AI Drug Discovery Lessons from Calico’s New Chapter

In 2013, Google launched Calico with a moonshot premise: treat ageing and the diseases that come with it. Twelve years later, death is still undefeated—but Calico isn’t “just research” anymore. According to a recent interview in Nature Reviews Drug Discovery, Calico has moved 5 candidates into the clinic, built a ~20-asset preclinical pipeline, and published hundreds of papers. That’s not a side project.

Here’s the twist that matters for anyone building in AI in pharmaceuticals and drug discovery: Calico is entering a phase where execution matters more than ambition. Its long-standing collaboration model (notably with AbbVie) helped it scale science and translational capability. Now, with a new head of drug discovery, Philip Kym, and a more independent posture, Calico’s next moves will look a lot like what every biotech eventually faces—choosing indications, proving mechanisms, designing trials that don’t collapse under noise, and making hard calls on portfolio focus.

If you’re leading R&D strategy, translational science, or AI/ML enablement in pharma, Calico is a clean case study: ageing biology is the toughest possible search space, and it exposes exactly where AI can help—and where it won’t—across target discovery, molecule design, biomarkers, and clinical development.

What Calico’s “go it alone” moment really signals

Calico’s shift toward greater independence signals one thing: it wants tighter control over discovery priorities, development decisions, and the operating cadence needed to turn longevity biology into medicines. Partnerships can accelerate early building—capabilities, scale, clinical expertise—but they also add interfaces, governance, and compromises.

A more independent Calico likely means sharper choices in three areas:

• Indication selection: “Ageing” doesn’t file an IND. Specific diseases do.
• Evidence standards: Ageing pathways can be compelling in mice and messy in humans.
• Translation discipline: Biomarkers, patient selection, endpoints, and trial duration become existential.

This matters because many AI-first drug discovery programs stall not on model performance, but on decision performance—how quickly teams can converge on tractable hypotheses, de-risk mechanisms, and design proof-of-concept trials.

Ageing is a portfolio problem, not a single target problem

Most companies get this wrong: they treat longevity like one program with one “master switch.” The reality is that ageing-related decline spans proteostasis, mitochondrial function, senescence, immune ageing, epigenetic drift, metabolic signaling, and tissue regeneration. Any serious longevity pipeline behaves like an oncology portfolio: multiple mechanisms, multiple risk profiles, and multiple translational strategies.

AI’s value in this context is less about predicting a single “correct” target and more about structuring uncertainty—ranking hypotheses, identifying patient subgroups, and guiding which experiments will change your mind fastest.

Where AI can actually accelerate anti-ageing drug discovery

AI speeds up anti-ageing research when it’s used to reduce iteration time across the full loop: hypothesis → experiment → learn → next hypothesis. In longevity biology, iteration time is usually the enemy because endpoints are slow and effects can be subtle.

Below are the highest-leverage AI use cases for organisations pursuing ageing and age-related diseases.

AI for target discovery: from “interesting biology” to testable mechanisms

The direct answer: AI can prioritize targets by connecting multi-omics signals to causal hypotheses and druggability constraints.

Longevity datasets tend to be high-dimensional (single-cell omics, proteomics, perturbation screens, longitudinal cohorts). AI helps by:

• Integrating multi-omics to identify pathways that change consistently with age across tissues
• Causal inference to separate correlation (age-related change) from likely drivers (intervenable causes)
• Knowledge graph approaches to connect genes, pathways, phenotypes, and existing compounds
• Druggability-aware ranking so the “top” targets aren’t all transcription factors with no clear modality

The stance I’ll take: if your target list doesn’t change after you add druggability and safety constraints, you’re doing biology exploration—not drug discovery.

AI for molecule design: shortening the “chemistry loop”

The direct answer: AI-driven molecule design is most valuable when paired with rapid experimental feedback, not when used as a one-shot generator.

For ageing-related targets—often novel, sometimes poorly validated—you want AI systems that optimize for:

• Potency and selectivity without inducing liabilities that derail chronic dosing
• ADME/PK for long-term administration (tolerability and safety margins matter more than in acute settings)
• Polypharmacology control when ageing biology suggests network effects but safety demands precision

In practice, successful teams treat generative models as a proposal engine and win by running tight build-test-learn cycles with:

1. Clear target product profiles (including chronic safety assumptions)
2. Predictive models calibrated on internal data (not just public benchmarks)
3. A medicinal chemistry team that trusts the system because it’s auditable

AI for biomarkers: the make-or-break layer in longevity trials

The direct answer: biomarkers are the bottleneck for ageing programs, and AI is the fastest way to make them usable.

Ageing trials face a brutal constraint: waiting for “hard outcomes” (mortality, major morbidity) is expensive and slow. So programs depend on biomarkers—but only the right kind.

