2026 CMP Industrial Projects

Genes are the basic units of heredity and encode the information for the synthesis of proteins and other molecules that perform various functions in living organisms. Understanding the relationships between genes and their functions is a fundamental challenge in biology and medicine. One way to approach this challenge is to represent genes as numerical vectors, also known as embeddings, that capture some aspects of their biological properties and interactions. Embeddings can be derived from various sources of data, such as gene sequences, gene expression, gene ontology, protein-protein interactions, and literature. Embeddings can then be used for various tasks, such as gene clustering, gene function prediction, gene-disease association, and gene pathway analysis.

The project is part of a broader and strategically critic goal for AI applied to computational biology in R&D in Novo Nordisk, which relates to the use, validation and control of gene embeddings for novel target and biomarker discovery and functional contextualization.

Aim
The aim of this project is to explore how to define embeddings for genes and how they relate to each other, and to evaluate which embeddings are most useful and which data sources should be included. The specific objectives are:

• To review the existing methods and tools for generating gene embeddings from different data sources
• To compare and contrast different types of gene embeddings, such as sequence-based, expression-based, ontology-based, interaction-based, and literature-based. Of particular interest will be recent work on augmenting large language models with domain-specific tools such as database utilities for more precise access to specialized knowledge (e.g. GeneGPT),
• To apply and test different gene embeddings on various biological analysis tasks, such as gene function prediction, gene-disease association gene pathway analysis and drug target prediction.

Methodology
The methodology of this project consists of the following steps:

• To query existing public and internal databases for the general characterization of a gene including
  • Which tissue/cell type is the gene expressed in.
  • Which pathways is it part of and which other genes are in that pathway
  • Which diseases is it known to be associated to?
  • What is the interaction network of this gene in a particular tissue?
  • Has the gene been previously explored?
  • Is there patent data and/or human clinical data for the gene?
  • What assay, cell-type and conditions should be used for validation? 
• To select and implement the appropriate methods and tools for generating gene embeddings from the different data sources, (eg. word2vec, doc2vec, autoencoders, graph neural networks) and with particular emphasis on transformers / large language models to represent literature information.
• To evaluate and compare the quality and performance of different gene embeddings on various biological analysis tasks, such as gene function prediction, gene-disease association, gene pathway analysis and drug target predictions, using appropriate metrics and benchmarks.
• Of particular interest, is the downstream comparison of the a-priori based information embeddings to the gene embeddings of cellular genetic perturbation in vitro imaging screening assays that Novo Nordisk has inhouse.

Expected Outcomes 

• A comprehensive review of the existing methods and tools for generating gene embeddings from different data sources.
• A comparative analysis of the different types of gene embeddings and their applications on various biological analysis tasks.
• A critical evaluation of the strengths and limitations of different gene embeddings and data sources, and suggestions for possible improvements and extensions.
• An extremely valuable comparison of the a-priori embeddings to the embeddings Novo Nordisk has inhouse from our perturbation assays.

The implications of this project are:

• To provide a better understanding of the relationships between genes and their functions, and to facilitate the discovery of new biological insights and hypotheses.
• To contribute to the advancement of the field of gene embeddings and their applications in biology and medicine.
• To demonstrate the potential and challenges of applying natural language processing and machine learning techniques to biological data.

A virtual cell is an in-silico model (a kind of “digital twin”) that lets us predict how a living cell will respond to interventions (e.g. adding a drug). This sits in a fast-moving area of BioAI, with community benchmarks such as the Virtual Cell Challenge [1] pushing models toward realistic generalisation settings.

At the core of the virtual cell is biological perturbation prediction: given a baseline cell state, predict how the cell changes after an intervention (e.g., knocking out a gene). Conceptually, this is a causal effect problem, made hard by biological confounding, incomplete measurements, and distribution shift (new cell types, new perturbations, new experimental settings).

