BioByte 061: AI for Alzheimer's prediction, NeurIPS 2023 recap, gene therapy for hearing loss, autonomous chemical discovery, and 3D molecular generation
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!
Full speed ahead for bio. We’ve got you covered:
FDA Approves AI Software for Alzheimer's Prediction: The FDA approved "BrainSee," an AI software by Darmiyan, predicting Alzheimer's disease (AD) progression in individuals with mild cognitive impairment. Utilizing MRI data and cognitive test scores, BrainSee estimates AD dementia risk within five years, with an accuracy of 88% to 91%.
NeurIPS 2023 Highlights on Generative Protein Design: Simon Barnett from Dimension Capital recapped ML x Bio research, focusing on protein design models like IgDesign and AntiFold. Key developments include diffusion models for conditional protein modeling and combining ML with energy-based models for faster and more efficient protein binding simulations.
Gene Therapy Enables Hearing in a Child: Akouos, acquired by Lilly, developed AK-OTOF, a gene therapy that successfully treated an 11-year-old boy with congenital hearing loss due to a mutation in the otoferlin gene. The therapy, involving a dual AAV-based system, restored hearing across all frequencies within 30 days. This breakthrough establishes gene therapy as a viable treatment for congenital hearing loss and opens possibilities for treating a range of sensorineural hearing conditions.
Advancements in Molecular Discovery and Spatial Proteomics Imaging: MIT and the Broad Institute developed an autonomous chemical discovery platform, significantly advancing small molecule and dye discovery. PEPSI, a novel technique, analyzes tumor microenvironments by measuring subcellular protein localization, improving predictions of patient survival outcomes and enhancing precision medicine. And finally, the study of Lingo3DMol in Nature Machine Intelligence highlights the generation of 3D molecules for drug design, marking a significant step in structure-based drug development.
Key Deals in the Biotech Industry: Notable partnerships announced include Novo Nordisk's licensing deal with EraCel Therapeutics and GenEdit's collaboration with Genentech. Sanofi's acquisition of Inhibrx for $2.2 billion and BridgeBio's $1.25 billion funding for commercializing its oral heart pill are significant M&A and financing movements. The biotech sector continues to see active investment with IMU Biosciences, Accent Therapeutics, and Locus Biosciences among those securing substantial funding for innovative projects.
What we read
FDA Approves BrainSee, an AI Software That Purportedly Predicts AD [AlzForum, January 2024]
The FDA recently approved BrainSee, an innovative AI software developed by Darmiyan, designed to predict the likelihood of Alzheimer's disease (AD) progression. This technology, targeting individuals with mild cognitive impairment, uses MRI data combined with cognitive test scores to estimate the risk of developing AD dementia within five years.
Renowned researchers like Nick Fox of University College London have highlighted the significance of hippocampal atrophy, a key MRI feature, in predicting AD from amnestic Mild Cognitive Impairment (aMCI). Although BrainSee's AI capabilities extend beyond this single measure, its effectiveness, with a reported accuracy between 88% to 91%, still demands further peer-reviewed studies for comprehensive understanding.
Kate Papp of Brigham and Women’s Hospital emphasizes the need for a multi-modal approach in predicting dementia, integrating data from various sources including blood biomarkers, cognitive function, and genetics. Despite the promising role of blood biomarkers in detecting AD pathology, they are still awaiting regulatory approval.
AI's ability to handle large computational tasks makes it ideal for identifying subtle brain atrophy patterns, a precursor to AD. BrainSee, leveraging AI, has been trained on extensive brain scan data and cognitive scores. It generates an AD dementia probability score, ranging from 0 to 100, based on inputs from standard whole-brain MRI, MMSE (Mini Mental State Examination), and CDR-SB (Clinical Dementia Rating scale sum of boxes).
Accessible to a wide range of physicians, BrainSee offers a quick and user-friendly platform for calculating the probability of AD dementia. Priced at $1,500, it is offered at a reduced cost of $300 pending Medicare coverage. The software not only serves as an initial screening tool but also assists in guiding further diagnostic and treatment decisions for patients with varying risks of AD dementia.
