AI drug discovery models need more than accurate predictions of how molecules interact with biological targets. They also need to identify compounds with properties that make them suitable for further development, including solubility and permeability. Boltz PBC is partnering with Excelra to strengthen that part of its small-molecule discovery pipeline. The collaboration brings Excelra’s curated GOSTAR™ chemical and biological data into the ADMET models used by BoltzMol-1, helping the system screen out compounds with unfavourable properties before researchers spend resources purchasing or synthesising them.
Curated Chemical Data Strengthens BoltzMol-1’s Screening Pipeline
The partnership focuses on three measurements from GOSTAR: kinetic solubility, lipophilicity and Caco-2 permeability. These properties help researchers assess how a candidate molecule may behave in biological systems. A compound that appears promising in a target-binding prediction can still be unsuitable for further development if it dissolves poorly or has unfavourable permeability characteristics.
BoltzMol-1 uses the resulting ADMET models as an early triage layer. ADMET stands for absorption, distribution, metabolism, excretion and toxicity. By incorporating these predictions into the discovery workflow, Boltz aims to identify compounds that are not only predicted to interact with a target but also have more favourable physicochemical and absorption-related properties.
This distinction matters in computational drug discovery. Predicting whether a molecule binds to a protein is only one part of identifying a useful starting point for drug development. Researchers must also consider whether the molecule has characteristics that justify further laboratory testing and optimisation.
Why Proprietary Data Matters for AI Drug Discovery
One of the persistent challenges in AI-driven drug discovery is the quality and coverage of training data.
Public chemical databases provide valuable information about molecular structures, biological activity and experimental measurements. However, datasets can vary in completeness, consistency and experimental context. Models trained on limited or unevenly curated data may struggle when presented with chemical structures that differ substantially from those represented in their training sets.
Excelra’s GOSTAR database is designed to address some of these limitations. Its small-molecule collection contains more than 10.6 million chemical structures and 35 million bioactivity records, covering approximately 79,000 biological targets. The company says its records undergo manual standardisation and a three-tier quality-control process conducted by medicinal chemistry specialists.
For machine learning developers, the value is not simply the volume of records. Structured measurements with consistent annotations can help models learn relationships between chemical structures and properties without requiring researchers to spend as much time cleaning and reconciling source data.
Boltz is using selected GOSTAR measurements to train its models for solubility, lipophilicity and permeability. The collaboration therefore addresses a specific part of the drug discovery pipeline rather than replacing the broader molecular modelling architecture behind BoltzMol-1.
BoltzMol-1 Moves From Prediction Toward Experimental Validation
BoltzMol-1 is Boltz’s small-molecule hit-discovery pipeline, built around an optimised version of its Boltz-2 model. It is designed to prioritise candidate molecules for experimental testing, using model-based ranking and access to purchasable or make-on-demand compound libraries.
The platform’s earlier technical work reported prospective experiments across ten challenging biological targets, most of which had limited representation in the model’s underlying affinity training data. Boltz reported confirmed binders or active compounds for six targets, using relatively small experimental screening sets.
The company’s results are notable because prospective validation tests whether a model can identify useful compounds in new experiments, rather than simply reproduce patterns in previously collected datasets.
The partnership announcement also says the GOSTAR-trained ADMET models matched or outperformed a leading open-source ADMET package on strictly out-of-distribution chemistry, with the largest improvements in solubility and permeability.
That finding is relevant because chemical structures outside a model’s training distribution can be particularly difficult to predict accurately. Better generalisation could help researchers evaluate less familiar molecular structures and avoid spending experimental resources on candidates with unfavourable properties.
However, these findings should be understood as reported model and experimental results, not evidence that the system has discovered approved medicines or established clinical effectiveness.
Reducing the Cost of Early-Stage Screening
Traditional high-throughput screening can involve testing thousands or even millions of compounds against a biological target. Although these campaigns remain important in pharmaceutical research, they can require significant laboratory resources, time and expenditure.