Our recent work (“LangPert”, ICLR 2025 workshop spotlight [2]) suggests a practical path forward: use LLMs to retrieve and synthesise mechanistic biological context (gene function, pathways, interactions, etc.) and condition predictive models on that context. The key benefit is zero-shot or low-data generalisation to perturbations the model has not seen during training, while also producing explanations that are at least partially aligned with known biology. In parallel, LLM-powered causal analysis motivates a causality-first approach to virtual cells [3, 4].

The project will explore LLM-informed causal modelling for perturbation prediction. The high-level aim is to use LLM knowledge as a contextual guidance—not as a replacement for data—so that models can better infer gene–gene relationships and predict outcomes of interventions, especially when faced with novel perturbations or shifted experimental conditions.

1. Context building with LLMs: Retrieve and summarise biologically relevant information for a given perturbation. Explore different ways of representing the context.
2. Causal / intervention-aware prediction: Combine LLM-derived context with state-of-the-art models for intervention prediction and causal effect estimation (e.g., Do-PFN [5]). Investigate how LLM context should enter the model (e.g., as an uncertainty-aware DAG prior), and whether identification-consistent in-context exemplars (baseline and singles) are sufficient for generalising to unseen combinations.
3. Generalisation and robustness: Evaluate performance on held-out perturbations and under distribution shift (e.g., new cell types). Run ablations and compare outcomes to causal identification expectations to measure when context helps vs. hurts.
4. Interpretability (if time): Assess whether model rationales and retrieved context align with known biology and identify failure modes (e.g. hallucinated context).

Because the AI landscape changes quickly the specific LLM, retrieval approach, and causal estimator will be updated at the project start to reflect the best available options.

Successful outcome:

• A reproducible experimental pipeline (datasets, baselines, evaluation splits, and ablations).
• A short technical report summarising findings and recommendations.
• If results are strong and pass internal review, the work may be included in an AI workshop paper submission.

Primary Objectives:

• Develop a robust methodology for aggregating order books across multiple liquidity pool
• Create an information content classification system that quantifies the predictive value of different order book levels
• Design metrics to identify which price levels contain the most relevant information for mid-price determination
• Build a real-time nowcasting model for current market prices based on aggregated order book data

Secondary Objectives:

• Analyze the relative importance of different liquidity pools in price discovery 
• Investigate how information content varies across different market conditions (high/low volatility, different trading sessions)
• Assess the temporal stability of information content metrics

Information Content Metrics:

• Apply information theory measures (entropy, mutual information) to quantify price level importance
• Implement machine learning techniques to identify patterns in order book informativeness
• Develop weighted scoring systems based on historical price impact and predictive accuracy

Validation Framework:

• Backtest the classification system against historical price movements
• Compare nowcasting accuracy against benchmark models
• Conduct out-of-sample testing to ensure robustness

Applications

• Algorithmic trading strategy optimization
• Risk management and position sizing
• Market making and liquidity provision strategies

APEX Horticulture Ltd. is a professional research and development business, offering bespoke testing services for cut flowers and plants. APEX has three purpose-built testing centres in the UK and US. APEX is a division in the wider MM group, where the primary business, MM Flowers, is one of the UK’s leading cut flower importer/processing companies, with vertically integrated ownership model and innovative practices. More recently, the MM group has diversified its activities, including supplying plants, bulbs and other gifting products to the retailers in the UK and Europe. MM is owned by the AM Fresh Group, a leading breeder, grower and distributor of citrus and grapes; Vegpro, East Africa’s largest flower and vegetable producer; and Elite, based in South America, and the leading flower grower globally.

APEX is at the optimal position in the chain, able to deliver high quality, independent research and close-to-market proximity matched with the invaluable insight into the true performance of flowers and plants subjected to actual supply chain conditions. The infrastructure and specialised personnel of APEX aims to deliver robust, standardised and consistent research every week of the year, together with the ability to undertake large scale projects to match all client requirements, influencing all elements of the cut flower supply chain.

