If traditional drug development is like trying countless keys in the dark, AI's first step is not to open the door itself, but to prioritize the keys most likely to fit. In 2026, several strong signals emerged for this narrative. In March, Insilico Medicine partnered with Eli Lilly, securing a $115 million upfront payment with a potential total of approximately $2.75 billion. By July, its AI-discovered idiopathic pulmonary fibrosis candidate, Rentosertib, had entered Phase III clinical trials. Earlier, Isomorphic Labs also announced collaborations with Eli Lilly and Novartis, with a combined potential value nearing $3 billion.

Another signal, more captivating for the tech industry, came in January 2026 when NVIDIA and Eli Lilly announced a joint AI innovation lab in the San Francisco Bay Area. They plan to invest up to $1 billion combined over five years in infrastructure and R&D, integrating BioNeMo, next-generation computing, robotics, and "physical AI" into drug discovery and manufacturing. In June, Anthropic launched Claude Science, connecting over 60 scientific databases, research software, and computing environments into an auditable scientific workbench. It also links to life science models like Evo 2, Boltz-2, and OpenFold3 via the NVIDIA BioNeMo Agent Toolkit. These two tech giants are pursuing different paths: NVIDIA provides both computing power and experimental infrastructure, while Anthropic aims to be the intelligent gateway for the entire research process. They both point to a key shift: competition in AI drug discovery is moving from "who has a better molecule model" to "who can organize models, data, experiments, and human judgment into a productive system."

What is AI Drug Discovery Exactly?

Often called AIDD (Artificial Intelligence-driven Drug Discovery), it embeds machine learning, generative models, multi-omics, physical computing, and automated experiments into the drug development process. It can participate in target discovery, protein structure prediction, molecular or antibody design, ADMET prediction, and synthesis route planning, and is now entering clinical trial design, patient stratification, and regulatory document preparation. Simply put, traditional R&D relies more on expert-driven, iterative trial and error. AI simultaneously considers constraints like activity, selectivity, toxicity, and synthesizability to generate and rank candidates more worthy of validation. Scientists provide the questions and judgment, while experiments provide the truth. Therefore, the core of AIDD is not a single "guess," but a continuously tightening cycle: Design, Make, Test, Analyze, and then redesign. Without wet-lab experiments, models may just repeat old experiences on old data. Without AI, experimental resources are easily wasted on low-probability directions. The speed of iteration between these two sides determines R&D efficiency.

What Benefits Can AIDD Bring?

First, it expands the search boundary. AI can explore chemical and protein spaces far beyond manual enumeration, simultaneously addressing multiple drug-likeness targets. Second, it enables failure to be front-loaded through reverse screening. Molecular dynamics, free energy calculations, ADMET predictions, and phenotypic models can eliminate some low-quality candidates before synthesis, animal testing, and clinical trials. Insilico Medicine published research showing that Rentosertib took about 18 months from target discovery to preclinical candidate nomination. In a Phase IIa study published in 2025, the highest dose group showed a positive trend in lung capacity over 12 weeks, but the sample size was limited and requires larger trials for validation. Third, it transforms R&D from a linear relay to a high-frequency cycle. Robotic labs continuously perform synthesis, purification, testing, and data feedback, while research agents propose the next round of experiments. The true efficiency gain lies in less waiting and less data loss between each iteration. However, biology does not become simpler just because computing power becomes cheaper. AI has a non-negotiable boundary: disease biology, human variability, and long-term safety cannot be eliminated by computation alone. The FDA's focus on AI in drug development has shifted from "can it be used" to whether the model is trustworthy for its specific use, emphasizing data governance, risk classification, performance evaluation, and lifecycle management.

