Beyond Legacy Tools: Defining Modern AI Drug Discovery for 2025 and Beyond

In this report:

• Intro: The New Framework
• “AI Drug Discovery” is About Holism
• "AI Drug Discovery" is About Building Software
• Access to Data is King
• Validation is Critical for AI Drug Discovery Platforms
• Grand Vision of AI Drug Discovery for 2025 and Beyond

Read our disclaimer

With the start of 2025, it seems there is still a lack of a robust definition of an emerging category of artificial intelligence-driven drug discovery companies (hereinafter, AIDD). 

The purpose of this report is to suggest a qualitative framework for classification of AIDD companies, combining the four key attributes that define the leading players in this area: 

• Focus on holism vs reductionism
• Creating robust AI platforms (software)
• Priority of data acquisition
• Technology validation (via demonstrable ability to discover novel targets, discovery and develop clinical-grade drug candidates rapidly, a track record of platform partnerships, scientific publications, patents, and so on)   

We will delve deeper into framework discussion below, but in a nutshell, it boils down to this: 

Indeed, abstracting from specific characteristics of a tech stack and platform design, there are three key value points of an AI platform on business outcome: 

1. Is a computational platform scalable and robust enough to impact the R&D workflow, people collaboration patterns, and daily decision making of a wide range of specialists of a given organization to make a tangible difference? 

2. Is it able to represent biology in silico down to sufficient depth, but also sufficient breadth to be able to grasp relevant and useful dependencies, patterns, network biology effects, to be able to impact scientific decision-making beyond the status quo patterns? 

3. Is the AI platform capable of addressing the above two questions in a repeatable, stable, standardized way across all levels of R&D workflows in the organization? Would a third-party collaborator be able to get sustainable value from using the AI software if they had access? 

In my opinion, AIDD is about being able to answer “yes” to all three questions.

In practical terms, the above framework filters off the overwhelming majority of companies that were labeled, or claimed to be, AIDD companies. As further analysis will show, in fact, there is no mainstream adoption of AI in drug discovery so far, unlike what is routinely presented in media and analytical reports. 

The application of the abovementioned logical framework to a list of companies, typically mentioned as AIDD companies in media or analytical reports as well as by companies themselves, leads to their fragmentation into several groups. 

1. Companies in the darker green area are the closest to what I would classify as actually working AIDD business models. These companies have mature AI software platforms, some of which are used commercially by third parties. They have exclusive data foundation (e.g. internal biobanks), and/or large scale in house data repositories or data generation capabilities (e.g. high throughput experimental biology labs). Companies in this group are targeting holistic biology, trying to map the entirety of it. Most of them are well-validated. 

2. Lighter green area includes strong emerging players in the AIDD space, with highly sophisticated broad scope platforms, but certain limitations in one of the factors (either less developed software, or narrower modeling focus, or smaller data foundation etc). Such companies, nonetheless, are representing the holistic approach in drug discovery.

3. Companies in white area represent either niche AI platforms (the first column), mostly focused on target-based drug discovery, chemical simulations, or specific areas of biological modeling (e.g. protein structure prediction and design). Companies in the right-hand part of the diagram are generally focused on holistic modeling using advanced AI algorithms, but might be earlier in their journey towards AI software products.

It goes without saying, all companies that made it to this table are using some of the most sophisticated algorithmic and architectural solutions there are in the cheminformatics and bioinformatics space. They are all strong players, but some of them are more representative of the modern AIDD strategy than others.

Disclaimer: While I did my best to classify companies to the best of my knowledge and sector understanding, there is always a possibility of misinterpretation of publicly available data, or confusion. If you believe some companies were inaccurately classified anywhere in the report, or missing from the report, kindly suggest supporting information via this form for possible updates. 

Now, let’s dive in and explore the suggested classification framework and AIDD perspective in more detail. 