AI helps in three ways:

• Composite biomarker construction: combining signals (proteomic panels, imaging, digital biomarkers) to reduce noise
• Surrogate endpoint discovery: finding measures that track with clinical outcomes strongly enough to justify development decisions
• Patient stratification: selecting cohorts most likely to respond based on molecular or phenotypic signatures

A practical rule: if your biomarker doesn’t change the trial design—enrichment, dosing, endpoints, duration—it’s a dashboard metric, not a development tool.

The hard part: clinical trials for ageing-related diseases

The direct answer: AI improves longevity trials by reducing variance and improving cohort selection, not by “predicting success.”

Anti-ageing programs often fail in the same predictable ways:

• Endpoints are too distal (you can’t afford the time horizon)
• Populations are too heterogeneous (signal disappears)
• Dosing is constrained by chronic tolerability (efficacy never shows)
• Mechanism is under-validated in humans (translation gap)

How AI optimizes trial design in practical terms

AI-enabled clinical development teams focus on decisions that cut time and increase statistical power:

• Eligibility criteria optimization: simulate inclusion/exclusion impacts on event rates and variability
• Site selection and recruitment forecasting: find sites with the right patient mix and reduce timeline risk
• Adaptive designs: use interim biomarker responses to adjust arms or enrichment strategies
• Synthetic control arms (when appropriate): reduce patient burden and cost, especially in rare or slow-progressing conditions

For ageing-related indications, the biggest win is often variance reduction. When effect sizes are modest, every bit of noise you remove is equivalent to adding budget.

The “go it alone” implication: owning the data flywheel

When a company runs more of the stack internally—discovery through early clinical—it also owns more of the data exhaust: negative results, dosing response curves, biomarker failures, subgroup insights.

That’s gold for AI.

• Better internal datasets improve model calibration.
• Better calibration improves decision quality.
• Better decisions create better experiments.

This flywheel is hard to build in fragmented collaborations where data access and standardization lag behind the science.

Actionable lessons for pharma teams building AI-enabled R&D

The direct answer: Calico’s trajectory reinforces that AI in drug discovery wins when it’s tied to operating discipline—portfolio strategy, translational planning, and measurable cycle-time improvements.

Here’s what I’d copy if I were building an AI program inside a pharma or biotech org.

1) Treat “ageing” as an engine for multiple indications

Don’t pitch ageing as one mega-asset. Build a map:

• 3–5 mechanisms you believe in
• 2–3 indications per mechanism (near-term and longer-term)
• A biomarker strategy for each indication
• A kill criteria document that leadership actually respects

2) Make your AI program accountable to cycle time

Model metrics are fine, but they don’t run a portfolio. Track:

• Time from hypothesis to first in vivo validation
• Time from hit to lead to candidate nomination
• Percent of programs with human translation evidence before IND

If cycle time isn’t improving, AI is probably sitting in a slide deck.

3) Build “translation-first” datasets

Longevity programs die in translation. Prioritize data collection that supports human relevance:

• Perturbation datasets with clear phenotypic readouts
• Human tissue or organoid systems that reflect aged biology
• Longitudinal biomarker measurements tied to clinical outcomes

4) Decide upfront what success looks like in Phase 1/2

For ageing-related diseases, Phase 1/2 should answer:

• Did we engage the target in humans?
• Did we move a mechanistically relevant biomarker?
• Did we see early clinical signal in an enriched subgroup?

AI helps most when these questions are operationalized into measurable thresholds.

People also ask: “Can we drug ageing itself?”

The direct answer: regulators approve drugs for diseases, not for ageing, so the practical path is treating ageing-related diseases with mechanisms that plausibly generalize.

That doesn’t mean the ambition is wrong. It means the development strategy has to be concrete:

• Pick indications with clear unmet need and measurable endpoints
• Use biomarkers to show mechanism and to justify broader claims later
• Design programs that can expand label over time, mechanism by mechanism

If longevity companies want credibility in 2026 and beyond, they’ll earn it in well-chosen indications—not in slogans.

What this means for the AI in Pharmaceuticals & Drug Discovery series

The direct answer: Calico’s next chapter is a stress test for AI-enabled drug discovery at scale—especially in biology where signal is slow and endpoints are hard.

Calico has already demonstrated persistence: a decade-plus runway, a real pipeline, and clinical assets. The “go it alone” posture increases the pressure to turn that foundation into repeatable development wins. That’s exactly where modern AI approaches—target ranking, generative chemistry, biomarker modeling, trial optimization—can pay off, but only if they’re integrated into how decisions get made.

If you’re building AI into your R&D org, take the hint: longevity isn’t a special case. It’s the clearest case. It forces you to connect models to outcomes, and outcomes to decisions.

If your team had to choose one ageing-related indication to prove your AI-enabled discovery engine in the next 18 months, what would you pick—and what data would you need to defend that choice?

If you want help pressure-testing an AI drug discovery roadmap (targets → molecules → biomarkers → trials), we can share a practical evaluation checklist we use to assess readiness and ROI in pharma R&D teams.