From Prediction to Validation: The Crucial Leap

The development of AI drug discovery can be summarized in three stages. Stage 1.0 is "Prediction Tools." Capabilities include virtual screening, structure prediction, and property calculation, with AI acting as an efficiency plugin for researchers. Stage 2.0 is "Generation Platforms." Models start generating molecules and proteins from scratch, and business models shift from software subscriptions to project-based delivery, pipeline licensing, and joint R&D. Stage 3.0 is "Experimental Closed-Loop + Clinical Validation." Models connect with automated labs, continuously generating high-quality data and enabling automated trial-and-error iteration. Candidates undergo Phase II and III clinical testing. Platforms must also demonstrate repeat purchases, milestones, and consistent delivery capabilities. Currently, AIDD is in the early transition from 2.0 to 3.0. Technical feasibility and business collaborations have been validated, and Rentosertib entering Phase III is a significant milestone. However, the question of whether AIDD can systematically improve mid-to-late-stage clinical success rates remains unanswered. Overstating the maturity of the phase would overestimate short-term profits, while treating AI as only a concept would underestimate the paradigm shift in R&D.

Who Will Benefit First?

A common claim is that since AI generates more molecules, the demand for wet-lab experiments will proportionally explode. This reasoning is incomplete. AI can also filter out candidates virtually, potentially reducing the number of molecules actually synthesized per project. Industrial gains are more likely to come from three multipliers: more projects can be initiated, each project iterates faster, and higher demands for experimental quality and data standards. First, scientific services and experimental infrastructure benefit first. In the era of large-scale AIDD infrastructure, gene and protein synthesis, recombinant proteins and antibody reagents, high-throughput screening, cellular and organoid models, and automated lab equipment are the first stops for turning model outputs into verifiable data. The most advantaged are not those simply selling reagents, but service providers who can offer standardized, traceable, and data-recyclable feedback to the models. Second, preclinical CROs and integrated CXOs benefit from pipeline spillover. As candidates advance to efficacy, pharmacokinetics, toxicology, safety evaluation, and CMC, demand for animal models, DMPK, safety assessment, process development, and CDMO will be released sequentially. The pace of benefit depends on whether AI pipelines can truly cross PCC and IND milestones, not on the hype of model launches. Third, innovative drug and AI platforms share the highest upside potential but also bear the greatest volatility. Companies that can connect disease understanding, proprietary data, algorithms, automated experiments, and clinical development into a closed loop can improve R&D output per unit of investment. However, single pipeline failures, missed BD milestones, and sustained financing pressures will directly impact valuations. Therefore, assessing true benefit requires four indicators: visible AI-related revenue and repeat purchases; a proprietary data closed-loop from experiments; continuous delivery of PCC, IND, and clinical milestones; and the platform's ability to deliver across multiple stages, not just showcase model parameters.

From Industry Trends to Index Tools: Three Risk-Return Expressions

For investors more focused on the synergy of the A-share innovative drug and R&D service chain, the CSI Innovative Drug Industry Index (931152.CSI) can be considered. It selects up to 50 representative stocks from companies primarily involved in innovative drug R&D, including the R&D service chain, suitable for expressing an "innovative drug + CXO" combination logic. For those more focused on the flexibility of Hong Kong-listed innovative drug platforms and pipeline assets, the CSI Hong Kong Innovative Drug Index (931787.CSI) offers higher purity and flexibility, but as a Hong Kong market index, it also carries risks from exchange rates, offshore market trading, and secondary market premiums. For investors seeking lower single-theme concentration, the CSI All-Share Medical and Health Index (000991.SH) offers broader coverage, with AI drug discovery being just one growth line. While it has lower purity, it offers higher diversification. It is important to emphasize that these indices are not narrowly defined "AI drug discovery themes." They are better understood as different risk-return expressions of the innovative drug and pharmaceutical industry chain. Ultimately, long-term returns will be determined by clinical data, product approvals, and commercialization capabilities, not by whether "AI" is in the company name. AIDD will not compress a decade of drug development into a single click, nor will it eliminate clinical failures. What it truly changes is making hypotheses faster, exposing failures earlier, and returning experimental data to the next decision cycle more quickly. In the short term, focus on scientific services and experimental infrastructure. In the medium term, watch pipelines and milestones. In the long term, look at clinical success rates and approved drugs. The value of AI drug discovery lies not in peak computing power or model parameters, but in the real drugs that reach patients, one after another.