July 2026
A new paper co-authored by ForecomAI Chief Technology Officer Prof. Miroslaw Bober shows how modern AI can turn cell-based imaging into quantitative biological readouts.
Rabies virus is a demanding test case: infection usually produces no visible cytopathic effect, so the morphological cues exploited by many AI methods are unavailable. The study instead analyses fluorescence from viral antigen in infected cell cultures.
The best-performing model reached 99.6% accuracy at well level. Applied without retraining to the virus neutralisation assay recommended by the World Health Organization (WHO) and the World Organisation for Animal Health (WOAH), it returned antibody titres matching those of the expert operator across the proficiency panel.
The wider point is the underlying problem: resolving weak but meaningful signals while remaining robust to artefacts and to biological and experimental variability. That balance is what makes microscopy and cell-based assays usable as quantitative measurements, and it is the direction we are pursuing with our partners across cellular imaging, morphology and phenotypic analysis in disease biology, toxicology and drug discovery.
The work was carried out in collaboration between the Animal and Plant Health Agency (APHA) and the University of Surrey.
June 2026
How long should anticoagulation continue after venous thromboembolism (VTE)? Continuing treatment prevents recurrence but causes bleeding, so the decision requires both risks to be quantified for the individual patient, and requantified as time on treatment accrues.
Two companion papers co-authored by ForecomAI were accepted in the Journal of Thrombosis and Haemostasis (JTH). One develops a dynamic prediction model for recurrent VTE (rVTE), the other for clinically significant bleeding (CSB). Both draw on over 20 years of UK real-world healthcare data, covering more than 50,000 patients and over 200,000 person-years of follow-up, and both account for the competing risk of death.
Each model uses a point-based score built from routinely recorded patient characteristics, and each prediction decomposes into the factors contributing to it. A clinician can therefore see why a patient scores as they do rather than accept an unexplained output.
Predicted risk varied widely. First-year recurrence risk ranged from below 1% to above 20%, and first-year bleeding risk from 0.6% to 18%. VTE patients do not separate cleanly into low-risk and high-risk groups; predicted risk lies on a broad continuum.
For us the significance is methodological. Extracting reliable evidence from real-world data (RWD) depends on handling competing risks, time-varying exposure and censoring correctly, and on producing outputs a clinician can interrogate. Those requirements hold whether the data are images, longitudinal patient records, or both.
The work was a collaboration between National Health Service (NHS) clinicians, academic researchers, industry partners and the ForecomAI team.
April 2025
Our paper, Bridging Self-Supervision and Mechanism of Action Discovery in Morphological Profiling, in collaboration with the University of Surrey, has been accepted for the 1st CVPR Workshop on Computer Vision For Drug Discovery (CVDD) to be held in Nashville.
It introduces TRex — a task-guided framework that enhances the biological relevance of morphological embeddings for mechanism of action classification. TRex doubles performance, and improves compound recognition and generalisation — all without retraining large models.
This underscores the value of lightweight, task-aware adaptation for unlocking the full biological relevance of morphological profiling.
August 2024
ForecomAI’s work predicting which patients with venous thromboembolism may benefit from an extended anticoagulation 3 months post diagnosis, with collaborators from Pfizer and Bristol Myers Squibb among others, was presented at the European Society of Cardiology Congress in London.
A Cohen, S Choudhuri, I Khan, K Pollock, M Bober-Irizar, Z Galias, A Irizar, M Bober, A novel dynamic risk score to predict venous thromboembolism (VTE) recurrences after 3 months of anticoagulation in patients with incident VTE and without active cancer, European Heart Journal, Volume 45, Issue Supplement_1, October 2024, ehae666.2214, https://doi.org/10.1093/eurheartj/ehae666.2214
May 2023
AI developed in the UK is the world leader in identifying the location and expression of proteins. Press release from University of Surrey.
May 2023
Our work on the development of the advanced AI system, HCPL (Hybrid subCellular Protein Localiser) for single-cell image analysis for protein localisation, is published in Communications Biology of the Nature Portfolio.
This technique can rapidly identify the subcellular localisation of proteins of interest in images, streamlining high throughput screening for various applications including the study of tumour heterogeneity, identification of disease and toxicity biomarkers, and drug discovery.
May 2021
ForecomAI scientists won silver medal in a global competition to design single-cell AI analysis method that can automatically determine subcellular distribution of proteins from microscopy images. The competition was based on the Human Protein Atlas (https://www.proteinatlas.org/) and over 700 teams competed worldwide on the ML platform Kaggle.
Automated cell analysis methods are essential to support the ongoing revolution in biology, which promises to advance understanding of how human cells function, how diseases develop and how to cure them.
October 2020
ForecomAI has embarked in a collaboration with a pharmaceutical alliance to undertake machine learning methodology to develop and evaluate a novel risk prediction algorithm for the cardiovascular risk of pharmacological treatments using CPRD real world data.