“AI Drug Discovery” is About Holism 

Starting to explore the newly suggested framework, one key difference lies in what we model and what we try to represent in silico, in contrast to what could be modelled and represented by legacy tools of the past.

A good first question is, where lies the silver lining between a strong computational tool like a 36 years old AutoDock Software, used by many pharma companies, which is typically not considered in the context of modern “AI drug discovery” and a computational platform like Pharma.AI by Boston-based Insilico Medicine, which does fall under the broad acceptance as end-to-end “AI drug discovery” solution? 

In simple terms, “traditional” or “legacy” cheminformatics and bioinformatics rely on human-driven approaches: cheminformatics uses predefined chemical descriptors (like molecular weight or logP), statistical methods and some machine learning approaches for tasks like QSAR modeling and docking, while bioinformatics applies statistical methods, including dimensionality reduction techniques, to analyze complex biological datasets (e.g., genomics, proteomics) and uncover potential drug targets. These methods are hypothesis-driven, modular, and work with smaller, well-structured datasets. 

Conceptually, legacy computational systems and simpler machine learning methods are useful in the paradigm of “biological reductionism.” And they do a great job there, even today.

Classical reductionist approach example is structure-based drug discovery, where it is believed modulating a specific protein is an answer to a drug discovery problem (it sometimes is). The computational part, therefore, is mostly focused on narrow-scope tasks like fitting a ligand into a protein pocket (docking), or, computationally identifying a new type of chemistry for a given target (ligand-based virtual screening).

In stark contrast, cutting edge AI-driven drug discovery companies attempt to shift to a systems biology level, a hypothesis-agnostic approach, using deep learning-based systems to integrate largely multimodal data (phenotype, omics, patient data, chemical structures, texts, images, etc) to construct complex and comprehensive biology representations (e.g. “knowledge graphs”).

For example, the scientific underpinnings of Insilico Medicine's platform Pharma.AI are rooted in a novel combination of policy-gradient-based reinforcement learning (RL) and generative models, enabling multi-objective optimization to balance parameters such as potency, toxicity, and novelty. 

According to the company, a target identification PandaOmics module leverages 1.9 trillion data points from over 10 million biological samples (including RNA sequencing and proteomics) and 40 million documents (such as patents and clinical trials), using NLP and machine learning to uncover and prioritize novel therapeutic targets. 

The Chemistry42 module applies deep learning, including generative adversarial networks (GANs) and reinforcement learning, to design novel drug-like molecules optimized for binding affinity, metabolic stability, and bioavailability. 

In the context of clinical development, inClinico predicts trial outcomes using historical and ongoing trial data, offering insights into patient selection and endpoint optimization. 

On an algorithm side of things, Pharma.AI incorporates advanced reward shaping, allowing it to fine-tune generated molecules to specific target profiles or polypharmacological goals. Additionally, Insilico emphasizes the use of knowledge graph embeddings, which encode biological relationships — such as gene–disease, gene–compound, and compound–target interactions — into vector spaces. 

These embeddings are augmented by attention-based neural architectures, inspired by Transformer models, to focus on biologically relevant subgraphs, refining hypotheses for target identification and biomarker discovery. 

The platform employs a continuous active learning and iterative feedback process, retraining models on new experimental data, including biochemical assays, phenotypic screens, and in vivo validations, to accelerate the design–make–test–analyze (DMTA) cycle by rapidly eliminating suboptimal candidates and enhancing lead generation. 

Furthermore, the platform’s multi-modal data fusion integrates textual information from published literature, patents, and clinical trial data with omics-level insights and chemical libraries. To this end, Natural Language Processing (NLP) models are used to extract relevant biological context and side-effect annotations from these textual sources, which are then enriched with phenotypic screening data, enabling a holistic view of the drug discovery process.

You can familiarize yourself with some of the aspects of the Pharma.AI platform by reading a recent paper “A small-molecule TNIK inhibitor targets fibrosis in preclinical and clinical models” (image below is from the paper).