APEX undertakes many different research projects covering the entire supply chain, from development of new flower types through to the manufacturing requirements for the final bouquets. Each of these projects generates a significant amount of data and insight, which is used to provide recommendations to the various stakeholders of each project.

Introductions to Lean:
https://alexkontorovich.github.io/2025F311H/
https://adam.math.hhu.de/

Agentic Theorem Prover (One of Many):
HILBERT: RECURSIVELY BUILDING FORMAL PROOFS WITH INFORMAL REASONING
https://arxiv.org/pdf/2509.22819

MM Flowers operates in a highly time-sensitive fresh flower supply chain, where product quality, availability and freshness are critical drivers of customer satisfaction. Even small inaccuracies in forecasting or delays in product movement can result in extended stock dwell time, increased waste at retailer level and ultimately higher customer complaint volumes. Roses, in particular, represent a high-volume and high-visibility product category where performance variability can have a significant commercial and reputational impact.

Forecast accuracy directly influences ordering decisions, inbound volumes and stock allocation. When actual arrivals deviate from forecasted volumes, this can lead to either stock shortages, impacting service levels, or excess stock levels, increasing dwell time and the risk of quality deterioration. Longer dwell times at MM Flowers or retail stores can accelerate deterioration, contribute to retailer waste and negatively affect the end consumer experience.

This project focuses specifically on yellow and white 40cm and 50cm roses, which are core SKUs within the MM Flowers portfolio and are particularly sensitive to demand variability and shelf-life constraints. By analysing the relationships between forecast accuracy, stock dwell time, retailer waste and customer complaints for these products, the project aims to identify whether measurable correlations exist across the supply chain.

The project is interesting and valuable because it connects operational planning decisions with downstream quality outcomes and customer feedback. Understanding these relationships will support more data-driven forecasting, stock management and waste-reduction strategies, while also providing insight into how supply chain performance ultimately affects customer satisfaction. The findings have the potential to inform targeted improvements for key rose lines and contribute to broader continuous improvement initiatives within MM Flowers.

This project will investigate whether measurable correlations exist between forecast accuracy, stock dwell time, retailer waste and customer complaints for yellow and white 40cm and 50cm roses supplied by MM Flowers. The project is primarily data-driven and quantitative, making it well suited to a student with strengths in mathematics, statistics or data analysis.

The project will begin with a data familiarisation and definition phase, during which the student will work with historical supply chain data provided by MM Flowers. This will include forecast volumes, actual arrival quantities, stock dwell time metrics, retailer waste data and customer complaint records. The student will be responsible for defining and calculating appropriate performance measures, such as forecast accuracy metrics (e.g. absolute error or percentage error), dwell time distributions and waste rates.

In the next phase, the student will apply statistical and mathematical techniques to explore relationships between variables. This may include:

• Descriptive statistics to summarise trends and variability;
• Correlation analysis to quantify the strength and direction of relationships between forecast accuracy, dwell time, waste and complaints;
• Regression analysis to assess the relative impact of each variable on customer complaints;
• Time-series or lag analysis to investigate delayed effects, such as whether longer dwell times or excess stock in one period lead to increased complaints in subsequent periods.

The project is open-ended in nature, allowing findings from the initial analysis to guide deeper investigation. For example, if strong correlations are identified for certain rose lengths or colours, the student may focus further analysis on those segments or explore threshold effects where performance begins to deteriorate significantly.

A successful outcome would be:

• A clear, evidence-based assessment of whether and how forecast accuracy and dwell time influence retailer waste and customer complaints;
• Identification of key drivers or risk indicators that are most strongly associated with complaints;
• Practical, data-backed insights that MM Flowers could use to improve forecasting, reduce waste and enhance customer satisfaction for core rose lines.

