Protein FAQ

There are several areas where AI could add value, but one of the most exciting is helping us translate complex experimental results into actionable insights for drug discovery teams. We see computational and experimental approaches as complementary rather than competing technologies. In vitro assays such as iSLAseizure generate highly relevant biological data, while AI and machine learning can help uncover patterns within that data that might otherwise be difficult to identify. One example is work I recently presented at the BioNow BioAI Symposium, in which we explored the use of structure-activity relationships (SAR) and machine learning models alongside iSLAseizure assay data. Supported by colleagues, including Dr. Louisa Zolkiewski and Dr. Kim Rockley, ApconiX has developed an in-house seizure liability dataset that we analyzed in collaboration with MedChemica. Using machine learning and chemical informatics approaches, we identified chemical substructures that show statistically significant relationships with seizure-related outcomes. The aim is not simply to predict risk, but to understand which molecular features may contribute to that risk and provide medicinal chemists with practical guidance they can apply much earlier in the discovery process. Looking ahead, I think this is where AI can have a particularly powerful impact. Rather than acting as a standalone prediction tool, it can work alongside human-relevant assays such as iSLAseizure to connect molecular structure, experimental biology, and safety outcomes. That combination has the potential to improve decision-making earlier in the discovery process, reduce late-stage safety surprises, and ultimately support the development of safer medicines for patients.

Toxicology is a very diverse field, spanning chemistry, biology, and clinical science. Data varies enormously in accessibility, quantity, quality, terminology, and format. That makes it a complex domain for AI, but also one where AI can be very useful when applied carefully. A realistic AI-assisted safety pipeline today is not about replacing the expert toxicologist. It is about augmentation. AI agents and Large Language Models (LLMs) can autonomously search for data, collate information, generate summaries, and produce plots. This allows safety scientists to bring their expertise to bear more efficiently, while AI helps uncover and organize the evidence. We should also remember task-specific models. DeepDILI, for example, combined classical and deep learning approaches to predict drug-induced liver injury (DILI), a major cause of post-marketing drug withdrawals. Models like this can have a real impact when focused on specific safety pain points.

For molecular design, classical machine learning with traditional fingerprints is not going to generate accurate 3D structures in the way modern graph-based, diffusion, and foundation models can. Where we have large, high-quality datasets, graph neural networks, language models, and foundation models can also outperform classical methods for property prediction. We even see some attempts at generating very large and widely applicable QSAR models, although bespoke models for specific target chemistries and properties remain the main form of QSAR model in use today. However, classical machine learning models are still very useful, especially for small datasets, which are common in early discovery. Classical machine learning models often have fewer parameters and can generalize better in small chemical spaces than larger deep learning models, which may overfit in such spaces. Classical models are also often much faster to train and perform inference with using modest compute infrastructure, such as a laptop. They can also be easier to interpret because many descriptors have a chemical or physical rationale. I do not think classical machine learning belongs in the past. The method pool is broader now, and the researcher's or engineer’s understanding of the data and desired outputs should guide the modeling decision.

Data-driven models, such as Quantitative Structure Activity Relationships (QSAR), are incredibly attractive because they offer rapid methods for predicting molecular properties, including those that are difficult to access through fundamental chemical and physical theory. For new molecules, in related regions of chemical space to the training data, they can often be very accurate. The challenge is that they typically do not generalize across large diverse chemical and biological spaces. Therefore real care is needed when applying them to novel chemistry or biology. Some attempts to build more widely applicable QSAR models have been made in recent years, but most models in use today are still built for specific chemistries and properties owing to the scale and complexity of chemical space. When I started this work, especially for solubility predictions, the datasets were modest by today’s standards, often hundreds to a few thousand molecules. One of the first things that teaches you is to explore the data carefully, assess its quality and quantity, and only then move forward with modeling. The aim is to capture genuine relationships between numerical molecular descriptions and important target endpoints, but those relationships may be only locally generalizable. I also think the phrase “simple descriptors” is interesting. Some are simple, such as atom counts or bond counts, but others rely on detailed parameterization, group contributions, and graph-theoretical techniques. We still see these descriptors used today, sometimes alongside embeddings from graph neural networks or language models. The main lesson for me is that there is no single workflow. Each dataset, model, and end-use case needs careful thought.

