Z. Parlak, P. MacNichol, D. French
QATCH Technologies,
United States
Keywords: viscosity, injection, subQ, developability
Summary:
High-concentration protein formulations (typically 100–200 mg/mL) are essential for subcutaneous (SC) delivery of modern biologics, yet their viscosities are notoriously difficult to predict. A single formulation can vary from 5 cP to >100 cP with small changes in buffer, pH, or excipients, creating manufacturing barriers and limiting which biologics can ever become SC injectable. Conventional viscometers require 40–1,000 μL per test, so viscosity is often measured only after scale-up, months after key development decisions have already been made. Up to 50% of biologics ultimately fail to achieve an injectable profile, forcing companies into IV formulations that increase cost and reduce patient access. We present visQ.AI, the first formulation-aware ML engine trained on experimental high-shear-rate viscosity data collected using the nanovisQ microfluidic viscometer, which requires only 4 μL per measurement and produces full shear-rate curves (100–15,000,000 s⁻¹) in <5 minutes. visQ.AI combines this microliter-scale data with a hybrid ML architecture designed for the extreme ‘small-data’ reality of protein formulation. Unlike sequence-only models, which require MD simulations or thousands of labeled viscosity points, visQ.AI learns from only 3–5 anchor formulations per protein and delivers predictive viscosity maps across concentration, buffer, excipient, and temperature spaces. Our current seed dataset includes >600 high-concentration formulations spanning monoclonal, FC-Fusion, and polyclonal antibodies, multiple buffer systems (PBS, acetate, histidine), sugars (sucrose/trehalose), surfactants (Tween-20/80), and excipients (arginine, proline). In internal benchmarks, visQ.AI achieves ≤20% global MAPE when anchored with 3–5 nanovisQ measurements (few-shot mode). Technically, visQ.AI integrates physics-informed feature space with experiment-derived viscosity anchors, meta-learning and gradient-boosted meta-models designed for sparse experimental updates. visQ.AI is local, on-premise retraining, enabling secure use on proprietary formulations without cloud upload. This work demonstrates that high-concentration viscosity, which is a key developability attribute, can be predicted with practical accuracy using sparse, real-world experimental data.