Another relevant example of what can be classified as an AI drug discovery approach is Recursion Pharmaceuticals’ OS Platform.

While different from Insilico Medicine in model architectures and workflows, Recursion is, however, focused on the same key objective: to create a comprehensive representation of biology to be able to mine crucial insights for drug discovery:

Key models of Recursion OS include Phenom-2, a 1.9 billion-parameter ViT-G/8 MAE trained on 8 billion microscopy images, achieving a 60% improvement in genetic perturbation separability. MolPhenix, winner of NeurIPS 2024 Best Paper, predicts molecule-phenotype effects with a 10X improvement over baselines. MolGPS, a 3-billion-parameter model, excels in molecular property prediction and integrates proprietary phenomics data, outperforming benchmarks in 12 of 22 ADMET tasks. MolE, trained on 842 million molecular graphs, leads in 10 of 22 ADMET tasks. Additionally, OpenPhenom-S/16, publicly available on Google Cloud and Hugging Face, improves biological relationship recall using over 3 million microscopy images. 

Conceptually, “AI drug discovery”, in contrast to “legacy” computational systems refers to a modern computational tech stack, usually a multimodal ensemble, that is capable of modeling biology holistically, including molecular, phenotypic, and clinical data of all types and sizes (chemical, omics, text, images (e.g. cell staining), EHR, etc) -- all at once, or substantial part of variety. 

Another crucial aspect differing modern AIDD from earlier computational tools is generative capabilities. 

While companies like Insilico Medicine pioneered the use of Generative Adversarial Networks (GANs) for generative chemistry back in 2016, by leveraging their ability to model complex molecular distributions and propose novel chemical structures, it is the introduction of transformers and attention mechanisms in 2017, particularly with the advent of models like BERT and GPT, that in my opinion rendered a paradigm shift of generative modeling across domains.

I consider 2017 as a pillar year for generative AI, including chemistry and biology, after the landmark paper “Attention is all you need”.

These architectures, pioneered by Google, and later developed by OpenAI, Antropic, Mistral AI, and others, demonstrated unparalleled scalability and capacity for capturing long-range dependencies in sequential data.

By pretraining on vast corpora of text (hundreds of billions and even trillions of parameters) and employing self-attention to dynamically weight input relationships, transformers enabled large-scale generative models such as GPT-3 and GPT-4 to generate highly coherent and contextually accurate outputs.

Yes, “hallucinations” are still a major issue. But the shift is paramount. The pioneering commercial products in this regard are ChatGPT for primarily text-to-text, Midjourney for text-to-image generation, and many others for text-to-video, text-to-music, etc.

The emergence of practically feasible transformers and large language models catalyzed a sort of race in computational chemistry and biology towards so-called foundation models. The article 19 Companies Pioneering AI Foundation Models in Pharma and Biotech summarizes some of the initiatives in this domain. 

For instance, Insilico Medicine's latest tools include nach0, a multimodal encoder-decoder large language model pre-trained on scientific literature, patents, and molecular data using NVIDIA's NeMo framework, excelling in tasks like molecular generation, synthesis, and biomedical question answering. The model employs instruction tuning and achieves superior performance across single-domain and cross-domain benchmarks. Additionally, Precious3GPT is a multimodal transformer model enabling applications like tissue-specific age prediction, compound effect analysis, and experiment simulation.

To summarize, here is a simple generalizable framework to draw a silver lining between legacy CADD and modern AIDD:

Next, as we have reviewed what “AI drug discovery” attempts to model (holistic biology vs mainstream “reductionism”), and what kind of models are generally capable of doing it, let’s discuss another crucial aspect -- AI platform maturity as a software product.

"AI Drug Discovery" is About Building Software

A characteristic feature of leading AIDD platforms versus “superficial” AI-companies, is the demonstrable focus on building actual software, a simple but somehow overlooked observation by many analysts, journalists, and commentators in this area. 

