Why Better Biological Data Could Decide the Future of AI Drug Discovery

 

Artificial intelligence may be transforming pharmaceutical research, but powerful algorithms alone cannot discover the next generation of medicines.

The quality of the biological data used to train those systems could matter even more.

British pharmaceutical company GSK has entered into a research collaboration with biotechnology firm Relation Therapeutics worth up to $110 million, expanding their existing partnership in AI-assisted drug discovery.

Under the agreement, Relation Therapeutics will produce large-scale biological datasets showing how human cells respond to genetic modifications and drug interventions. These datasets will be used to train AI models designed to identify promising drug targets, including models within Relation’s MORGAN platform.

The collaboration highlights a major shift in AI-powered pharmaceutical research. Instead of relying only on existing databases, companies are increasingly investing in laboratory experiments that generate new, disease-specific biological information.

AI Needs More Than Existing Data

Many AI systems are trained using information gathered from public biological repositories.

Resources such as CZ CELLxGENE, the Human Cell Atlas, and the National Center for Biotechnology Information’s Gene Expression Omnibus provide researchers with access to massive collections of single-cell data.

CZ CELLxGENE alone reportedly contains more than 100 million standardised cells.

However, having access to more data does not automatically produce a better AI model.

Biological datasets are often created using different sampling methods, sequencing technologies, laboratory procedures, and processing pipelines. These differences can introduce technical noise, inconsistencies, and experimental artefacts.

The same or similar cellular samples may also appear across multiple databases. This can cause certain data to be overrepresented during AI training.

It may also create data-leakage problems when overlapping samples are used in both training and testing datasets, making a model appear more accurate than it really is.

For this reason, dataset selection, filtering, balancing, and quality control are becoming just as important as the design of the AI model itself.

Bigger Datasets Do Not Always Create Smarter Models

A study published in Nature Methods in June 2026 examined how the size and diversity of pretraining data affected single-cell foundation models.

Researchers used a dataset containing 22.2 million cells, trained 400 models, and conducted approximately 6,400 evaluation experiments.

The researchers found that many current single-cell foundation models reached performance plateaus after being trained on only a portion of the available data.

This differs from the scaling patterns commonly associated with large language models, where increasing the amount of training data and computing power can often lead to measurable improvements.

In biological AI systems, simply adding millions of additional cells did not consistently improve performance.

Researchers concluded that model capacity, dataset quality, computational resources, and biological diversity must be properly balanced.

Another study published in Genome Biology in 2025 evaluated Geneformer and scGPT, two foundation models developed for single-cell biological analysis.

The models did not consistently outperform simpler methods across several zero-shot evaluation tasks. Researchers also identified problems involving batch effects and warned against assuming that larger pretrained models automatically produce better representations of biological systems.

These findings do not mean that small or privately owned datasets are always better. They show that biological relevance and data quality may be more valuable than raw volume.

How Relation Generates New Biological Data

Relation Therapeutics uses what it calls a Lab-in-the-Loop approach, combining physical laboratory experimentation with computational analysis.

Its work includes:

• Human tissue profiling
• Single-cell and spatial transcriptomics
• Genetic sequencing
• Functional biological assays
• Target identification and validation
• Machine-learning-assisted experimental design

The company also conducts perturbation experiments, which measure how genetic changes affect cellular characteristics connected to disease.

The results are then analysed alongside human genetic information and patient-derived biological data.

This creates a continuous research cycle. AI helps researchers design experiments, laboratory experiments generate new data, and the resulting data are used to improve the AI models.

The approach allows researchers to move beyond simply analysing information that already exists.

Building Disease-Specific Data Assets

Relation has already applied this strategy to Osteomics, which the company describes as a proprietary functional single-cell atlas focused on bone biology.

The project uses patient-derived samples and combines:

• Single-cell and spatial omics
• Medical imaging
• Genomic information
• Proteomics
• Clinical phenotype data

Osteomics is being used to study osteoporosis, potential therapeutic targets, biomarkers, disease mechanisms, and different patient subgroups.

Hospitals and research organisations in the United Kingdom and Australia are participating in the observational study.

Research published in Nature Genetics in July 2026 also examined the cellular and genetic factors behind skeletal disease using single-cell analysis, human genetic information, and functional validation. Several researchers from Relation Therapeutics contributed to the study.

Pharmaceutical Companies Are Investing in Specialised Data

The GSK and Relation partnership is part of a wider pharmaceutical industry trend.

A 2025 analysis published in Nature Biotechnology identified specialised biological dataset providers as an emerging force in AI-focused biopharmaceutical partnerships.

High-quality, disease-specific datasets are becoming increasingly valuable for causal and generative machine-learning models.

GSK previously entered into a separate agreement with Ochre Bio worth approximately $37.5 million. The agreement involved access to human liver single-cell data and information generated from perfused human organs.

In another major partnership, AstraZeneca and Pathos AI entered into a $200 million agreement with Tempus in 2025.

Under the arrangement, Pathos was expected to develop oncology foundation models using de-identified clinical, genomic, and imaging information from more than 150,000 patients.

These agreements show that pharmaceutical companies are no longer competing only for the best AI platforms.

They are also competing for access to the most useful biological datasets.

Data Access Remains a Major Bottleneck

Access to reliable and relevant training data remains one of the biggest obstacles in AI drug discovery.

Pharmaceutical companies hold valuable proprietary information, but privacy rules, intellectual-property concerns, commercial restrictions, and incompatible data systems can limit how that information is shared.

Federated learning may help address some of these limitations by allowing organisations to train shared AI models without transferring all of their sensitive data into one central database.

However, technical, regulatory, and organisational challenges remain.

As a result, AI-biopharma agreements now take several different forms. Some focus on access to AI platforms, while others involve joint research, data licensing, laboratory experimentation, or the creation of completely new biological datasets.

The GSK and Relation agreement combines both data generation and model development.

Relation will produce human cellular datasets and use them to train AI systems capable of identifying and validating potential drug targets.

The Real Competitive Advantage

The future of AI drug discovery may not be decided by which company builds the largest model.

It may be decided by which company can produce the most accurate, diverse, clinically relevant, and experimentally validated biological data.

For the Philippines, this development also presents an important lesson.

Filipino universities, hospitals, biotechnology researchers, and healthcare institutions should not focus only on adopting foreign AI platforms. The country must also invest in building secure and ethically governed biological datasets that represent Filipino genetics, disease patterns, environmental conditions, and treatment outcomes.

Without locally representative data, medical AI systems may fail to fully understand the biological realities of Filipino patients.

In AI-powered healthcare, data are not merely fuel for an algorithm.

They are the scientific foundation on which every prediction, drug target, and potential treatment depends.