AI is moving so quickly that in the gap between answering this and publication, I could be proved wrong! With that caveat, there is a lot of hype around AI drug discovery campaigns. AI is certainly a substantial part of many discovery programs today and has been applied in drug discovery for many years. However, we need to be realistic about what AI is doing, where humans remain central, and how that balance may vary. AI agents are an area that has emerged relatively recently and continues to evolve quickly. They can complete tasks for example data extraction and reporting with guidance from experts, helping to generate larger, more standardized datasets over time. They can also help build closed-loop systems that connect computation and experiment directly, with strategic direction still coming from expert scientists. However, care and oversight is needed with these systems to validate their actions. I also think people often equate AI with generative AI, but there are many other AI systems which are underappreciated. For example, the GNN property prediction models and task-specific predictive tools we discussed earlier can be extremely valuable. Some of these methods receive less attention simply because they have already become familiar in drug discovery.

Having worked across different parts of the discovery process, I think broad exposure changes the way you think. I have developed quantum chemistry, artificial intelligence (AI) and machine learning (ML) methods in research settings, but I have also had the chance to apply them in real discovery environments for materials and pharmaceuticals. That experience has shown me how different parts of the discovery pipeline interact, what their motivations are, and how they communicate. AI and ML have been used in early discovery for many years, but the biggest recent shift is in areas dominated by human processes and communication. Computers can now interact directly with human language and analytical images, opening new applications such as AI research assistants and AI-assisted dossier drafting. That means AI can start to work more collaboratively with scientists, although expert oversight and guardrails remain essential. In nonclinical safety, I think the biggest impact will be AI working alongside safety scientists to unearth key data, connect safety science with earlier discovery and later clinical work, and help process, predict, and summarize important outcomes. That could help safety science integrate more tightly with design, make, and test cycles in discovery, while also interfacing earlier with clinical science. Save a Copy of This Exclusive Interview by Downloading the Free PDF

One major example is protein structure. Methods such as AlphaFold and Boltz-2 have transformed how we generate 3D structural data. Accurate protein structure prediction has been a decades-old challenge, and while there is still work to do on underrepresented protein classes, these models have opened an extraordinary pathway. We have also seen progress in multi-output models. Just over a decade ago (circa 2013), toxicology and property prediction models often focused on one endpoint at a time. Now, multi-output models, especially from methods such as graph neural networks (GNNs) and other deep learning methods, can predict multiple endpoints simultaneously. This allows for transfer learning between related properties. That creates opportunities across ADMET (absorption, distribution, metabolism, excretion, and toxicity) prediction and toxicology. Another exciting area is machine learning potentials. Molecular simulations traditionally rely on fixed mathematical functions to describe molecular interactions, but neural potentials can now replace or augment those functions. They can, in many cases, provide highly accurate simulations and structures at lower computational cost than quantum chemistry. However, its not just down to model architectures. The biggest blind spot behind all of these advances remains careful data curation, annotation, and collection. Open databases for example ChEMBL, PubChem and Chemspider together with data initiatives such as OpenADEMT and OpenBind to name but a few are making notable progress here. We still see some data for example on formulation chemistry often receiving less attention. Its is not just about data scale, but quality that is critical here to. Accessible high quality data sets are the fuel of AI methods. Architecture matters, but high-quality, accessible and abundant data is also absolutely critical.

There has been a regulatory drive to reduce animal testing for some time, including through frameworks such as the European Union’s REACH directive. More recently, we have seen this strengthened with roadmaps to phase out animal testing for chemical safety. In silico and in vitro methods are likely to be central to that transition. These tools can provide efficient, human-relevant, and multiscale evidence. One could imagine an approach similar to the carcinogenicity weight-of-evidence approach, in which in vitro and in silico data are gathered to justify reduced or no animal testing at specific points in the approval process. Regulatory acceptance remains challenging because standards vary around the world. For AI, there are also questions about model approval, model updates, benchmarking, and version control. It will be very interesting to see how regulators test and define these pathways.

