AI Virtual Screening Hits 750x the Rate of Your HTS Lab. At 90% Lower Cost.

The Week Pharma Declared Its Intentions

NVIDIA GTC 2026 ran from March 16 to 19 in San Jose. By the time the conference ended, the pharmaceutical industry had announced more AI infrastructure in four days than it deployed in the prior four years.

Eli Lilly unveiled LillyPod: 1,016 NVIDIA Blackwell Ultra GPUs, 290 terabytes of high bandwidth GPU memory available to genomics teams, 700 terabytes of genomic data accessible through the system, all assembled in four months. Yue Wang Webster, VP of R&D Informatics at Lilly, stated the supercomputer "breaks the physical limit of the wet lab." That phrasing was deliberate. LillyPod runs protein diffusion model training, small molecule graph neural network models, genomics foundation models, and agentic workflows that direct research lab agents. It is not a screening supplement. It is a screening replacement operating at computational scale.

Then came Roche. The company now operates the pharmaceutical industry's largest announced hybrid cloud AI factory, totaling more than 3,500 NVIDIA Blackwell GPUs deployed across hybrid cloud and on premises infrastructure in the US and Europe, with 2,176 new Blackwell GPUs added in the latest expansion. Nearly 90% of Genentech's eligible small molecule programs now integrate artificial intelligence drug discovery tools. That number is not aspirational. It is operational. One oncology degrader molecule was designed 25% faster using AI assisted methods. A backup drug candidate was delivered in seven months instead of the previous two plus years, a reduction of over 70%.

IQVIA launched IQVIA.ai, a unified agentic AI platform with 150 intelligent agents already embedded in 19 of the top 20 pharma companies. NVIDIA introduced nvQSP, a GPU accelerated simulation engine for Quantitative Systems Pharmacology delivering 77x faster performance than traditional CPU simulations. Researchers can now analyze hundreds of dose levels and patient subpopulations in the time it previously took to simulate a handful.

Behind all of this sits the $1 billion NVIDIA Lilly co innovation lab announced in January at J.P. Morgan, co locating Lilly domain experts with NVIDIA's AI engineers in South San Francisco. And LillyTuneLab, launched September 2025, giving biotech companies access to ai tools for research trained on over $1 billion worth of proprietary Lilly data through federated learning, with partners including Insilico Medicine, Schrodinger, and Chai Discovery.

[LillyPod] breaks the physical limit of the wet lab.

Yue Wang Webster, VP of R&D Informatics, Eli Lilly

GTC sent one message: the companies building supercomputers and deploying ai tools for research at industrial scale are writing the next chapter of the drug discovery process. The ones buying liquid handling workstations are reading the last one.

High Throughput Screening by the Numbers

High throughput screening has been the backbone of pharmaceutical drug screening methods for three decades. Automated robotic systems test large chemical libraries against biological targets, measuring binding or activity across thousands of wells per plate. At peak throughput on 1536 well plates, a well equipped HTS facility screens 200,000 compounds per day. That sounds fast until you look at the hit rates.

The market is not small. At $27 to $36 billion in 2025 and projected to reach $71 billion by 2035, high throughput screening remains a massive industry. But its composition is shifting. AI integrated HTS is growing faster than traditional approaches, and standalone physical screening without computational pre filtering is becoming obsolete.

The cost structure tells the story. Equipment alone runs $80,000 to $150,000 for flow cytometers and $100,000 to $500,000 for liquid handling workstations. Phenotypic screening costs over $1.50 per well. A single campaign requires large quantities of purified target protein. And after all that, the median outcome is a hit rate below 1%, meaning 99% or more of the compounds tested produce no useful signal.

For thirty years, this was the only option. It no longer is.

AI Virtual Screening Goes Head to Head

The most comprehensive comparison of AI virtual screening against traditional high throughput screening comes from a landmark study published in Nature Scientific Reports in April 2024, covering 318 biological targets screened using Atomwise's AtomNet convolutional neural network.

The study concluded that "machine learning approaches have reached a computational accuracy that can replace HTS as the first step of small molecule drug discovery." That was April 2024. The models released since then widened the gap further.

