Our Take
AI accelerates discovery by finding candidates faster, but it optimizes for potency and selectivity, not developability—pushing sponsors toward solubility, bioavailability, and manufacturability problems that slow everything down later.
Why it matters
As pharmaceutical discovery speed increases, development bottlenecks shift from a luxury problem to a cost multiplier. Teams that don't integrate formulation strategy into early screening will waste API, delay programs, and burn capital on late-stage rework.
Do this week
Development: audit your early-phase screening criteria this week to identify which legacy assumptions are now filtering out viable candidates from AI platforms, then rebuild the intake checklist around dissolution, absorption, and stability before chemistry prioritizes.
AI discovers faster; formulation lag widens
Investment in AI-driven drug discovery platforms grew steadily through the first half of 2026, with new deals focused on small molecule discovery and broader R&D capability buildout across the pharmaceutical sector. The acceleration in silico screening has not made the molecules easier to develop. As AI expands chemical space and surfaces more complex candidates, sponsors encounter a familiar problem with higher stakes: poor solubility, limited bioavailability, stability constraints, and manufacturability issues that determine whether an asset can move forward.
The core issue is structural. Generative AI platforms optimize for target potency and selectivity, which biases their output toward higher molecular weight, higher logP (lipophilicity), and reduced aqueous solubility. A molecule may screen strong in early assays but become difficult once formulation scientists begin building a stable dosage form, generating meaningful exposure, or designing a manufacturing process that scales beyond the bench.
Greater discovery output means more programs encounter the same developability problems, but later. When discovery velocity exceeds development readiness, the cost of getting development wrong compounds: faster supply of candidates means more API wasted on poor formulation strategy, more revisions to manufacturing processes, and more program delays.
Development decisions now determine which discoveries matter
The bottleneck has moved upstream. Early development no longer functions as a supporting service that validates what discovery handed off; it must now determine whether a strong molecule can move forward under realistic timelines and cost constraints.
Three questions now demand answers much earlier in the program: Can the molecule reach therapeutic exposure at a practical dose? Can it remain stable through processing and storage? Is there a manufacturing path that holds up as the program advances? These determine where delays emerge, where material gets wasted, and where teams revisit work that should have been settled sooner.
Legacy screening assumptions amplify the risk. Many were built around a narrower range of compounds. Applied rigidly, they push sponsors to deprioritize molecules that may have a viable path forward with the right formulation strategy. In an AI-driven workflow, that is costly: some of the most promising candidates may also be the least likely to fit older ideas of what a developable molecule should look like.
Integrate formulation latitude into candidate prioritization
An integrated decision loop must anchor development, not prediction alone. AI can narrow options and improve prioritization, but it cannot show whether a candidate will hold up once formulation work begins. That still requires in vitro screening, in silico PBPK modeling, and in vivo confirmation.
Formulation flexibility becomes the real lever. As more AI-surfaced candidates fall outside the comfort zone of standard approaches, sponsors need expanded latitude to work around the molecule's properties. This includes bioavailability-enabling approaches (amorphous solid dispersions for poorly soluble compounds), processing flexibility (solvent-free fusion or nonstandard routes when heat or solvent constraints limit conventional paths), and broader formulation design space (greater excipient flexibility and multi-component systems).
Early development teams that can define practical limits around dissolution, absorption, and stability—then assess whether a formulation approach improves performance in a way that supports the program—generate stronger decisions, more efficient use of limited API, and fewer late-stage surprises. Development still decides what advances. AI will continue changing how molecules are discovered, but the candidates that generate the most interest in silico still have to succeed under real development conditions.