Uncertainty and model limitations are critical in safety assessment. There is clearly greater potential for animal or human harm from an incorrect safety assessment than from making the wrong molecule in early discovery, although uncertainty matters in both contexts. It is important to distinguish different types of uncertainty. Aleatoric uncertainty reflects irreducible randomness in a task, while epistemic uncertainty is model-related, such as hallucinations or lack of knowledge. For LLM systems, we use multiple methods to measure consistency and we have built fact-checking approaches in which a separate LLM independently verifies the primary engine's outputs. For QSAR models, including our SAR work around seizure liability, we can use statistical metrics such as confidence intervals to quantify and communicate uncertainty. Understanding that uncertainty exists and communicating it in both safety and discovery is critical. Uncertainty metrics tell us where potential blind spots may exist in our modelling allowing us to take action to improve over time. This applies to both safety and discovery. The aim is not to pretend AI removes uncertainty, but to make uncertainty visible so scientists can make better-informed decisions. Download the Full PDF to Revisit James McDonagh's Insights on AI in Drug Safety Discovery

The tools we are building and applying at ApconiX are currently internal-facing, supporting our scientists and collaborators in delivering thoroughly researched, evidence-based, and authoritative toxicology information in a timely manner. Even though I did not directly interact with ApconiX while I was at Redx, my experience in biopharma strongly shapes how I think about usability. At ApconiX, we constantly consider how others will interpret and use the data and insights we generate. For example, our early SAR work around seizure liability has identified preliminary relationships between chemical substructural patterns and seizure risk. Although this work is in its infancy, these are the kinds of insights that discovery chemists could interpret directly and apply from day one to help mitigate risk. For target safety assessments, the ApconiX data science team have collated a large data sets including our Acuity expression atlas for expression data, enabling our data scientists and bioinformaticians to generate highly informative, data-driven insights and visualizations. By leveraging AI tools, such as LLM workflows and agents, we can efficiently support consistent and thorough analyses, making safety information more accessible and actionable.

Since 6% of young children and 3% to 4% of adults experience food allergies, scientists have been hard at work figuring out exactly what elicits these allergic reactions to foods that should be safe. So far, their efforts have revealed specific proteins in top allergens, like peanut and egg, that cause adverse immune reactions. These proteins are recognized by antibodies, which then activate two of the immune system's fast-acting inflammatory cells, mast cells and basophils. If we know how and what the immune system reacts to during an allergy, shouldn't we know how and what the immune system reacts to during tolerance? Scientists have the "how" mostly covered. Already, there is a solid understanding that regulatory T cells are involved in tolerance. Prior research has revealed the anti-inflammatory and general immune-suppressive role that regulatory T cells play in tolerance, but "what" proteins prompt this non-reaction has remained unknown.

The study began with a bowl of mouse chow. Rather than start piecewise, one food after another, the researchers screened regulatory T cells from mice given a normal diet. They looked for what the regulatory T cells were attaching to, then mapped them backward to specific parts of the chow. They found three proteins-more specifically, they found small, specific bits of those proteins called epitopes-that the regulatory T cells recognized. The epitopes were found in three different food proteins: one from corn, one from wheat, and one from soybean. Notably, all three epitopes are from seed proteins, suggesting that these highly abundant plant proteins are commonly recognized by the immune system's tolerance mechanisms. Furthermore, the most abundant T cells were those reactive to the corn epitope, which makes sense given that corn is not a common allergy. Soy, on the other hand, is one of the major allergies in humans, so the identification of a soybean epitope is especially exciting, notes Blum. Additionally, the mammalian receptor that interacts with the identified soybean epitope also interacts with sesame, helping explain cross-tolerance, or when a tolerance to one food infers a tolerance to another. With the new epitopes identified, the researchers had a few follow-up questions, like where do these regulatory T cells live? And how do they perform in an inflamed versus a healthy environment? They used mice and cell culture models to answer these questions, finding that the regulatory T cells are primarily located in the gut and their activities vary based on whether they are in an inflamed or healthy environment, either working to reduce inflammation or sustain an absence of inflammation.