The project is interesting and useful because it links mathematical analysis directly to real-world operational and commercial outcomes. Students will gain experience applying statistical methods to complex, imperfect industry data, while MM Flowers will benefit from improved understanding of how planning and stock decisions affect product quality and customer experience across the supply chain.

The student will work independently on the core analytical aspects of the project, with regular guidance and supervision from the project lead at MM Flowers. The project will be based within a business and operational environment rather than a laboratory, giving the student exposure to real-world supply chain data and decision-making contexts.

In addition to the primary supervisor, the student will have opportunities to engage with forecasting, supply chain planning and quality teams at MM Flowers, allowing them to discuss data definitions, operational processes and practical implications of their findings. While there is no formal academic research group on site, the student will be supported through regular check-ins and access to subject matter experts across the business.

Working hours will be flexible, aligned with standard office hours, and can be adjusted to accommodate academic commitments. The project can be conducted in a hybrid format, combining remote analytical work with occasional on-site days at MM Flowers when beneficial for data access, collaboration and project reviews.

Day-to-day work will primarily involve data analysis, modelling and interpretation, with time allocated for meetings, progress reviews and refinement of the analytical approach. The student will be encouraged to manage their own time, structure their analysis and propose next steps, mirroring the autonomy expected in both industry and academic research roles.

This working environment offers a balance of independent mathematical problem-solving and practical business engagement, providing a supportive setting for a mathematics student to apply theoretical skills to a real operational challenge.

The project will build an end-to-end workflow that ingests option-chain quotes, produces a stable implied volatility representation across strike and maturity, and generates diagnostics for quality and stability. The student will implement a robust construction method, and assess sensitivity to data quality, filtering, and weighting choices. A PCA decomposition will then be built from a historical time series of model outputs evaluated on a fixed strike–maturity grid.

Successful outcome:

• Fitting implied volatility surfaces under different modelling assumptions
• Designing and evaluating quote weighting schemes for surface fitting (e.g., bid/ask and, liquidity/volume/open-interest based weights, spread- and vega-weighting, robust loss functions to downweight outliers and stale quotes).
• Run surface quality controls including fit-to-bid/ask diagnostics and static arbitrage checks, with clear metrics and flagging.
• Build a PCA decomposition from a historical time series of model outputs evaluated on a fixed strike–maturity grid, extract interpretable factors such as level, skew, and curvature-type moves, and produce factor time series.
• A well-documented workflow and reproducible codebase.

In a fast-moving consumer goods environment, it is vital that safety assessments are conducted to ensure products are safe for humans and the environment. Historically, these assessments have required in vivo animal testing and so there is a pressing ethical and scientific need to develop of non-animal methods to support product safety risk assessment. For more than 20 years, Unilever’s Safety, Environmental and Regulatory Science (SERS) Group has been developing novel in silico and in vitro methods, which leverage recent advances in biology, genetics, computing, mathematics and statistics, to conduct safety assessments without the use of animal testing. [1, 2].

The current evolution in the risk assessment paradigm, presents new opportunities in terms of applying new deep learning and AI-based approaches. A key part of risk assessment is to characterise the potential effects that a chemical may have on different cell types, which would typically involve using high throughput transcriptomics (HTTr) to measure the transcriptional response of cells to different concentrations of a test chemical. Such data can be expensive to generate, particularly if it needs to be generated for multiple chemicals and cell types. Therefore, the use of data-driven modelling which maximises the utility of all the available data is a high priority to achieve cost effective implementation of non-animal approaches.

Many approaches in risk assessment of an unknown compound are based on determining the similarity to a known compound. Recent advances in deep learning have introduced powerful methods for generating embeddings, numerical representations that capture complex relationships between entities. By combining transcriptional and chemical information (such as structural representations like molecular fingerprints), these embeddings may provide valuable insights beyond traditional similarity metrics, even being able to predict responses for unseen chemicals [3, 4].

