Summary

• Cambridge-based Potential Sciences has emerged from stealth with an eight-figure package involving financing and computing support, including investment from the UK's Sovereign AI Fund.
• The company is developing GPU-accelerated quantum chemistry software that it claims can reduce the cost of certain high-accuracy molecular calculations by a factor of 100.
• Access to the AI Research Resource, including Bristol's Isambard-AI, will support dataset generation and the development of atomistic foundation models.

Cambridge-based Potential Sciences has secured government investment and access to national supercomputing resources as it develops software intended to make high-accuracy molecular calculations affordable enough to generate training data for artificial intelligence in drug discovery.

The company emerged from stealth on 5 November with an eight-figure package of financing and computing support, including investment from the UK’s Sovereign AI Fund. Other backers include Twin Path Ventures, Nauta Capital, Odyssey, Portfolio Ventures, Blue Wire Capital, rule30, and Lightfield Equity.

The government is also providing access to the AI Research Resource, which includes the Isambard-AI supercomputer in Bristol. The value of Sovereign AI’s individual investment has not been disclosed, while the published information does not provide a complete breakdown separating cash financing from the value of computing support.

Potential Sciences is developing GPU-accelerated quantum chemistry software capable of producing data about molecular behaviour. It claims its implementation can reduce the computational cost of certain high-accuracy calculations by approximately 100 times, potentially allowing AI models to be trained on substantially larger collections of scientifically derived information.

Molecular modelling faces a data constraint

Drug discovery involves identifying molecules that interact with biological targets in useful ways without producing unacceptable effects elsewhere. Researchers can test candidate compounds experimentally, but laboratory work requires time, specialist resources, and expenditure on substances that may ultimately prove unsuitable.

AI systems can help identify potential candidates by learning relationships between molecular structures and their properties. However, the available training information is constrained by the experiments researchers have performed and the results they have chosen or been able to publish.

Computational chemistry offers another source of information by calculating molecular behaviour using mathematical descriptions of physical interactions. Such methods can generate data for molecules that have not yet been synthesised, although their usefulness depends on the accuracy and limitations of the underlying calculations.

High-accuracy quantum chemistry methods are computationally demanding. Simpler approximations can be applied to larger datasets at lower cost, but they may struggle with complex chemical interactions or fail to reproduce properties sufficiently accurately for demanding scientific applications.

Potential Sciences wants to make more rigorous calculations economical enough for use in constructing AI training datasets. Its approach centres on implementing established quantum chemistry techniques efficiently on modern graphics processing units, which can perform many numerical operations in parallel.

The company specifically identifies CCSD(T), a computational method used as a high-accuracy reference in quantum chemistry. It says its software can perform such calculations at dramatically lower cost than traditional implementations.

That claim concerns computational efficiency for particular methods and implementations. It does not establish that every molecular simulation will become 100 times faster, that all calculations will achieve experimental accuracy, or that drug discovery programmes will experience equivalent reductions in cost.

Training data can be generated from physical models

The company’s planned AI models will learn from data generated through computational physics rather than depending exclusively on historical laboratory measurements. Such information could expand the range of molecular structures represented in training data, including compounds for which experimental results are sparse or unavailable.

Potential Sciences intends to develop atomistic foundation models, systems designed to learn patterns in how atoms interact and use those representations to predict chemical behaviour.

Its first commercial applications are expected to involve small-molecule drug discovery, including lead optimisation, reaction prediction, and covalent drug design. These areas require different kinds of scientific predictions, although each can involve comparing candidate molecular structures before selecting experiments.

Lead optimisation seeks to improve compounds that already show promising activity. Researchers may change a molecule’s structure to improve its interaction with a target or reduce unwanted effects, but evaluating possible variations experimentally can be expensive.

Reaction prediction involves estimating how chemicals may interact under particular conditions. Covalent drug design adds another requirement because candidate compounds are intended to form chemical bonds with their biological targets, creating questions about selectivity and unwanted reactions.

Computational predictions could help prioritise which molecules deserve laboratory testing, although they cannot eliminate the need for experimental evidence. The behaviour of a molecule in a biological system depends on factors that may not be captured by a particular quantum chemistry calculation.

The company will therefore need to demonstrate that its generated data produces models that perform reliably on relevant scientific tasks, including compounds and conditions not adequately represented in their training sets.

GPU efficiency changes research economics

Modern AI training already relies heavily on graphics processors, but many scientific software packages were originally developed for different computing architectures. Reworking calculations to use GPUs efficiently can allow researchers to extract more scientific work from available hardware.

Potential Sciences founder and chief executive Hatem Helal said: “We rebuilt quantum chemistry for GPUs so that gold-standard accuracy is affordable at the scale AI needs.”

The company argues that its software could support the generation of very large molecular datasets without the computing budgets traditionally associated with high-accuracy calculations. Its backers identify the ability to produce data for millions of molecules as a longer-term objective.

The eventual value will depend on the range of molecules that can be modelled, the accuracy achieved, the cost of generating each calculation, and the performance of AI systems trained using the resulting data.

Lower computational cost could also allow researchers to investigate unsuccessful or less promising molecular candidates rather than concentrating only on compounds with existing experimental results. That may help models learn distinctions between potentially useful structures and those unlikely to justify further testing.

However, calculated information remains an approximation based on selected physical methods. Independent benchmarking and comparisons with experimental measurements will be necessary to establish where the new software offers a meaningful advantage.

National computing capacity supports development

The government’s Sovereign AI Fund is participating as an investor while also making computing resources available through the national AI Research Resource. This combination is intended to support technically demanding companies whose development requires substantial computing before commercial revenue becomes established.

Access to Isambard-AI could help Potential Sciences generate larger datasets and train its first foundation models without immediately financing an equivalent private computing installation.

Such support is particularly relevant to scientific AI companies, where development may require both specialist researchers and extensive processing capacity. Funding alone does not necessarily provide access to the hardware needed at the appropriate time.

Potential Sciences is based in Cambridge, with researchers in Bristol and Tübingen, Germany. Its team includes scientists and engineers with experience at research institutions and technology companies, including Cambridge, Caltech, CERN, Graphcore, Recursion, and Dayhoff Labs.

The company plans to use its new resources to expand the research team, increase the scale of its datasets, and train atomistic foundation models. Its longer-term ambitions extend beyond pharmaceuticals into catalysts, enzymes, and materials research.

Each field will require evidence that the underlying calculations remain suitable for the relevant chemical systems. A model developed for selected small molecules cannot automatically be assumed to perform equally well on substantially different structures or applications.

The government has not disclosed the precise amount invested through Sovereign AI, and the company’s financing should not be represented as a specific cash total without further information.

Potential Sciences has secured the resources to test whether improved computational efficiency can change the economics of scientific AI development. Its next evidence will come from the accuracy, cost, and practical usefulness of the molecular predictions produced by its software and trained models.