Layla Hosseini-Gerami

Michael, I agree with your assessment that to discover and develop successful drugs, it requires more than AI tools. It requires experienced scientists and leaders with actual experience delivering safe and effective drugs are complete regulatory review and approval. Many promises have been made about how AI will improve the success rate and speed to discovery. My publication in 2010 in Science Translational Medicine discussed them and still has relevance to achieve success based on fundamental principles of confidence in rationale and confidence in safety. I hope you all the best and success in raising all the funding you need! Best regards.

5

1 Comment

** Data rant incoming ** Your digitised data is worthless without structure and contextualisation and hence your whole AI strategy on top of this unstructured data is prone to fail. So, if you are only prepared to fund fancy AI solutions and neglect the need for continuous(!) funding of adequate data strategies and architecture you will always be disappointed by the outcomes of your AI solutions. I have often seen how loose money is for nice new robots and more recently AI solutions, but when you want to 'save' some money, the data strategies come first. #Data #AI

83

3 Comments

The recent rise of histopathology foundation models has drastically changed what is possible to achieve in diagnostic AI - having spent years researching the challenging clinical task of ovarian cancer subtyping, I was genuinely shocked to see the impact these models made. In our new preprint (https://lnkd.in/eceZrKy8), we performed the most comprehensive single-task validation of histopathology foundation models (and the largest ovarian cancer AI subtyping study) to date. It was absolutely clear that any of the modern ViT-based foundation models offered a drastic improvement over the traditional approach of using generic feature extractors. We specifically found that H-optimus-0 gave the best performance, with several other models giving comparable performance with reduced computational requirements (Lunit, Kaiko-B8, UNI). The performance of these models has given us confidence that AI will be able to make a real clinical impact in the diagnosis of ovarian cancer in the coming years, which previously felt like a very distant prospect.

122

1 Comment

Want faster neural networks without new hardware? Shrink them to 3% of the bits and 18% of the weights! This article delves into a groundbreaking method for making neural networks more efficient by significantly reducing their size. Ever wondered how to make neural networks more efficient without needing specialised hardware? The recent paper on Self-Compressing Neural Networks by Szabolcs Cséfalvay and James Imber offers a fascinating solution. The idea is simple yet powerful: reduce the size of neural networks by removing redundant weights and minimising the bits needed to represent the remaining ones. In their experiments, they achieved floating point accuracy with just 3% of the bits and 18% of the weights. This means faster execution times, lower power consumption, and reduced memory footprint—all without compromising performance. For those grappling with the challenges of deploying large-scale neural networks, this approach could be a game-changer. It simplifies the process, making it more accessible and efficient. If you're struggling with the complexities of neural network optimisation, this might be the breakthrough you need. Dive into the details and see how these insights can transform your projects. Check this article: https://lnkd.in/deZE3JYS #NeuralNetworks #MachineLearning #AIResearch #DeepLearning #TechInnovation #DataScience #Efficiency ====================== Hit 'Connect' now - I may have opportunities for business and jobs in the future If you're interested in content like this, hit 'Observe' button now.

1

This article exploring the concept of the ‘curse of dimensionality’, some interesting phenomenon and the maths associated with it, was a quick and fun Sunday evening read. In the realm of biomedical data and bioinformatics, where high-dimensional datasets are the norm, this curse manifests as data sparsity, overfitting, and computational complexities. It underscores the importance of feature selection, dimensionality reduction techniques, and robust statistical methods in analyzing complex biological data to derive actionable insights. #BiomedicalData #Bioinformatics #DataScience #CurseofDimensionality 🔗 https://lnkd.in/eYEMdDUn

6

Another busy week in the world of AI.. I've summarized the key points for you here 🔥👇 👉 UK PM Starmer set to introduce new 'AI Bill' - the new government looks to move slightly more aggressively with regards to AI regulation. I wonder what the ramifications of this will be... 🧐 👉 The DNA Sequencing world gets a shake up! Element Biosciences raise a $277 million series D. It's to help with the commercialization of existing products, but also for a new upcoming launch. Watch out Illumina.. 👉 Roche gets the go ahead for AI powered Glucose monitoring Systesm 💥Whilst technically not an AI Drug, it's still good to see AI making tangible impacts on peoples lives! 👉 Flagship Pioneering close a new fund worth $3.6 billion 😲The VC group are really committed to AI. Their new fund is aimed at supporting the development of 25+ start-ups in the Biotech space, with a reported focus on AI! 😎 👉 Inceptive CEO Jakob Uszkoreit appears on CNBC to discuss the impact of Generative AI in Drug Development. It's safe to say him and the Inceptive team are bullish on the impact AI will have on Drug Discovery - I can't wait to see what they do! 😁 Follow #beyondthemodel to stay up to date with the ever shifting world of AI and please let me know your thoughts below 👇😁🙌