NeurIPS 2023 Roundup: Generative Protein Design [Simon Barnett, Dimension Capital, Jan 2024]
Dimension’s Head of Research Simon Barnett published Part 1 of a 3 part series recapping ML x Bio research at NeurIPS last month. The piece is both sufficiently technical and accessible, and is well worth a read. Some highlights:
Many of the best performing protein design models (eg IgDesign from Absci, AntiFold) incorporate inverse folding modules (which predict 1D amino acid sequence based on 3D structure) which constrain generated antibody sequences to only those sequences that can fold into a given structure, for a 3D structure that forms a tight antibody-antigen complex. One additional advantage of this approach is that data distributions of protein structure are smoother and therefore easier to optimize.
Diffusion models for protein design are shifting towards conditional modeling, in which the generative process can be steered towards specific properties or structures of interest. For example, a method called motif scaffolding for generating a protein scaffold around a fixed, desired functional motif of interest, such as a specific enzyme active site.
The conference also featured some exciting work combining ML with energy-based models which learn the energetics and thermodynamics of protein conformations. DSMBind uses ML to learn a data-driven binding energy function which is orders of magnitude faster than traditional physics-based methods such as Schrodinger’s FEP+ which require compute-intensive molecular dynamic simulations of the protein binding process. The efficiency of DSMBind makes it more appropriate for virtual screening applications or large libraries.
Gene Therapy Allows an 11-Year-Old Boy to Hear for the First Time [Gina Kolata, NYT]
Congenital hearing loss – loss of hearing at birth – is one of the most common conditions in children. It has several etiologies, ranging from congenital infections and malformations, to inherited genetically-defined loss of certain proteins crucial for hearing. Several mutations can cause congenital hearing loss, one of which is in the otoferlin gene; otoferlin is a transmembrane protein responsible for the exocytosis of synaptic vesicles at the auditory inner hair cell ribbon synapse. Loss of function mutations in otoferlin cause profound sensorineural hearing loss, and treatment (including hearing aids and cochlear implantations) have limited efficacy in improving symptoms.
Cue Akouos – a biotech company founded in 2016, and subsequently acquired by Lilly in 2022 – who developed AK-OTOF, a dual AAV-based gene therapy meant to replace defective otoferlin. It uses AAVAnc80, a capsid with high transduction for inner hair cells (the sensory receptors of the ear, containing the majority of auditory nerve fibers that project to the brain), and engineered with a strong, ubiquitous promoter to drive expression of otoferlin. In combination, they developed a novel, minimally invasive surgical delivery approach that effectively delivers AAV throughout the length of the cochlea, providing AAV to each hair cell.
The intervention worked exceptionally well in an 11-year-old boy with profound hearing loss from birth, who experienced restored hearing within 30 days of AK-OTOF treatment, with hearing restored across all testing frequencies. This work establishes gene therapy as a powerful way to treat congenital hearing loss, and sets the stage for future interventions across sensorineural hearing loss.
Autonomous, multiproperty-driven molecular discovery: From predictions to measurements and back [Koscher et al., Science, 2023]
Why it matters: efficient exploration of chemical space is critical to identify new small molecules, polymers and materials with desirable properties. Currently, there is a gap between in silico chemical property prediction and retrosynthesis planning tools and the chemical automation synthesis platforms. In this paper, the authors combine both to create a (semi)autonomous chemical discovery and measurement platform.
The authors from MIT and the Broad Institute demonstrated the platform in two discovery use cases for small molecule organic dyes: i) exploration of unknown chemical spaces and ii) exploitation of a known chemical space. Human intervention was kept to a minimum: managerial issues (such as error recovery) and custodial (providing consumables). The platform attempted >3000 reactions with >30% yielding the predicting reaction product and also synthesizing 303 unreported dye-like molecules.
The platform automates across the DMTA cycle:
Front-end predictions: the platform proposes candidate structures based on specific conditions, plans reaction pathways and reaction conditions, predicts candidate properties and selects candidates to realize. For each generated candidate, multiple synthesis pathways are automatically planned; using retrosynthesis recommendations to ensure candidates are synthesizable. By executing multistep reaction pathways, more of the multidimensional property space becomes accessible.
Automation and experimental execution: reaction pathways are translated into synthesis and characterization workflows to be executed in 96-well plates. Reaction execution, analysis of reaction outcomes, isolation of target products and characterization of the isolated molecules are automatically carried out. The platform has access to a liquid handler, a HPLC with mass spec, a robotic arm and a unit that contains a storage carousel, a high temperature realtor, a plate reader and photodegradation reactor.