Boltz 2, released in June 2025 by MIT and Recursion Pharmaceuticals, is the first model to combine structure prediction and binding affinity prediction in a single framework. On the standard FEP+ affinity benchmark, Boltz 2 matches the accuracy of OpenFE, a widely adopted free energy perturbation pipeline, while running over 1,000 times faster. A drug screening campaign that would take months with physics based methods completes in hours. It doubles the average precision of standard ML and docking baselines. It runs on a single consumer GPU under an MIT license.

Chai 2, from Chai Discovery (backed by $225 million including participation from OpenAI), designs fully de novo antibody complementarity determining regions based solely on a given epitope and target structure, with no templates or scaffolds. The antibody hit rate: 16% to 20%, compared to less than 0.16% for previous computational methods. A 100x improvement. Across 52 diverse antigens, 50% of targets yielded at least one hit from just 20 designed candidates, validated in a standard 24 well plate in under two weeks.

The pharmaceutical analysis is straightforward: AI virtual screening reduces campaign costs by over 90%, compresses timelines from months to days, tests fewer than 500 physical compounds instead of millions, and delivers hit rates five to seven hundred times higher. Computational chemistry software now outperforms the pipette on every measure except physical confirmation.

Who Has Already Moved

The companies that have restructured their screening operations around artificial intelligence drug discovery are not startups running pilot programs. They are the largest pharmaceutical organizations on earth.

Recursion and Exscientia merged in November 2024 for $688 million, creating the most comprehensive AI drug discovery platform in the industry. The combined entity integrates Exscientia's automated precision chemistry with Recursion's 60 plus petabytes of proprietary biological data, spanning 10 clinical and preclinical programs, 10 advanced discovery programs, and approximately $450 million in received upfront and milestone payments, with $20 billion in additional potential milestones.

Roche and Genentech have reached 90% AI integration across eligible small molecule programs. This is not a pilot. Genentech's Lab in the Loop strategy connects experiments, data, and AI in continuous cycles of hypothesis generation and experimental validation, replacing the sequential screen, analyze, iterate workflow that defined pharmaceutical R&D for decades.

Novartis built data42, the largest corporate database in biopharma. Using generative ai drug discovery tools, Novartis computationally designed 15 million potential compounds, used predictive models to filter for brain penetration and other properties, and worked with only approximately 60 molecules in the lab to arrive at a potent molecular scaffold. From 15 million to 60. That ratio tells you everything about where drug screening methods are heading.

Insilico Medicine operates the Chemistry42 generative platform, used by over 20 pharmaceutical companies. Its lead compound, rentosertib, published Phase IIa results in Nature Medicine in June 2025: +98.4 mL improvement in lung function versus a 20.3 mL decline with placebo, the first AI designed drug for an AI discovered target to demonstrate clinical proof of concept.

Schrodinger operates the most commercially successful computational chemistry software platform in pharma, used by all top 20 pharmaceutical companies with a 100% retention rate for high value customers. Drug discovery revenue surged 295% in Q3 2025. Pfizer committed $11 billion in R&D spending for 2026, pairing AI engineers with scientists across every R&D function. AstraZeneca committed over $200 million across AI partnerships with BenevolentAI, VantAI, and Immunai.

Each small molecule discovery is like a work of art. If we can make that an engineering problem, versus this sort of discovery, this artisanal drug making problem, think of the impact on human life.

David Ricks, CEO, Eli Lilly

The workforce is shifting in parallel. Novo Nordisk announced plans to cut approximately 9,000 jobs to save $1.26 billion per year. Seventy percent of life science organizations now report some AI use. The HTS lab is not disappearing, but its role is migrating from primary discovery to validation of computational predictions, a fundamentally different job requiring computational biology and ML engineering alongside wet lab expertise.

The Infrastructure Buildout

What separates this moment from every prior wave of AI enthusiasm in pharma is the capital expenditure. These are not software licenses. They are nine figure infrastructure commitments.

LillyPod is a production system, not an experiment. Lilly spent a billion dollars in proprietary research data training models for TuneLab, then gave biotech companies access through federated learning, no IP requirements attached. This is a platform strategy designed to make Lilly the center of gravity for ai for scientists working in small molecule and biologics discovery.