23

3 Comments

Rory Cellan-Jones kickstarts his AI in the NHS series. The focus and hype around AI - in the NHS or elsewhere - can sometimes be misplaced. AI can absolutely improve things for patients and healthcare professionals. Used pragmatically, it could have a transformative impact on the structural capacity problems in the NHS. The key is focusing on scalable interventions, allowing people to focus on what people are good at - caring for each other. We're reducing did not attends, and helping the most vulnerable patients using prediction algorithms within our patient engagement platform. Staff can use this to identify potential missed appointments and plan accordingly. Clinics run better and patients who need help are more likely to get it. I encourage everyone to read our conversation on Rory's Substack. Looking forward to seeing the next edition of the series! https://lnkd.in/eHaStpHt #healthtech #aiinthenhs #healthcare #futureofoutpatients

63

5 Comments

Testing out new Chat GPT 4.0... It's remarkably good at summarising and producing structured answers. Hey Chap GTP, What impact do you think iPSCs will have on ALS drug discovery? ChatGPT Induced pluripotent stem cells (iPSCs) are expected to significantly impact amyotrophic lateral sclerosis (ALS) drug discovery in several key ways: Disease Modeling: iPSCs can be generated from patients with ALS and differentiated into motor neurons, the cell type primarily affected by the disease. This allows researchers to create patient-specific models of ALS in the lab. These models are invaluable for studying disease mechanisms and for screening potential drug candidates in a context that closely mimics human pathology. Personalized Medicine: Because iPSCs can be derived from individual patients, they facilitate the study of genetic variability in ALS. Researchers can explore how different genetic mutations contribute to the disease and how patients with different genetic backgrounds respond to treatments. This paves the way for personalized medicine approaches, where treatments can be tailored to the genetic profile of individual patients. High-Throughput Drug Screening: iPSC-derived neurons can be used in high-throughput screening to test thousands of compounds for potential therapeutic effects. This approach can rapidly identify promising drug candidates that can then be further validated in more complex models and eventually in clinical trials. Understanding Disease Mechanisms: iPSCs provide a platform for investigating the cellular and molecular mechanisms underlying ALS. By differentiating iPSCs into various neural cell types, scientists can study the interactions between motor neurons, astrocytes, and other cell types involved in ALS pathology. This comprehensive understanding is crucial for identifying new therapeutic targets. Testing Drug Toxicity and Efficacy: iPSC-derived motor neurons can be used to test the toxicity and efficacy of drug candidates before they enter clinical trials. This helps ensure that only the most promising and safest drugs move forward in the drug development pipeline, potentially reducing the time and cost associated with bringing new therapies to market. In summary, iPSCs offer a versatile and powerful tool for ALS drug discovery. They enable the creation of disease models that closely replicate human pathology, facilitate personalized medicine, support high-throughput drug screening, improve our understanding of disease mechanisms, and offer new avenues for testing drug safety and efficacy. These advantages collectively accelerate the pace of discovery and increase the likelihood of finding effective treatments for ALS. #ipscs #chatgpt

30

2 Comments

Ever wonder how AI is betting progressively better at understanding human language, including the nuances of medical language (still something of a work in progress)? Embeddings are the key! Basically a series of of floating-point numbers (think of long decimal values that are either positive or negative), they can translate complex medical terms into a numerical form that AI can process. Embeddings can capture relationships between words: Knowing "heart attack" is closer to "cardiomyopathy" than "common cold." Link to the interactive visual in the comments.

4

3 Comments

Yet another diffusion based protein structure prediction work. It's great to see many tools now available for protein engineering. What a fantastic time to be working in this area! Article: De novo protein design with a denoising diffusion network independent of pretrained structure prediction models Abstract: The recent success of RFdiffusion, a method for protein structure design with a denoising diffusion probabilistic model, has relied on fine-tuning the RoseTTAFold structure prediction network for protein backbone denoising. Here, we introduce SCUBA-diffusion (SCUBA-D), a protein backbone denoising diffusion probabilistic model freshly trained by considering co-diffusion of sequence representation to enhance model regularization and adversarial losses to minimize data-out-of-distribution errors. While matching the performance of the pretrained RoseTTAFold-based RFdiffusion in generating experimentally realizable protein structures, SCUBA-D readily generates protein structures with not-yet-observed overall folds that are different from those predictable with RoseTTAFold. The accuracy of SCUBA-D was confirmed by the X-ray structures of 16 designed proteins and a protein complex, and by experiments validating designed heme-binding proteins and Ras-binding proteins. Our work shows that deep generative models of images or texts can be fruitfully extended to complex physical objects like protein structures by addressing outstanding issues such as the data-out-of-distribution errors. https://lnkd.in/eNYvV8GR

