As governments and regulators across the UK, EU and beyond move to reduce reliance on animal testing, the need for reliable and scientifically robust in silico approaches has become increasingly important. For organosilicon chemistry, limited representation in existing models has traditionally constrained predictive confidence. Through targeted, pre-competitive data sharing between industry and independent scientific experts, recent work has expanded chemical space coverage, improved model performance, and strengthened the role of in silico tools in regulatory decision-making.
This work is also now documented in a newly published peer-reviewed paper, which explores the expansion of chemical space and performance of in silico models for organosilicon compounds in more depth:
Enhancing in silico prediction of Ames mutagenicity for organosilicon compounds and exploring chemical space boundaries
Addressing a recognised scientific gap
Organosilicon compounds present a well-recognised challenge for in silico toxicity prediction. Their unique chemistry and historically limited representation in traditional training sets mean they often sit outside the applicability domain of many (Q)SAR models, reducing prediction confidence for this class of substances.
At the same time, in silico approaches play an increasingly important role in supporting regulatory submissions for organosilicon compounds across frameworks such as REACH, cosmetics, food contact materials, and healthcare applications. This makes improving model coverage and reliability for this chemistry an important area of ongoing scientific collaboration.
A collaboration built on shared data and shared purpose
Building on this need, collaboration between Silicones Europe and Lhasa Limited has focused on strengthening predictive performance through targeted data sharing and model development.
At its core, this work is an example of what pre-competitive data sharing can achieve. Silicones Europe, a non-profit trade association representing major silicone producers across Europe, contributed carefully curated, industry-generated datasets to Lhasa Limited, enabling targeted improvements to both Derek Nexus and Sarah Nexus in areas where model coverage for organosilicon chemistry was previously limited.
This collaboration, and the outcomes it has enabled, are explored in detail in the newly published paper, which provides a systematic assessment of model performance and chemical space coverage for organosilicon compounds.
This kind of structured data donation – where industry expertise and experimental data are combined with independent scientific experience to improve tools that benefit the wider scientific community – reflects a model for collaboration that we believe will become increasingly central to the delivery of non-animal methods at scale.
Improvements to skin sensitisation and mutagenicity models
Recent updates to Derek Nexus include refinements to structural alerts for skin sensitisation and mutagenicity prediction, informed by donated organosilicon data. These improvements enhance the relevance and interpretability of predictions for this challenging chemical class and support their use in regulatory contexts such as REACH dossiers and industry weight-of-evidence assessments.
In parallel, Sarah Nexus has been strengthened through the integration of additional mutagenicity data generated through this collaboration, further improving model performance for Ames mutagenicity prediction within organosilicon chemical space.
Together, these enhancements reflect a shared objective: increasing confidence in predictions, improving transparency, and supporting the continued reduction of animal testing where scientifically appropriate.
Increasing confidence in non-animal approaches
Organosilicon compounds can be particularly challenging to assess for toxicity using conventional in vitro methods, making robust in silico approaches a critical component of integrated safety assessment strategies.
By expanding chemical space coverage and strengthening the underlying data foundation, this collaboration increases confidence in computational predictions and reinforces their role within weight-of-evidence approaches.
Importantly, improved structural alert coverage also enhances transparency, enabling users to better understand the basis of predictions and supporting more informed scientific and regulatory decision-making.
Supporting regulatory readiness and future submissions
While fully defined approaches for skin sensitisation are not yet universally applicable across all areas of organosilicon chemistry, continued improvements in model performance are helping to bridge this gap.
The enhanced predictive capability of Derek Nexus will support more robust weight-of-evidence submissions under frameworks such as REACH, enabling industry to reduce reliance on animal testing while maintaining regulatory confidence.
Benchmarking performance across organosilicon chemistry
To understand how in silico tools perform for organosilicon chemistry more broadly, the collaboration has also contributed to a systematic benchmarking exercise assessing 14 different in silico applications to predict Ames mutagenicity, including varying algorithms and training sets, using a curated dataset of mono-constituent organosilicon compounds.
This analysis provides a broad assessment of model performance across organosilicon chemical space, highlighting consistency and variability in predictions between approaches.
This work will be presented at the upcoming ESTIV conference (29 June – 2 July, Maastricht, Netherlands) in a poster titled:
“Enhancing in silico prediction of Ames mutagenicity for organosilicon compounds and exploring chemical space boundaries”
The poster, presented by Barbara Schmitt (Evonik), explores predictive performance using two curated datasets:
- Dataset A: 121 organosilicon compounds (including 50% alkoxysilanes, 17% siloxanes, 18% chlorosilanes, and 15% other silicon-based substances), with 12% classified as Ames mutagens
- Dataset B: 115 substances following removal of six compounds with discordant or biologically implausible results, with 8% mutagens
For the (curated) Dataset B, Derek Nexus and Sarah Nexus demonstrated strong and consistent performance across key metrics, including coverage, balanced accuracy, sensitivity, and specificity.
The organosilicon datasets have since been incorporated into the Sarah Nexus training set, further improving predictive capability within this chemical space.
What this means for regulatory practice
For organisations working with organosilicon compounds, improved in silico performance translates directly into more defensible regulatory submissions. Robust computational predictions can support weight-of-evidence assessments under REACH, strengthen safety dossiers for cosmetics and food contact materials, and in pharmaceutical and medical device contexts contribute to ICH M7-aligned impurity assessments and the evaluation of extractables and leachables from silicone-based materials.
More broadly, work of this kind demonstrates that with well-curated data and expert scientific collaboration, the applicability domain of non-animal approaches can be meaningfully extended, reducing reliance on animal testing without compromising the scientific confidence regulators require.
Get in touch
If you are working with organosilicon compounds and would like to understand how recent improvements to Derek Nexus and Sarah Nexus could support your regulatory submissions, or if you are interested in discussing data-sharing collaborations of this kind, please contact us to learn more.
Last Updated on June 9, 2026 by lhasalimited