Roche's AI factory pairs its 3,500 GPU deployment with NVIDIA Omniverse digital twins of production facilities, already accelerating development of a new GLP 1 manufacturing facility in North Carolina. AI is extending into regulatory documentation, quality assurance, and production scheduling. The gap between computational drug discovery and computational manufacturing is closing.

New models announced at GTC reinforce the trend. Proteina Complexa, a generative model for protein binder design, is already in use at Novo Nordisk, Viva Biotech, and Manifold Bio. nvMolKit provides GPU accelerated cheminformatics. ReaSyn v2 addresses synthesizability, one of the persistent limitations of generative ai chemistry. And the AlphaFold Protein Structure Database expanded by 1.7 million high confidence predicted protein complexes, with 100x faster inference through NVIDIA TensorRT integration.

What AI Screening Still Gets Wrong

The data favoring AI virtual screening over traditional high throughput screening is strong. It is also incomplete. Anyone building a drug development strategy around the assumption that computational screening has solved the discovery problem is making a mistake that the data does not support.

False positives remain the norm. While AI hit rates of 5% to 8% are far higher than the sub 1% rates of traditional HTS, the majority of computationally predicted hits still fail experimental validation. Virtual screening depends on molecular docking algorithms and scoring functions that can mispredict true binding affinities. Every predicted interaction requires physical confirmation.

Training data bias is structural. Drug development databases contain hundreds to thousands of times more inactive compounds than active ones. Models trained on biased datasets may not generalize to underrepresented chemical space. Deep learning models require 10,000 compounds or more to extract meaningful features, and high quality labeled data remains scarce for rare ADMET endpoints.

Phenotypic screening cannot be replaced computationally. Phenotypic screens capture multi pathway complexity that target based AI models cannot. Only 9.4% of small molecule drugs have been discovered through target based approaches alone. Target conditioned AI requires experimental data for each target and cannot account for multi target interactions.

ADMET prediction is improving but unreliable. Approximately 40% of preclinical candidates fail due to insufficient ADMET profiles. Roughly 30% of marketed drugs are withdrawn due to unforeseen toxic reactions. Prediction accuracy sits at R squared values of 0.60 to 0.90, too wide a range to replace in vivo testing. The FDA included AI based toxicity models under the New Approach Methodologies framework in April 2025, but with strict validation requirements.

The validation bottleneck is organizational, not computational. Many AI drug discovery initiatives stalled at the pilot stage in 2025, exposing gaps in data readiness, integration, and governance rather than model capability.

Machine learning approaches have reached a computational accuracy that can replace HTS as the first step of small molecule drug discovery.

Nature Scientific Reports, April 2024; with the caveat that the last mile from computational hit to clinical candidate remains stubbornly physical

The emerging model is clear: screen computationally, validate physically. AI narrows the funnel so the wet lab can focus on compounds worth testing. The 318 target study tested 85 to 440 compounds per project instead of hundreds of thousands. Organizations that treat AI screening as a complete replacement for pharmaceutical analysis rather than a vastly more efficient first pass, powered by ai tools for research and validated by physical experiment, will learn that lesson expensively.

What This Means for Formulation Science

Every announcement at GTC 2026 focused on the same phase of the drug development pipeline: finding molecules. LillyPod finds molecules. Roche's AI factory finds molecules. Boltz 2 predicts how molecules bind. Chai 2 designs antibodies. Chemistry42 generates candidates. None of them formulates a drug product.

This gap will dominate the next five years of pharmaceutical development.

Over 200 AI discovered drugs are now in clinical development: 94 in Phase I, 56 in Phase II, 15 in Phase III. Phase I success rates for AI discovered drugs run at 80% to 90%, roughly double the traditional baseline. Relay Therapeutics' zovegalisib has FDA Breakthrough Therapy Designation with 11.1 month median progression free survival.

The clinical failure rate sits at approximately 90% and has not moved. Faster discovery does not fix formulation. If AI compresses early discovery from 4.5 years to 12 to 18 months but formulation development, stability testing, and manufacturing scale up still consume the same years they always have, the relative burden of formulation increases. More candidates arriving faster at the formulation stage means more pressure on a process that remains largely manual and empirical.