These seed epitopes are an exciting new addition to our understanding of oral tolerance. Scientists have already considered regulatory T cells as a promising immunotherapy route for people with severe food allergies. It may one day be possible to create regulatory T cells that are pre-programmed to tolerate certain foods and dampen immune responses to common allergens. "Diet is our most intimate interaction with our environment," says Blum. "Correctly recognizing foods as safe creates an anti-inflammatory environment to support nutrient acquisition and prevent allergy. Our research advances scientific understanding of the major dietary allergens, and points us toward future therapeutic interventions that could redirect allergic and autoimmune states." In the less-distant future, the researchers are excited to see their workflow for mapping proteins adapted to humans. The reagent they developed to track their proteins is now available for others to use, so they're hopeful that new insights into regulatory T cell-mediated oral tolerance are soon to come. Blum J. E., et al. (2026). Identification and characterization of dietary antigens in oral tolerance. Science Immunology. DOI: 10.1126/sciimmunol.aeb4684, https://www.science.org/doi/10.1126/sciimmunol.aeb4684.

The endogenous opioids and their receptors are widely distributed throughout the central and peripheral nervous systems, particularly the parts of these systems that regulate pain, emotion, reward, stress responses, motivation, drug addiction, and autonomic control. The differential expression and location of the various receptor subtypes across different neurons account for the wide range of opioid-related behaviors. The activation of µ-opioid receptors is mainly known for playing a role in pain relief. Still, research has also indicated it may be involved in behaviors related to survival, such as appetite and reproduction. The activity of µ-opioid receptors is also known to play a critical role in responses to social stimuli by modulating responses to social rejection or social acceptance, for example. Activation of the δ-opioid receptors and κ-opioid receptors is also known to be involved in pain modulation. Also, studies have shown that NOP activation is involved in pain mechanisms and several behaviors related to psychological stress. Alterations in the endogenous opioid system are suspected to be involved in Parkinson's disease, seizures, neuroprotective mechanisms, and depression.

COTI-2 is a computer-designed small molecule that was discovered by our artificial intelligence drug discovery platform called CHEMSAS. The molecule has undergone extensive preclinical evaluation at multiple centres in Canada, Europe and the USA. It has been well tolerated in preclinical testing and very effective against human tumours with p53 mutations and/or mutations in the PI3K/AKT pathway. We have positive results from more than 10 different human tumour xenograft models using COTI-2 alone or in combination with conventional therapy. COTI-2 is now in late preclinical development being prepared for a Phase 1 study and should be ready for clinical development by the end of this year.

These finding support proceeding to clinical trials with COTI-2 in patients with cancers having p53 mutations. Dr. Mills has established that COTI-2 is potentially effective in a wide range of p53 mutation types.

Extensive data from the experiments have demonstrated that:

I would suggest starting with our web site. You can also follow @CriticalOutcome on Twitter and Facebook.

COTI-2 has an effect on mutated p53 protein that is similar to that documented for a few other members of the same class of compounds. COTI-2 is a third generation thiosemicarbazone specifically engineered for low toxicity. Other drugs in development target the mdm2 protein leading to increased functional p53 by decreasing the destruction of normal p53 protein. COTI-2 also has a unique effect on the PI3K/AKT pathway via a direct gene level up regulation of a specific protein (PIK3IP) that is known to inhibit PI3K. This means that COTI-2 is potentially effective in cancers with p53 mutations as well as in mutations of the PI3K/AKT pathway. This effect has been confirmed in an animal model of human ovarian cancer.

Dr. Mills first evaluated the activity of COTI-2 in a number of human cancer cells lines that we had not previously studied. He was able to evaluate the killing activity of COTI-2 in the context of p53 mutations since the p53 mutation status of these cells was known. He then went on to introduce 32 different types of p53 mutations into mouse embryonic fibroblasts (MEFs), a cell-based system for studying the effect of gene mutations. Once the p53 gene mutations were established in the MEFs the effect of COTI-2 on those cells was evaluated. Changes in normal p53 protein levels together with changes in other proteins known to be direct cellular targets of normal p53 protein were also evaluated for each of 32 different types of p53 mutations following treatment with COTI-2.