BioByte 076: AlphaFold3, engineering antibody agonists, NfL as a biomarker for neuro disease, diffusion models for molecular linker design
Welcome to Decoding Bio, a writing collective focused on the latest scientific advancements, news, and people building at the intersection of tech x bio. Happy decoding!
What we read
Blogs
Rational Drug Design with AlphaFold3 [Google Deepmind, Isomorphic Labs, 2024]
Back in October, the Isomorphic Labs and Google Deepmind teams released a teaser regarding the capabilities of the next generation of AlphaFold. Compared to its predecessor AlphaFold 2 which could only predict the structure of static proteins, the now fully released AlphaFold 3 can predict the static structure of a protein in complex with a range of biological modalities such as ions, small molecules, antibodies and nucleic acids. This enables powerful applications such as structure-based-drug-design and further understanding of the protein in its biological context. The latest release includes a different model to the teaser but with only slightly better performance. The team have released a limited AF Server to test out simple folds but have not released the code. Whether the teams will release the code remains a question, AlphaFold2’s code was released, enabling significant strides in the scientific community and leading to the Deepmind team receiving substantial recognition for their work. However, Isomorphic Labs has two high profile pharma partnerships based on their AI capabilities and so this may hinder the ability for the teams to fully released their suite of tools. Whilst very exciting, it still remains to be seen whether such models can directly enable the design of a novel therapeutic against a difficult to drug target.
Receptors Unite: Agonism with Antibodies [Dimension Research, May 2024]
Antibody therapeutics have revolutionized medicine, and are now one of the most important therapeutic modalities alongside small molecules. Essentially all of the approved antibodies act as antagonists, including the cancer immunotherapy pembrolizumab (Keytruda), which was the top selling drug in 2023, generating >$25B in revenue.
This recent essay from Simon Barnett at Dimension Capital dives into the emerging landscape of agonist antibodies. There are a number of challenges when designing agonist antibodies. Binding of a receptor by an antibody is a necessary but not sufficient step for agonism. Agonists must also bind multiple receptors to induce a conformational change or receptor complex formation (eg dimerization) to activate the receptor. Commonly used technologies for designing antibody antagonists such as phage display or ML methods for antibody design focus do not necessarily extend to agonists as these methods are primarily focused on designing binders. Functional and/or activity-based assays are required to go a step further and confirm agonist activity.
Novel antibody formats are being explored to address challenges in engineering antibody agonists. One notable innovation, termed "Contorsbody", involves introducing linkers that bend the Fab arms of an antibody, such as trastuzumab, into a compact, multivalent structure that promotes receptor dimerization rather than blocking it, demonstrating a proliferative effect on Her2+ breast cancer cell lines. Another approach by Genentech involves creating "i-shaped antibodies" that have additional hinge linkages to align the Fab arms parallelly, enhancing receptor agonism in models like OX40 and IL-2 pathways crucial in cancer and autoimmunity. These iAbs show significant potential in promoting agonism in a concentration-dependent manner, with effects on cell proliferation and gene expression profiles similar to native ligands, indicating their potential as receptor agonists across various receptor types.
ChatNT: A Multimodal Conversational Agent for DNA, RNA and Protein Tasks [Instadeep, 2024]
Instadeep, an AI solutions firm previously acquired by BioNTech, has unveiled ChatNT, a multimodal conversational agent. The model has been developed to tackle a myriad of biological tasks, leveraging DNA, RNA and protein sequences.
The need for foundational models is clear: there is an increasing abundance of raw data due to decreasing sequencing costs yet labeling data is still time-consuming and expensive. So by using techniques such as masked- or next-token prediction, deep learning models can work on unlabeled data. However, such models typically require fine-tuning to be effective on a specific task and are usually difficult to use or interpret.
ChatNT integrates a pre-trained DNA coder, an English language model decoder, and a novel projection layer to seamlessly bridge the gap between DNA and language embedding spaces. To validate its efficacy, researchers curated a comprehensive genomics instructions dataset spanning 28 tasks across DNA, RNA, and proteins from various species, framed in natural language. The model is also conversational like ChatGPT, increasing the ability for wider adoption.
The world is relying on the United States to get value-based drug pricing right [Shafrin et. al, Stat News, May 2024]
This article emphasizes the urgent need for the U.S. to refine its approach to drug pricing, adopting a method that prioritizes value. The piece sheds light on the significant role the U.S. plays in influencing global pharmaceutical innovation through its pricing strategies. Traditionally, America has supported innovation by allowing drug prices to be set based on the costs of R&D. This method, however, often leads to prices that don't necessarily match the actual health benefits of the drugs.
The push for value-based pricing is about ensuring that drug prices are directly linked to their real-world health and economic benefits. This approach aims to make sure that the pricing not only reflects the value these drugs bring to patients and the broader society but also keeps pharmaceutical companies motivated to invest in new, effective treatments. The authors argue that if drugs are priced below their value, it might cut down on innovation, whereas pricing them above their value could mean that consumers are paying too much.
The article also discusses the need for a wider view in health technology assessments (HTAs). Currently, many countries, including the U.S., focus these assessments on the cost-effectiveness of new drugs from the healthcare payer's perspective. This often overlooks broader benefits such as the impact on productivity or easing the burden on caregivers, which should be factored into the value assessments to get a full picture. The authors advocate for a broader, more dynamic approach to drug pricing that considers the complexity of diseases, the pharmaceutical market's nature, and the extended impacts beyond just the health system. By moving towards this more holistic evaluation, the U.S. can better align its drug pricing policies to ensure they're both supporting ongoing medical innovation and delivering true value. This approach is crucial not only domestically but globally, as U.S. drug pricing strategies have a profound impact on global access to new therapies and the direction of pharmaceutical R&D.