We should expect that AIDD company has to be able to demonstrate the presence of a robust, self-contained software platform that supports critical functionalities—ranging from user-friendly interfaces (GUI) for data input and parameter tuning, to configurable machine learning modules (including algorithm selection, hyperparameter adjustment, and visualization of model performance). 

Such a platform should integrate standardized data ingestion pipelines (e.g., for omics data, small-molecule libraries, or clinical metadata) with back-end components enabling dynamic model training, validation, and iterative optimization (e.g., active learning, reinforcement learning loops). 

A well-documented application programming interface (API) is also essential for interoperation with external tools, ensuring end users can automate workflows and seamlessly exchange data between software components. Additionally, a proper end-to-end solution should incorporate security, data integrity measures (version control, audit trails, encryption), and deployment options (on-premises or cloud-based) to fit diverse organizational needs. 

For platforms that are meeting the definition of “AI drug discovery” suggested by a new framework, you can actually see and even access a demo of their software, and sometimes licence it and use it for your internal projects. 

Examples include Insilico Medicine’s platform here, Recursion Pharmaceuticals’s here, CytoReason’s software here, for NOETIK’s here, for BPGbio here, for OWKIN here, and so on.

I wasn’t able to identify information about software characteristics or live demos from the overwhelming majority of companies claiming to be AI-driven businesses.

Now, why would I focus on this so much? Because at the moment, there is no AI solution on the market, where you just press a button, and it outputs a clinical-grade therapeutic asset for a specific sub-group of patients. All cutting edge AI systems are, in essence, still co-pilots to assist human scientists behind the scenes. 

So, in the present role of corporate co-pilots, the AI system has to be meaningfully integrated into the organizational workflow. Only a properly designed software can provide this level of interoperability to actually impact the rate and quality of innovation at the organizational level. So, it is not a matter of formality, it is a matter of operational impact on R&D success rate. 

Only a robust, well tested AI software product, integrated across the entirety of the company R&D processes and enabling collaborative workflows, should be classified as AIDD.

A big difference between modern AIDD software solutions and good old cheminformatics/bioinformatics software packages are:

1. Scale and complexity of data AIDD platforms can handle and be trained on;
2. Cutting edge predictive and generative model ensembles under the hood;
3. Interoperability between various stages of drug discovery, combining hypothesis construction, biology modeling, molecular simulations, lead optimization, preclinical insights, and clinical predictions.  

Access to Data is King

In 2025, I believe the key theme in the AIDD space could revolve around companies and deals related to data generation and gathering, as well as any new innovative tools and techniques for experimental data mining (e.g. NGS, next gen proteomics, mass spectrometry, 3D-NMR, cryo-EM, organ-on-chip systems, robotic labs, etc). Companies like Tempus will be dominating headlines, in my opinion.

From early on, AI drug discovery companies like Insilico Medicine, BPGbio, Recursion, and more recently, NOETIK, have been investing in data acquisition as a cornerstone of their assets valuation. 

For instance, Insilico Medicine's so-called “6th-generation” intelligent robotics drug discovery laboratory, launched in Suzhou BioBAY in January 2023, integrates AI-powered decision-making with fully automated robotic modules for target discovery, compound screening, precision medicine development, and translational research. 

According to company claims, by combining its Pharma.AI platform with six functional modules—spanning automated cell culture, high-throughput screening, next-generation sequencing (NGS), and high-content imaging—the lab forms a closed-loop system that validates novel targets, optimizes lead compounds, and generates high-quality biological data to train and refine AI models. 

In another example, Recursion's data foundation includes 60+ petabytes of proprietary multiomics data, such as phenomics, transcriptomics, proteomics, ADME, InVivomics, genomics, and patient data. Internally, they’ve generated ~36 petabytes via 2.2 million weekly high-throughput experiments, using CRISPR-Cas9 editing and Brightfield imaging to create one of the largest pharma-related datasets. This data is embedded using AI models for advanced biological analysis.

