Transparent and Responsible AI Lab (TRAIL)

Chakraborty Group / TRAIL

TRAIL is a small team of early career researchers shared between the UCL Cancer Institute and Alan Turing Institute (the UK's National Institute for AI).

It is led by Dr Tapabrata Rohan Chakraborty, a Principal Research Fellow at UCL Cancer Institute and a Theme/Group Lead in Health AI at the Alan Turing Institute.

The long term research vision / mission of TRAIL is to answer, at least in part, the question: "What is Cancer that AI may know it, and what is AI that it may know Cancer?"

Those interested in the historical development of AI will realise the reference to the famous quote from Warren McCulloch (proponent of the perceptron model for artificial neurons).

Moving from the long term philosophical focus to a more concrete short term focus, the current North Star of TRAIL can be summarised by this research agenda.

Theme 1: Multimodal AI with Cancer Imaging + Omics

In this theme, we look at pan-cancer integration of digital pathology images and omics features in different ways using multimodal AI.

One way is to extract clinically relevant features from digital pathology images using vision transformer based foundation models, combine these with transcriptomics based features via co-attention and then take a predictive decision pertaining to cancer diagnosis (grading, sub-typing) or prognosis (staging, survival risk).

Another way is to use state of the art generative AI models like stable diffusion to synthesize transcriptomic factors and genomic expressions from digital pathology images either at bulk or spatial level. While histopathology images are routinely collected in the clinic and are the current gold standard for cancer diagnostics, transcriptomic data (bulk or spatial) are rarely collected in the public healthcare system due to high cost, despite having huge potential for personalised treatment and precision medicine.

If the insights from in-silico transcriptomes translate to better individualised treatment selection and patient outcome, then this would have a transformative health impact.

Featured publications

• Dey S, Banerji CRS ... Chakraborti T. Generating crossmodal gene expression from cancer histology predicts multimodal decision utility. arXiv:2502.00568, 2025. (Under review in Nature Communications) [Code]
• Paul D, Saha S, Basu S, Chakraborti T. Computational analysis of pathogen-host interactome for fast and low-risk in-silico drug repurposing in emerging viral threats like Mpox. Sci Rep. 2024 Aug 12;14(1): 18736. [Code]

Theme 2: Transparent AI with Concept Bottleneck Models

The high impact potential of multimodal cancer AI is evident from Theme 1, but for such powerful AI systems to be widely adopted in the clinic and reach their full transformative potential, they need to be trusted given the high risk nature of healthcare applications concerning patient lives. However, these deep learning based AI systems are known to be complex and opaque, for such systems to be translated to the clinic, the AI black box must be opened.

There are many approaches to develop transparent AI models. One method we are particularly interested in are concept bottleneck models, where clinically established concepts for a particular task are learnt along with the predictive output to ensure that the decision is clinically interpretable. One additional advantage of such models is that they provide the user (here, clinician) the opportunity to intervene on the concepts if wrong and thus provide a clinician-in-the-loop feedback to improve the system over time.

Furthermore, one can add an unknown concept which might help to discover new clinically relevant information in an exploratory manner. This would enable AI systems to align better with principles of responsible and transparent AI as being mandated by many emerging legislations in AI safety around the world.

Featured publications

• Banerji C, Chakraborti T, MacArthur B, Legally compliant clinical AI must learn to receive help. Nature, 2025 [accepted, in press]
• Baptista A, Barp A, Chakraborti T, Harbron C, MacArthur BD, Banerji CRS. Deep learning as Ricci flow. Sci Rep. 2024 Oct 8;14(1): 23383. [Code] [Blog]
• Mitra R, McGough SF, Chakraborti T, et al. Learning from data with structured missingness. Nature Machine Intelligence 5 (2023), 13-23. [Blog]

Theme 3: Responsible AI with Conformal Uncertainty Quantification

Having transparency in clinical AI systems is a necessary but not sufficient condition. In addition to the decision process being explainable and the decision itself being accurate at a population level, it also needs to have a high level of confidence or certainty for each individual. If not, that uncertainty needs to be carefully quantified and bounded, so that the clinician can make the decision to take or leave the AI system's suggestion. At the end of the day, accountability will lie with the human expert using the AI.

We are developing conformal prediction based methods to quantify uncertainty by providing marginal coverage guarantees as well as patient level bounds. A simple way to understand would be cancer subtype classification, where conformal calibration at a certain level of significance alpha would provide the minimal set of predictions for each test sample in the inference stage to ensure that the correct prediction is included in the conformal set, thus the length of the set of predictions itself is a quantification of predictive uncertainty for that patient.

This of course has strong implications for AI based predictive models in personalised healthcare and precision medicine. If the AI system provides a wrong suggestion with high certainty, that would have the potential of causing patient harm, thus breaking the Hippocratic oath of "first do no harm".

Featured publications

• Chakraborti T, Banerji C, et al. Personalised Uncertainty Quantification. Nature Machine Intelligence, 2025 [accepted, in press].
• Banerji CRS, Chakraborti T, Harbron C, MacArthur BD. Clinical AI tools must convey predictive uncertainty for each individual patient. Nat Med. 2023 Dec;29(12): 2996-2998.

Group Members

Current members of the Transparent and Responsible AI Lab (TRAIL), led by Dr Tapabrata Rohan Chakraborty, including those co-supervised through his research network.