets can be modelled with machine learning. We also demonstrated that machine learning can predict drug synergy on cancer cell lines as long as data inconsistencies between screening centres are taken into account. We also found evidence suggesting that estimating the uncertainty of individual predictions will be an effective way to enhance virtual screening on cancer cell lines. Ongoing research also reveals that graph neural networks outperform other algorithms.

Discovering novel antibiotics for human pathogens is a priority to face the challenge posed by the growing antimicrobial resistance trend. We thus investigate chemistry-informed AI models to screen ultralarge libraries for novel antibiotics against major human pathogens.

This is important to reveal the targets of phenotypic hits or better understand the polypharmacology and mechanism of action of drugs with already some known targets. We proposed a benchmark specific to target prediction and explain why the widespread use of virtual screening benchmarks is inadequate. These issues arise from not having the same objective: target prediction aims at finding the targets bound by a molecule, while virtual screening aims at finding the molecules binding to a target. We have also shown that reliability prediction boosts performance and provided a webserver implementing the method, which we are currently updating.

We developed USR, an ultrafast method to identify molecules with similar 3D shape to a template with activity for the target. We later provided a webserver to carry out virtual screening using this and related methods. We also made a hardware implementation of USR, which is even faster and more energy-efficient than these software implementations at the cost of being less versatile.

Also known as structure-based virtual screening, it is an effective way to discover potential drug leads from their docked poses against the 3D structure of the target. This is typically posed as a binary classification problem and the resulting models are called scoring functions. Virtual screening is hard due to extreme class imbalance and data biases. We showed how machine learning and data augmentation can improve scoring functions for virtual screening. Strong progress in machine learning for structure-based virtual screening has been made in recent years, also in the prospective applications of these methods. Scoring functions built with machine learning, instead of the classical ones assuming linearity between features and binding strength, provide higher hit rates while identifying molecules with higher potency too.

We still cannot reliably estimate the prospective performance of a method from retrospective analysis. We have contributed to this effort by explaining how to reduce decoy bias or providing a user-friendly protocol to build and benchmark machine-learning scoring functions (In Press). Rigorous in-depth analyses are also important to make progress on this key problem.

Not optimal for virtual screening, as non-binders are practically not considered. However, these structure-based models can be useful for optimising the potency of known small-molecule binders against a target. Having the least confounding factors, this type of data leads to the most robust results. In 2010, we discovered that machine-learning scoring functions are substantially more predictive than classical scoring functions at this task. This performance gap grows with larger training sets, which is still true when one removes test complexes with similar molecules or targets to those in the training set. We also found that expert-based selection of features, the innovation engine of classical scoring functions, is actually suboptimal for machine-learning scoring functions.

Machine-learning scoring functions as a part of holistic efforts to find other uses for existing drugs, for instance as anti-aging treatments.

Cancer drugs are only effective in a fraction of patients. We thus need to predict which patients will be resistant to a given drug. Such prediction is based on the presence of a somatic mutation (mutation marker) or the a priori combined expression of expert-selected genes (gene expression signature), but considering other genes, profiles and model-building approaches improves the prediction. Especially in terms of recall. A promising alternative is using AI to predict how patients respond to a drug from the multiple molecular profiles (multi-omics) of their tumours. This AI approach has so far provided enhanced prediction in paclitaxel-treated breast cancer patients, doxorubicin-treated breast cancer patients, gemcitabine-treated pancreatic cancer patients and temolozomide-treated low-grade-glioma patients. Overall, we found that the best learning algorithm - molecular profile combination changes across drugs and cancer types and is also easy to miss as almost all other combinations are not predictive. This highlights the critical importance of considering many algorithms and profiles when analysing these clinical pharmaco-omic datasets.

Prospectively applying our developed methods is important as a reality check, understanding how to improve methods further and actually making biomedical discoveries. However, this requires finding a collaborator to carry out in vitro tests and also provide expertise on the target. We are hence keen to enter this type of collaborations and have been able to complete some over the years. These are some examples below (we do not include the prospective applications by others using our methods):

Screening over a billion make-on-demand compounds with our AI models has led to the identification of nanomolar inhibitors of colorectal and glioma cell lines (all with novel chemical scaffolds). In vitro tests are not yet completed.

The targets of these phenotypic hits will be determined with our target prediction method, which already discovered low-nanomolar targets of drugs such as mebendazole and actarit.

USR-based virtual screening has discovered novel low-micromolar inhibitors for a breast cancer target , a colorectal metastasis target and antibacterial targets, among others. This ligand-based approach was also successful when combined with a generic machine-learning scoring function on these antibacterial targets. We have also applied prospectively PARP1-specific machine-learning scoring functions, which have led so far to cell-active PARP1 small-molecule inhibitors as potent as 1 nanomolar.

On the other hand, we are keen on collaborating with clinicians in drug-based omics-informed precision oncology. The main obstacle to making an impact in this area is now clinical data sharing. Full data from two cohorts of patients with the same cancer type, with their tumours profiled with the same omics technology and treated with the same drug is hard to access. Collaborations providing such data to get our predictive models closer to the clinic is our priority. We would be also happy to consider other drug-cancer type cases for analysis if non-public data is available. Lastly, our AI protocols can be adapted to the analysis of longitudinal molecularly-profiled liquid-biopsy data for early detection of cancer. We are also keen to consider collaborators in this area.