According to an exclusive interview with a company representative, Recursion processes 2.2+ petabytes of transcriptomics data and integrates 20 petabytes of patient data from Helix and Tempus, covering whole genome and exome sequencing from hundreds of thousands of cases. 

In cell manufacturing, Recursion produces 1 trillion hiPSC-derived neuronal cells, creating the "Neuromap" for neuroscience and oncology programs with Roche and Genentech, spanning 40 therapeutic programs.

Next example, CytoReason constructs its data foundation by integrating extensive public and proprietary datasets, encompassing bulk and single-cell transcriptomics, proteomics, and clinical data, into a unified AI-driven Disease Model Platform.

This platform employs advanced machine learning algorithms to map and compare treatments, patient groups, and disease mechanisms at cellular and molecular levels, enabling comprehensive analyses across various diseases and tissues.

Yet another example is Berg Health (now BPGbio), which established an extensive biobank comprising over 100,000 clinically annotated human specimens, including biofluids and tissue samples, to fuel their AI-driven drug discovery platform.

The company conducted comprehensive multi-omics profiling of the specimens—encompassing genomics, proteomics, metabolomics, and lipidomics—to capture a holistic view of human biology. 

The resulting high-dimensional datasets were analyzed using their proprietary NAi Interrogative Biology® platform, which integrates Bayesian artificial intelligence learning algorithms to identify disease-specific biomarkers and therapeutic targets. This, arguably, led to the company’s quite successful clinical trial launches over the years. 

Finally, a more recent but promising approach is how NOETIK is building their data foundation by sourcing curated human tumor specimens for its in-house biobank, applying stringent quality controls on parameters like ischemia time, necrosis percentage, and sample age, and ensuring each sample is pathologist-reviewed before inclusion. 

The company generates multimodal datasets through advanced techniques such as spatial transcriptomics for single-cell RNA expression, whole exome sequencing for genomic alterations, and custom protein panels to map tumor-immune microenvironment interactions, all anchored by spatially randomized tissue microarrays to mitigate slide-level artifacts.

This vast data pipeline, coupled with their patent-pending processes, enables the creation of high-quality, self-supervised training datasets that power their AI engine OCTO, designed to model tumor biology and predict patient-specific therapeutic responses.

Validation is Critical for AI Drug Discovery Platforms

Finally, a central measure of credibility in AIDD is platform validation, which typically involves demonstrating tangible outcomes and reproducibility across diverse use cases. 

Possible ways to validate a platform include a combination of the following: 

(1) by advancing internal pipelines of novel therapeutics, where the AI engine is used support R&D team in discovering, designing, and optimizing lead molecules that progress through preclinical and, in some cases, clinical development. 

(2) through partnerships with established pharmaceutical or biotech organizations, enabling third-party testing of the AI platform’s predictive power and generative capabilities on proprietary datasets. A track-record of public milestone announcements is critical. 

(3) via public software demos or proof-of-concept studies published in peer reviewed journals, and patents. 

(4) via regular publishing of AIDD case studies in peer reviewed journals  

Below is a table showing historical pipeline growth dynamics for several selected companies, including BenevolentAI, Healx, Insilico Medicine, Iambic Therapeutics, Schrodinger, Relay Therapeutics, Recursion, Valo Health, Verge Genomics, and Exscientia which was acquired by Recursion in 2024:

Beyond just the number of drug candidates, it is interesting to look at the target novelty landscape of the companies in the AIDD category:

Commenting on the above data from the table and the infographics, Insilico Medicine has shown remarkable pipeline growth over the past five years. Specifically, the company has launched 31 therapeutic programs targeting diverse indications, with 21 preclinical candidates nominated from 2021 and nine more in 2022, and a total of 10 pipelines receiving IND approval. At present, the leading Insilico program for idiopathic pulmonary fibrosis (IPF) was discovered from concept to Phase I trials in just under 30 months and is now in Phase 2 clinical trials in both the United States and China, while five other programs are in Phase 1.

