In traditional drug discovery, typical timelines for successful drugs stretch from 10 to 15 years, or sometimes even longer, to bring a drug from discovery to approval. In addition to that, the cost of bringing a drug to market can range from hundreds of millions to multiple billions of dollars. In fact, the inflation-adjusted cost of drug development has roughly doubled every nine years.
Credit: Soley Therapeutics
A major reason for clinical failure is lack of efficacy. An analysis of clinical trial data from 2010 to 2017 showed that of the drugs that fail in clinical trials, 40% to 50% do so due to lack of efficacy, revealing how poorly preclinical assays reflect a drug’s biology.
The AI drug discovery field has attracted investment and optimism, but clinical validation remains limited and mixed. In 2024 and 2025, the sector attracted 612 venture rounds and approximately $19.9 billion in total capital, according to DealForma’s sector review. Despite this, AI has not demonstrably improved the 90% clinical failure rate.
The snapshot assay problem
One source of clinical trial failure is the structural limitations of conventional drug discovery assays. Current methods reduce complex, evolving cellular behavior into static measurements. Additionally, cells are destroyed during transcriptome profiling, making it hard to track the dynamic gene expression of an individual cell at multiple time points.
These limitations have caused a resurgence of interest in phenotypic drug discovery, approaches that observe complex cellular behavior rather than modulating a single target. However, these approaches have their own challenges, including hit validation, target deconvolution and difficulty translating phenotypic signals into mechanistic insight.
Live cell dynamics
Live Cell Dynamics (LCD) is a self-supervised machine learning pipeline that extracts dose- and time-dependent cellular state information from continuous brightfield images without stains or labels. This method was published in a January 2026 Scientific Reports paper by scientists from Soley Therapeutics.
“By treating cellular response as time-resolved information rather than a static snapshot, LCD enables mechanism classification, compound comparison, and detection of complex biology through measurable trajectories,” explained Kurosh Ameri, co-founder and CSO of Soley Therapeutics. “This provides early forward-looking biological signal rather than a late binary readout, shifting drug discovery from observing damage to forecasting a drug’s direction and future impact.”
The study pre-trained on 189 compounds and evaluated performance on 81 additional held-out compounds across 10 mechanisms of action, using a single human osteosarcoma cell line (U2OS). LCD outperformed cell count and CellProfiler-based feature extraction for phenotypic activity detection across all doses and time points, with the largest advantages at early time points and low doses. Incorporating multiple doses and time points incrementally improves mechanism-of-action classification, disentangling mechanisms that appear similar at late stages.
“Learned representations from LCD preserved signal in those early regimes and performed strongly across dose and time, while the CellProfiler baseline tended to be comparable only later, or lower at early time points,” said Ameri.
Many drugs affect multiple biological targets simultaneously, a phenomenon called polypharmacology that is common but notoriously difficult to detect. Identifying it conventionally requires extensive, costly assay panels. Using only brightfield imaging, the model flagged both Aurora kinase and JAK inhibitor activity that was consistent with prior studies that required extensive kinome profiling to reach the same conclusion.
“Brightfield is difficult because the signal is subtle, not evident to the naked eye, contrast is low, and small changes in optics, focus, plate position, or day-to-day setup can create batch effects that swamp biology,” said Ameri.
The paper outlines two training innovations that help LCD overcome these challenges: plane-agnostic augmentation, which teaches the model to recognize biology rather than focus-plane artifacts; and cross-batch sampling, which forces the model to learn features stable across experimental runs, separating biological signal from technical noise.
The results demonstrate that “LCD can represent compound behavior as a profile across dose and time, not a single label. Those profiles contain enough structure to separate closely related mechanisms and expose mixed activity, which is exactly the kind of complexity that shows up in development,” said Ameri.
The study used a single, well-characterized cancer cell line under laboratory conditions, meaning LCD’s performance in primary cells, patient-derived organoids or disease-relevant models remains unknown. Whether the performance advantages observed in a controlled compound library will hold up across the messier, more heterogeneous biology of disease models is the central question the work leaves open.
According to Soley, the next step is to expand LCD to additional cell types, including primary and disease-relevant models, broader mechanism coverage and prospective use in active drug programs. LCD will need to be validated in settings closer to human disease before claims about its clinical impact can be fairly evaluated.
Filed Under: AI Meets Life Sci
