In silico: biosimulation, QSAR and AI

PBPK and QSP

Physiologically based pharmacokinetic (PBPK) models predict drug exposure. QSP models link mechanism to outcome. This is the most commercially proven NAM segment. Certara, owner of Simcyp, reported $418.8M of FY2024 revenue, and Simcyp became the first PBPK platform with an EMA qualification opinion in August 2025. Simulations Plus reported $79.1M of FY2025 revenue after cutting guidance.

Consolidation has been steady. Simulations Plus bought Immunetrics (2023). Certara bought Applied BioMath (2023) and Chemaxon (2024), then agreed to sell its writing business to focus on modeling (2026).

QSAR, read-across and expert systems

Rule-based and statistical models for endpoints such as mutagenicity have long been accepted, for example for mutagenic impurities under ICH M7. Instem, taken private by ARCHIMED in 2023, runs an in silico toxicology service.

AI toxicity and ADMET

Startups train models on proprietary human-relevant data:

• Axiom Bio: liver toxicity models trained on imaging of primary human hepatocytes, with a public dataset of 130,000+ molecules.
• Inductive Bio (Brooklyn): ADMET predictions trained on a precompetitive data consortium. Won the 2025 Polaris ADMET benchmark and the 2026 OpenADMET-ExpansionRx blind challenge. ARPA-H award of up to $21M for liver and heart toxicity models.
• Deep Origin: leads an ARPA-H consortium, up to $31.7M, building in silico and ex vivo models of drug safety and dosing.
• Quris-AI: models trained on patients-on-a-chip data.

In silico is already the largest category of NAM data in FDA drug submissions, about half by FDA's count.

Open questions

Data access, more than algorithms, separates players. Regulators will expect a clear context of use and performance on chemistry the model has not seen.