In both case studies, the platform improved the property prediction quality after several iterations but they warn that “although the models’ predictions improved for each scaffold, there was little change in predictive ability across general chemical spaces, demonstrating the local nature of these structure property spaces” and “despite a sizable number of datapoints from literature, the initial property models still did not extrapolate well to unexplored scaffolds.” The platform could also be further automated via the use of reinforcement learning to develop new operations.
PEPSI: Polarity measurements from spatial proteomics imaging suggest immune cell engagement [Wu et al., Pacific Symposium on Biocomputing 2024]
Why it matters: PEPSI's ability to detect and quantify subtle changes in subcellular protein expressions offers a groundbreaking approach in precision medicine. By effectively linking immune cell states to patient outcomes, it opens new avenues for understanding and potentially altering disease progression. This technique could revolutionize how clinicians assess and treat tumors, moving closer to personalized treatment strategies based on individual cellular landscapes.
This study introduces PEPSI (Protein Expression Polarity Subtyping in Immunostains), a novel approach for analyzing tumor microenvironments. It focuses on measuring subcellular protein localization to characterize immune cell states within tumors. Applied to over two million cells from 600 patient samples, PEPSI utilizes immunofluorescence imaging data to define surface protein polarity, identifying distinct immune cell states. Incorporating these polarity-defined subtypes significantly improves the performance of deep learning models in predicting patient survival outcomes. This method provides insights into immune cell functional states and their relationship to disease prognosis.
Generation of 3D molecules in pockets via a language model [Feng et al., Nature Machine Intelligence, January 2024]
Why it matters: Lingo3DMol's advancement in generating 3D molecules is pivotal for structure-based drug design. Its ability to accurately predict molecular properties and binding modes can significantly accelerate the discovery of effective therapeutic compounds. This technology could lead to more efficient drug development processes, potentially reducing costs and time-to-market for new medications.
The study introduces Lingo3DMol, a novel AI-driven method for 3D molecular generation in drug design. It surpasses traditional models by efficiently creating molecules with desirable 3D structures and drug-like properties. Lingo3DMol incorporates a new molecular representation, FSMILES, and a non-covalent interaction predictor, enhancing its ability to generate molecules with accurate topology and spatial positions. The method demonstrated superior performance in drug likeness, synthetic accessibility, and binding mode prediction, using the Directory of Useful Decoys-Enhanced dataset for evaluation.
Partnerships / Licensing:
Novo Nordisk has licensed an experimental oral small molecule from EraCel Therapeutics for an aggregate of €235M, continuing to back fill its obesity pipeline. EraCel’s platform enables zebrafish-based phenotypic drug discovery.
Unnatural products has inked a $220M partnership with Merck, to tackle an oncology target using their synthetic macrocycle platform.
Delivery start-up GenEdit partners with Genentech for $15M upfront to to create new nanoparticles for nucleic acid-based medicines for autoimmune diseases.
Sanofi acquires Inhibrx for an aggregate $2.2B for its lead program INBRX101 - a Phase II human recombinant protein designed to treat alpha-1 antitrypsin deficiency. Inhibrx’s other programs will be spin into a separate entity called New Inhibrx which Sanofi will retain 8%
IMU Biosciences secures £11.5M Series A to build an AI-powered immune atlas
Accent Therapeutics raises $75M to enter the clinic with its novel DHX9 and KIF18A inhibitors
Korean Mirae launches $50M U.S biotech fund for Seed to Series C start-ups
Comanche Biopharma raises $75M to advance its siRNA-based therapy against pre-eclampsia
BridgeBio raises $1.25B from investors to fund the commercialisation of its oral heart pill acoramidis and to continue running clinical trials across its portfolio. The financial breaks down to: $500M in cash from Blue Owl in exchange for 5% royalty on acoramidis future sales, $450M credit facility and the possibility of a further $300M in incremental facilities to support strategic pipeline expansion.
Ratio Therapeutics raises $50M to develop imaging agents and radiopharmaceutical candidates for various tumors.
Locus biosciences received $24M from BARDA to conduct phase II study of its CRISPR-enhanced bacteriophage therapy for UTI treatment
What we listened to
What we liked on socials channels
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