75

AI tools such as ChemCrow and DU8ML are changing the way we approach chemistry, but how to protect these innovations is an ongoing legal debate. Last November the UK High Court declared an artificial neural network (ANN) patentable for the first time and in the months since, the UK Patent Office saw a boom in applications for AI-related inventions. However, the Court of Appeal has now reversed this judgment, reclassifying ANNs as computer programs and therefore excluded from patentability. Worryingly, the court’s new definitions of ‘computer’ and ‘computer program’ also threaten the protection of existing technologies used across the research sector, including the circuitry involved in analytical equipment such as NMR machines and mass spectrometers. In my latest story with Chemistry World, Bruce Dearling explains how loose definitions and a lack of technical understanding brought the court to this decision and Keith Pickford shares what this could mean for future innovation in the UK. https://lnkd.in/epTNf4PK #AI #Chemistry #ANN #Patentability #IPI #EmotionalPerceptionAI #HepworthBrowne #ChemCrow #DU8ML

12

1 Comment

Can network biology help us find molecules to target disease states? Find out in the lastest Computing Biology post, where I look at network communities in the context of drug repurposing to find some really interesting things. #drugdiscovery #computationalbiology #machinelearning https://lnkd.in/ekWRTJf8

26

Major upgrade to #AlphaFold! The latest tool, AlphaFold 3, AI models how proteins interact with other molecules — but Google DeepMind restricts access to the tool. The latest version of AlphaFold, described on 8 May in Nature (https://lnkd.in/gsmgmJiv) gives scientists the ability to predict the structures of proteins during interactions with other molecules. But whereas DeepMind made the 2021 version of the tool freely available to researchers without restriction, AlphaFold3 is limited to non-commercial use through a DeepMind website. https://lnkd.in/gmKzM6zV

9

New blog: How AI can help reform UK NHS The UK government wants to reform the UK health system by digitisation, shifting care to communities, and focusing on prevention. I think there is a lot of potential for AI to help with this, if AI/Medicine researchers become less fixated on using AI to improve diagnoses in hospitals. https://lnkd.in/eN2ZfTsk

19

💊 This is going to be a game changer in the field of drug discovery. Michael Peel from Financial Times discusses discusses the unveiling of AlphaFold 3 by Google DeepMind, an advanced AI model designed to enhance understanding of the fundamental building blocks of life and their interactions within cells. This latest iteration of the AlphaFold technology, initially developed in 2018, provides the most detailed predictions yet on the structure and behaviour of biological molecules, expanding its analysis beyond proteins to include DNA, RNA, and ligands—key components in cellular processes and disease markers. 🧠 Developed in collaboration with DeepMind's drug discovery arm, Isomorphic Labs, and in partnership with pharmaceutical giants like Eli Lilly and Company and Novartis, AlphaFold 3 offers researchers the tools to rapidly identify potential new drug molecules. This significantly accelerates the hypothesis testing and drug design process, which traditionally could take months or years. The model’s predictive accuracy has shown considerable improvement over existing specialized tools, including its predecessors. 👩‍🔬 AlphaFold 3 has potential applications beyond drug discovery, such as improving knowledge of plant biology and enhancing food security. Despite these advancements, the predictions made by AlphaFold 3 still require experimental validation and must undergo normal clinical trial processes. DeepMind plans to make the majority of AlphaFold 3’s functionality available for free to academic non-commercial users, which could transform the way biological research and experiments are conducted. The model also presents some limitations, such as difficulties with chiral molecules and occasional inaccurate predictions, which are addressed by assigning confidence measures to its predictions. 📰 A study by Boston Consulting Group (BCG) suggests that drugs discovered by AI have a higher rate of success in the earlier stages of trials compared to drugs discovered through alternative methods. ⚕ Sir Demis Hassabis (DeepMind’s CEO and co-founder) believes that we must understand the different interactions of cells and molecules to fully understand biological properties, and in his own words “And you can think of AlphaFold 3 as our first big step towards that.” https://lnkd.in/eX7Un-pf

12

2 Comments