The solubility numbers have not changed. Between 40% and 70% of all new hits have solubility too low for complete gastrointestinal absorption. Approximately 40% of new chemical entities are insoluble in water. The majority fall into BCS Class II. AI designed molecules frequently have high melting points and poor solvent solubility, compounding the formulation challenge.

Computational approaches to formulation screening are emerging but fragmented. AI models predict polymer solubility directly from SMILES representations. NVIDIA's nvQSP engine, with its 77x speedup, enables formulation scientists to simulate how different excipient selections, release profiles, and particle size distributions affect in vivo performance. Digital twins bridge molecular level prediction, process level simulation, and patient level modeling. But no dominant AI platform ties these capabilities together for formulation teams the way Schrodinger or Recursion's platform serves discovery teams. The discovery layer is well funded and increasingly consolidated. The formulation layer is fragmented.

DeepC Builds the Bridge

The AI drug discovery market is projected to grow from $2.9 billion to as much as $160 billion by 2035. Over 530 companies focus on AI drug discovery. Over $11 billion in venture capital flowed into AI ML drug discovery in 2025 alone. The formulation development side of the pipeline, where drugs actually fail, receives a fraction of that attention and investment.

DeepC occupies the space that LillyPod, Roche's AI factory, Boltz 2, Chai 2, and Chemistry42 do not touch: formulation intelligence. It is an AI co scientist for formulation scientists, built for the work that happens after a molecule is found.

SMILES2SPEC predicts LC MS/MS spectra from molecular structures, sitting in the emerging computational chemistry software stack between molecular design and formulation optimization. As pharma moves to AI first screening, spectral prediction becomes a critical validation layer. Formulation scientists can anticipate analytical challenges, identify potential degradation products computationally, and predict spectral signatures of excipient API combinations before synthesis. This is ai chemistry applied at the point where discovery hands off to formulation.

MiCQ provides ML powered Critical Quality Attribute imputation using KNN, MICE, MissForest, and CatBoost to address incomplete CQA datasets. With ADMET prediction models requiring 10,000 compounds or more and struggling with sparse data, intelligent imputation of critical quality attributes is directly valuable. It turns a persistent pharmaceutical analysis blocker into a solvable problem.

The Formulation Agent handles excipient selection, compatibility assessment, and formulation design using pharmaceutical machine learning models trained on real world formulation data. It reduces the traditional 3 to 6 month empirical screening process to a prioritized, evidence based shortlist. Computational excipient screening at this specificity does not exist in the discovery AI stack.

Elute extracts structured data from over 50 document formats: CMC documents, batch records, stability reports, analytical certificates, regulatory filings. Document intelligence at scale for organizations generating thousands of stability data points across dozens of regulatory databases.

Every recommendation DeepC generates carries data lineage from source to output, grounded in the FDA Inactive Ingredient Database, FAERS, DailyMed (154,834 drug labeling records), ClinicalTrials.gov, and PubMed. That traceability is the baseline requirement established by the FDA/EMA joint AI principles, the FDA's January 2025 credibility assessment framework, and EMA's Annex 22.

• LillyPod tests billions of molecular ideas in parallel. It cannot predict whether any of them will dissolve.
• Boltz 2 achieves binding affinity predictions 1,000x faster than physics based methods. It does not predict excipient compatibility.
• Novartis screened 15 million compounds computationally to arrive at 60 for the lab. Those 60 still need formulation, stability testing, and manufacturing scale up.
• Over 200 AI designed programs are in clinical development. The overall clinical failure rate remains approximately 90%.
• Eroom's Law, the observation that drug development costs double roughly every nine years, will not reverse by accelerating one phase while leaving the rest manual.

The convergence of AI tools from molecular design through property prediction to formulation optimization represents the end to end computational workflow the industry needs. Discovery is becoming engineering. Formulation has not made that transition. DeepC is building the platform that closes that gap, extending ai for scientists and generative ai drug discovery from the molecular design phase into the formulation lifecycle where drugs actually succeed or fail.