Embeddings map complex (high-dimensional) data into a simplified (lower-dimensional) latent space, while preserving semantic relationships. Thereby enabling the discovery of relationships between datasets previously masked by the high dimensionality. Internal studies of these methods for biochemical in vitro responses have yielded promising results which we seek to further apply to relevant datasets.

For this project, the student(s) should start by familiarising themselves with literature surrounding this topic, upskilling around the use of representation learning /embedding models, transcriptomics and toxicology – with support from SERS experts. With this foundation, we recommend that the student progress this work by choosing one or more of the following paths;

1. Methods: We are interested in the further exploration of literature to identify, implement and compare methods with those already covered, and the application of these methods to internal datasets. However, the student(s) will not be limited to using pre-built models – any produced model capable of producing high quality embeddings that can be used for measuring biological similarity are considered a success.
2. Evaluation & Interpretability: Assessing the predictive capability of different methods can be difficult due to the black-box nature of these deep learning based approached. Testing embedding based approaches and ensuring the scientific justification of results is a key feature in the furthering the use of data driven approaches in all scientific research. The student would therefore endeavour to assess the robustness of any developed methods within a framework that allows for scientific interpretability and understanding where possible.
3. Optimisation: As deep learning models of this nature require large volumes of data, it is commonplace to use broad, public datasets that may differ from the specific target scenarios of interest. Therefore, further downstream modifications and optimisation should be applied to fine-tune the models. Methods such as transfer learning could be applied to apply the models to specific risk assessment problems. For example, modelling differences in transcriptomics data methodologies (e.g. RNA-Seq vs Temp-O-Seq) or post-training optimisation based on internally labelled risk classifications.

Aims:
For this specific project, we are first looking into internalizing transcriptomics foundation models of our choice. Then, we are aiming to fine-tune them, using our internal, high-performance AI training infrastructure. Fine-tuning tasks shall be prediction of cell niches and cell-cell communication, both in a broader tissue context. To predict cell niches, we are interested in cell densities, cell type composition, source tissue and more. For cell-cell communication, we are interested in key signaling molecules influencing the different cell types in a niche.

Key Tasks: 

• Set up the foundation models on our infrastructure.
• Prepare training, validation and test datasets, e.g. explore different tokenization strategies, summary statistic...
• Explore potential fine-tuning architecture (e.g. Multitask MLP, dedicated loss functions for our data of interest), and training paradigms for the different downstream tasks.
• Fine-tune foundation model, to achieve state of the art performance.
• Evaluate model performance against relevant benchmarks and baselines.

What summer students will learn:

• Exposure to biomedical foundation models.
• Exposure to large scale foundation model training/fine-tuning infrastructure.
• Exposure to biomedical data landscape and cancer biology.
• Get to know a cutting-edge, AI-first research setting in a large, pharmaceutical enterprise.
• New foundation model architecture design and model evaluations.
• Coding skills, mostly Python (PyTorch, JAX).
• Exposure to using coding assistants in a research environment.
• Potential opportunity to publish in a scientific journal.

Single cell foundation model references:
Tahoe-x1: https://doi.org/10.1101/2025.10.23.683759
STATE: https://doi.org/10.1101/2025.06.26.661135
STACK: https://www.biorxiv.org/content/10.64898/2026.01.09.698608
VariantFormer: https://www.biorxiv.org/content/10.1101/2025.10.31.685862v1

Niche detection:
CellLetter: https://academic.oup.com/bib/article/26/6/bbaf693/8405044
NicheFormer: https://www.nature.com/articles/s41592-025-02814-z

Required/useful skills:

• Basic Python skills (PyTorch, JAX would be a plus).
• Basic understanding of neural network architectures.
• Interest in applying AI to problems in biology.
• Interest in machine learning and statistics.
• Interest in linear algebra and optimization.
• Fits students who would like to follow their own ideas and also those who like a more engineering focus.
• Enthusiasm for foundation models and and interest in biomedical research questions