Additionally, several clinical assets have been out-licensed or co-developed with third parties, including recent milestone payment. These accomplishments appear to highlight a significant productivity boost, yet the question remains whether they definitively prove that AI technologies — rather than more conventional R&D structures and partnerships — are the key drivers behind this rapid expansion.

A similar phenomenon is observed with Schrodinger’s platform, which, although not explicitly marketed as “AI,” has enabled significant pipeline development. While software capabilities can help streamline decision-making — such as accelerating target identification and optimizing lead compounds — it is not a simple matter of “printing” successful molecules. Determining the true impact of AI requires assessing the extent to which such platforms reduce discovery cycles or increase success rates in a statistically meaningful way. One potentially stronger validation approach is to examine how many third-party organizations license and effectively use these tools for drug development, thereby providing external feedback and real-world performance benchmarks.

Pragmatic Recommendations for AIDD Investors and Stakeholders:

• Look Beyond Candidate Counts: Merely tallying the number of pipeline assets does not capture the incremental value AI platforms may provide. Faster program initiation or more accurate attrition rates, for instance, could be more telling indicators.

  A good example is CytoReason. The company is not involved in building internal pipelines of drug candidates, as it has a service business model. But the track record of the usage of its software by dozens of leading pharma players (big pharma) is reflective of the robustness and reproducibility of their AI software product.

• Evaluate Decision-Making Efficiency: Pinpoint where AI significantly shortens R&D workflows — e.g., by expediting hit-to-lead stages or improving target validation.

• Scrutinize External Adoption: Seek third-party evidence of productivity gains, such as collaboration announcements, successful milestones, or continued software licensing agreements. Tools that are openly licensed or sold commercially allow for real competitive benchmarking.

• Consider Contextual Factors: Keep in mind that corporate strategy, funding, and existing R&D infrastructure often play major roles in pipeline output. It is not always possible to isolate AI’s contribution without analyzing these concurrent influences.

Grand Vision of AI Drug Discovery for 2025 and Beyond

Having reviewed many claimed AI drug discovery companies over a decade of progress, hype and facing reality (like failed clinical trials of arguably AI-designed drug candidates), it is time to define the goal and method of AI drug discovery, and accept that the overwhelming majority of companies are not there yet.

The entire idea of the AIDD movement is, in fact, not about improving existing drug discovery processes, like structure-based drug discovery or virtual screening via using better models, advanced machine learning etc.

Obviously, it helps and most of the companies in this business are doing it. Using better models for screening or docking could have marginal improvements of research processes, but does not change the fundamental problem of drug discovery: poor translation of hypotheses to clinical results and high degree of clinical failures due to unexpected toxicity or poor efficacy in a selected patient sub-population.

The novelty and ambition of the AIDD approach is about redesigning the existing mainstream drug discovery paradigm into something different.

I suggest calling it “Rational Phenotypic-based Drug Discovery (RPDD).”

Starting from modeling the entirety of real-world data about patients (coming from specimens, analytical samples, and other biomedical data) and building the path down to a relevant underlying hypothesis on a molecular level. And then, walking that path in reverse -- from the newly discovered hypothesis, via drug design and development, back to the patient. Hopefully, with the improved probability of success.

In theory, fully realized RPDD vision would help eliminate the use of animal models in drug discovery. We are not there yet, but the industry is evolving, company by company, and now is the right time to invest and build in this area.

In my opinion, RPDD is the goal of the pharmaceutical industry for the coming years. AI is just a tool that could enable us to achieve that sooner than later. 

Some companies, like more established Insilico Medicine, and Recursion Pharmaceuticals, BPGbio, OWKIN, and more recent newcomers, like Iambic Therapeutics, Verge Genomics, NOETIK and others, are zooming in on the new way of doing things. 

Time will tell if AIDD proves to be the better way, I am cautiously optimistic...