What if drugs can affect NfL by clearance? [Eric Minikel, CureFFI.org]
Neurofilament light chain (NfL) is a well-known structural filament in neurons, and has been recently established as a biomarker for neural damage (FDA recently granted accelerated approval to Tofersen for SOD1 ALS using decreased NfL as a surrogate biomarker). NfL leaks out of neurons in pathologic processes, and makes this molecule useful in the diagnosis of disease, and in monitoring treatment.
Eric reports on data from his cohort study at MGH (evaluating people at risk for prion disease, of which they routinely evaluate NfL levels) where they noticed a patient had a 5.7x spike in CSF NfL in a short period of time. They learned that prior to the spike, the patient had received a 6 week course of minocycline, a widely used antibiotic in dermatology. Interestingly enough, prior trials had evaluated minocycline in neurological conditions as it inhibits microglial activation; in those patients, NfL levels went up, and it was hypothesized that minocycline damaged neurons. However, those patients did not have any negative impact on clinical outcomes, and another surrogate biomarker (GFAP) was not worse, either.
Eric and team subsequently ran a proteomics study on the spiked NfL patient, and found that the CSF proteome after minocycline only altered NfL and NfM (neurofilament medium chain), with little to no alterations in anything else. One would expect widespread dysregulation in the setting of neurodegeneration – suggesting that there may be a novel mechanism to minocycline’s effect on NfL. They additionally probed plasma NfL levels in a set of patients who were prescribed minocycline, and found that some patients had extremely elevated plasma NfL. In a mouse study, they found that treating mice with minocycline resulted in a 4-fold increase in plasma NfL after just 6 days of treatment, followed by a rapid drop after drug cessation. Proteomics profiling also was inconsistent with the hypothesis that minocycline results in neurodegeneration. This was supported by human cell culture studies, in which minocycline was not found to be toxic to IPSC-derived human neurons and microglia, but increased NfL levels 3-fold.
Given the above data, Eric and team hypothesized that minocycline actually inhibits clearance of NfL from CSF and plasma, resulting in a significant increase. This has significant implications across neurology drug development – the field assumes that alterations in NfL are an accurate reflection of neuronal damage (↑ = damage, ↓ = protection). However, there may be exceptions to this rule, and it will be important to elucidate true mechanisms of NfL mediated alterations by emerging drugs.
Academic papers
Equivariant 3D-conditional diffusion model for molecular linker design [Igashov et al., Nature Machine Intelligence, April 2024]
Why it matters: Fragment-based drug design (FBDD) has been used as an alternative to screening molecules that bind to relevant target proteins. The geometries for the identified fragments are crucial for the design of drugs. Computationally determining fragments and its linkers is cheaper and more efficient than other screening methods. In this paper, the authors developed a model to link fragments in a 3D context conditioned by the target protein pocket.Fragments, smaller molecular compounds (<20 heavy atoms), are screened to find those that can dock to a target protein. Once the relevant bound fragments are identified, they are connected to form a single chemical compound. Current computational methods for linker design (search and physical simulations) are computationally intensive. In addition, these models are not equivariant with respect to the permutation of atoms, can only combine pairs of fragments and don’t take into account the target protein pocket.
The authors introduced DiffLinker, a conditional diffusion model that generates molecular linkers for a set of input fragments represented as a 3D atomic point cloud. DiffLinker first generates the size of the linker and samples initial linker atom types and positions. Then, the linker atom types and coordinates are updated using a neural network conditioned on the input fragments. Finally, the denoised linker atoms and input fragments form a single compound. The generated linkers are also constrained by conditioning it on the target protein pocket, generating molecules that are structurally compatible with the pockets. DiffLinker demonstrated SOTA synthetic accessibility and drug-likeness.
Notable Deals
Flagship-backed Prologue Medicines raises $50M to use viral proteins in therapeutics - the company aims to exploit the fact that viruses have developed proteins over millions of years that are effective at evading our various immune defenses or tricking our cells to produce them, and so some of these proteins could be repurposed to tackle diseases such as cancer. The company will use a database of 6.2M viral protein sequences and use LLMs to predict their function and effect on human biology.
OverT Bio launches with $16M to search for genes that can help T cells become more effective in solid tumors. The company has a proprietary screening tool to achieve this - it was published in Nature in 2022.
BridgeBio spins out its oncology unit of RAS assets - the biotech has created BridgeBio Oncology Therapeutics and supplied it with $200M in fresh funding from a suite of high profile investors.
Reunion Neuroscience raises $100M+ to advance its synthetic psilocin post-partum depression therapeutic
Delphia raises $67M to develop “activation lethality” therapeutics - the concept focuses on over activating certain cancer-linked cell signaling pathways that lead to tumor cell death
Novartis buys Mariana Oncology for $1B - The swiss pharma is considered a pioneer in the space and Mariana sees this as a chance to tap into Novartis’ expertise in scale and commercialisation of radiotherapeutics.
What we liked on socials channels
Field Trip
Did we miss anything? Would you like to contribute to Decoding Bio by writing a guest post? Drop us a note here or chat with us on Twitter: @ameekapadia @ketanyerneni @morgancheatham @pablolubroth @patricksmalone
Fantastic read as always!!