AI-assisted virtual screening offers another route. Models can prioritise a smaller set of candidates for experimental validation, allowing researchers to focus laboratory work on molecules that appear more promising computationally.
Boltz’s technical work is an example of this approach. Rather than beginning with an exhaustive experimental search, its pipeline ranks candidate molecules and combines affinity predictions with developability-related filtering.
The Excelra partnership adds more curated experimental data to the property-prediction component of that workflow. The aim is to reduce the likelihood that researchers advance compounds that may show promising predicted binding but have poor solubility or permeability.
The commercial value will depend on whether those predictions consistently improve experimental hit selection across different target classes and chemical spaces. Smaller screening campaigns are useful only when the models maintain sufficient accuracy and do not systematically filter out potentially valuable compounds.
Data Partnerships Become Part of the AI Drug Discovery Stack
The collaboration reflects a broader trend in computational biology: AI drug discovery companies increasingly depend on specialised datasets alongside model architecture and computing resources.
Foundation models and molecular prediction systems can help identify patterns across complex biological and chemical information. But their performance remains closely tied to the quality, diversity and relevance of the data used to train and evaluate them.
This creates an opportunity for specialist data providers such as Excelra. Its GOSTAR database offers structured chemical and biological measurements that can be accessed through its application, APIs and downloadable datasets. The company positions the database as a resource for medicinal chemistry, cheminformatics and AI-based drug discovery. Excelra
For Boltz, the partnership strengthens a pipeline that already combines molecular modelling with experimental validation. For Excelra, it demonstrates how curated scientific datasets can become inputs to modern AI development rather than serving only as reference resources for researchers.
The next challenge is to establish how much the additional data improves model performance across independent targets, especially when the chemistry differs substantially from previous training examples.
AI is unlikely to eliminate the experimental work required to discover and develop medicines. But combining molecular prediction, high-quality chemical data and targeted laboratory testing could make early discovery campaigns more efficient.
The Boltz-Excelra collaboration highlights why data quality is becoming a competitive factor in AI-driven drug discovery: better models need not only more computing power, but also reliable experimental knowledge about the molecules they are designed to evaluate.
Market Landscape
AI-driven drug discovery is moving toward integrated platforms that combine molecular prediction, chemical data, property modelling and experimental validation.
Three developments are shaping this market:
- Data quality as a competitive advantage: Curated chemical and biological datasets are becoming important inputs for training models that must generalise beyond familiar molecular structures.
- Developability-aware screening: Drug discovery models are increasingly evaluating properties such as solubility, permeability and toxicity alongside predicted target binding.
- Prospective validation: Pharmaceutical AI companies are placing greater emphasis on laboratory experiments that test predictions against new targets and previously unseen chemistry.
The Boltz-Excelra partnership sits at the intersection of these trends. Boltz contributes molecular modelling and hit-discovery capabilities, while Excelra supplies structured experimental data for ADMET model training.
Other players in the AI drug discovery ecosystem are also combining models with specialised datasets, computational chemistry and laboratory workflows. Companies such as Recursion and Insilico Medicine are pursuing integrated approaches to computational drug development.
The competitive question is increasingly whether these systems can produce experimentally validated starting points more efficiently and reliably, rather than simply achieving strong retrospective prediction scores.
Top Insights
- Boltz is integrating Excelra’s GOSTAR data into BoltzMol-1’s ADMET models for kinetic solubility, lipophilicity and Caco-2 permeability.
- The models act as an early screening layer, helping researchers filter compounds with unfavourable properties before purchase or synthesis.
- GOSTAR contains more than 10.6 million chemical structures and 35 million bioactivity records, according to Excelra.
- Boltz reported prospective hit-discovery results across six of ten challenging targets, demonstrating the importance of experimental validation.
- The partnership highlights the growing role of curated scientific datasets in improving AI model performance beyond familiar training chemistry.
Power Tomorrow’s Intelligence — Build It with TechEdgeAI
