Derek Lowe has watched drug discovery technologies come and go for decades. His recent Science blog post cuts through the noise. It asks a simple question. So how is AI drug discovery doing, really?
The answer lands somewhere between guarded optimism and pointed skepticism. AI has delivered striking advances in protein structure prediction and molecule generation. Yet those tools have not yet rewritten the grim statistics of clinical failure. Short sentences. Long pauses. The field stands at a crossroads.
Consider the numbers. Traditional drug development stretches 12 to 15 years. Roughly 90 percent of candidates that reach Phase I never reach patients. AI was supposed to change that equation. Has it? Not obviously. Not yet.
Progress in the Lab Outpaces the Clinic
AlphaFold and its successors transformed how scientists see proteins. What once took years of crystallography now happens in hours. Generative models spit out novel chemical structures tuned to specific binding pockets. Virtual screening runs at scales once unimaginable.
But a Nature Reviews Drug Discovery perspective from August 2026 delivers a sobering line. “Evidence of their clinically relevant impact is, so far, disappointingly limited.” Andreas Bender and his co-authors, including Jack Scannell and David Shaywitz, know the territory. They call for benchmarks that test real decision-making. Not just model accuracy on clean datasets.
Data quality haunts the field. Chemical measurements prove noisy and inconsistent. ADMET predictions often fail prospective tests. Models overfit. They underspecify. Real biology refuses to sit still for neat algorithms. And yet companies keep pushing compounds forward.
Insilico Medicine’s rentosertib offers the clearest case study. This generative AI-designed TNIK inhibitor targets idiopathic pulmonary fibrosis. Phase 2a results in 2025 showed safety, tolerability, and early signals of improved lung function. Huspi’s analysis calls it the first generative-AI drug to reach mid-stage trials with encouraging data. A genuine milestone. One worth watching as it advances.
Other efforts follow. Recursion’s merger with Exscientia combines phenomics and chemistry automation. Schrödinger’s physics-based methods helped advance zasocitinib, now in Phase III. Nimbus originated the molecule. These examples show AI touching real pipelines. They do not yet prove faster approval or higher success rates across the board.
IQVIA’s Global R&D Trends 2026 report spots early signals. AI-enabled programs among emerging biopharma companies display stronger success rates. Reduced attrition could lift productivity. The analysis remains cautious. It describes a “tantalizing glimpse” rather than a proven transformation. 79 new active substances launched globally in 2025. Projections hold steady at 70 to 80 per year ahead. No explosion. Steady output.
Market forecasts paint a different picture. The AI-enabled drug discovery sector could grow from roughly $8.2 billion in 2026 to nearly $34 billion by 2036. A 15.3 percent compound annual growth rate, according to Future Market Insights. Other estimates place the 2025 base near $2.6 billion, expanding to $8-20 billion by 2030. Investors bet heavily. Valuations for AI-native biotechs run nearly double the industry median.
Big tech joins the fray. Anthropic launched an internal drug discovery program in June 2026 focused on neglected diseases. It released Claude Science tools for drugmakers. The move signals confidence. It also raises questions about who owns the resulting insights.
But Lowe’s post lingers on the gaps. No AI system yet mitigates the core risks of target selection or toxicology. Those failures drive most clinical attrition. Structure prediction helps. It does not replace decades of hard-won biological intuition. Generative chemistry produces candidates. It rarely predicts how those molecules will behave in a sick human body.
Fragment. The hype cycle continues. Fragment. Data remain the bottleneck. Fragment. Biology stays messy.
Recent articles echo this tension. A World Economic Forum piece from January 2026 outlines how AI reshapes target identification, compound generation, and safety prediction at firms like Novartis. It stresses human expertise remains essential. Amgen’s July 2026 commentary agrees. Technology accelerates insight. It cannot replace the scientist’s judgment.
A Nature perspective published just days before Lowe’s post recommends concrete steps. Move benchmarking beyond validation metrics. Focus on whether tools improve actual decisions. Generate better data through collaborative efforts like federated learning. Address biases. Prioritize context of use. These authors avoid both boosterism and blanket dismissal. They want practical progress.
So where does the field stand in mid-2026? Tools work better than ever in narrow domains. Protein folding. Antibody design. Certain virtual screens. Clinical translation lags. One or two AI-derived candidates show promise in Phase II. None have yet produced a landmark approval that traces its edge directly to artificial intelligence.
Pharma companies integrate these methods quietly. They fine-tune models on proprietary data. They combine AI with traditional assays. They avoid public overpromising. Startups chase headlines. Established players chase productivity.
The coming years will test the thesis. More AI-designed molecules will enter trials. Some will succeed. Others will fail for familiar reasons: wrong target, poor safety, unexpected biology. Success rates may tick upward modestly. Dramatic compression of timelines seems further off.
And the skepticism serves a purpose. It forces the community to confront limitations. To invest in better data infrastructure. To design experiments that truly challenge the models. To integrate human insight at every step.
Lowe ends on a note familiar to longtime readers. Short-term pessimist. Long-term optimist. The technology holds real power. Delivering on that power demands rigor, patience, and honest assessment. Exactly what his post provides.
Industry insiders watch closely. Billions in venture capital ride on the outcome. Patients wait for better medicines. The experiments continue. The data accumulate. The verdict, as always in drug discovery, will come from the clinic.