Shift Bioscience

Why hasn't AI cured diseases yet? What does it actually take to create the next 100 blockbusters? Learning from pharma & startup successes and failures. 🎙️ From Signal to Drug — Where It Actually Breaks @dives86 — CEO, @ShiftBioscience @jackcastle — CBO, @OchreBio Mythili Iyer

I'm excited to be speaking at @RNID's Hearing Therapeutics Summit 2025, which is helping to speed the development of therapeutics for hearing loss and tinnitus. With a strengthening link between age related hearing loss and the underlying aging process, hearing loss has become a

AI Virtual Cell vs Linear Model—who wins? 🤖 ⚔️ 📈 Our preprint, entitled "𝘋𝘦𝘦𝘱 𝘓𝘦𝘢𝘳𝘯𝘪𝘯𝘨-𝘉𝘢𝘴𝘦𝘥 𝘎𝘦𝘯𝘦𝘵𝘪𝘤 𝘗𝘦𝘳𝘵𝘶𝘳𝘣𝘢𝘵𝘪𝘰𝘯 𝘔𝘰𝘥𝘦𝘭𝘴 𝘋𝘰 𝘖𝘶𝘵𝘱𝘦𝘳𝘧𝘰𝘳𝘮 𝘜𝘯𝘪𝘯𝘧𝘰𝘳𝘮𝘢𝘵𝘪𝘷𝘦 𝘉𝘢𝘴𝘦𝘭𝘪𝘯𝘦𝘴 𝘰𝘯 𝘞𝘦𝘭𝘭-𝘊𝘢𝘭𝘪𝘣𝘳𝘢𝘵𝘦𝘥 𝘔𝘦𝘵𝘳𝘪𝘤𝘴", seeks to answer this question. Paper ▸ biorxiv.org/content/10.110… Code ▸ github.com/shiftbioscienc… 🎮 𝐋𝐚𝐜𝐤 𝐨𝐟 𝐜𝐨𝐧𝐭𝐫𝐨𝐥𝐬 Every wet-lab biologist knows that positive and negative controls are fundamental to assess whether an assay or experiment worked. However, the genetic-perturbation modeling field has been lacking these anchors to judge whether a model is actually learning the task. While the dataset mean is often used as a negative control, we propose an 𝐢𝐧𝐭𝐞𝐫𝐩𝐨𝐥𝐚𝐭𝐞𝐝-𝐝𝐮𝐩𝐥𝐢𝐜𝐚𝐭𝐞 baseline as a positive control, approximating the best achievable performance for a given dataset. 📐 𝐌𝐞𝐭𝐫𝐢𝐜 𝐦𝐢𝐬𝐜𝐚𝐥𝐢𝐛𝐫𝐚𝐭𝐢𝐨𝐧 With positive and negative controls, we analysed 14 perturbation datasets to see which metrics best separate the two. We call this difference the 𝐃𝐲𝐧𝐚𝐦𝐢𝐜 𝐑𝐚𝐧𝐠𝐞 𝐅𝐫𝐚𝐜𝐭𝐢𝐨𝐧 (𝐃𝐑𝐅). Strikingly, widely used metrics like MSE and Pearson Δ (relative to control) often show low DRF, indicating limited sensitivity to perturbation signals. Weighted MSE and normalized inverse ranking perform well. 🧠 𝐃𝐞𝐞𝐩 𝐥𝐞𝐚𝐫𝐧𝐢𝐧𝐠 𝐩𝐞𝐫𝐟𝐨𝐫𝐦𝐚𝐧𝐜𝐞 Using well-calibrated metrics, most deep-learning models outperform linear baselines—and even additive models on combinatorial tasks. This holds from GEARS and scGPT to MLPs built on foundation-model embeddings. Recently, a 𝐭𝐫𝐨𝐯𝐞 of papers has cast doubt on the utility of deep learning for building the so-called AI Virtual Cell. Here, we show these models learn useful biology when evaluated with well-calibrated metrics. Looking forward to the bright future of the field! Congrats to the first authors @Henrymiller2012, @g27182818, and @Fleblanc_3 for the absolutely fantastic work at @ShiftBioscience, alongside @BrendanMSwain and @BoWang87! 🔥

🚀 We’re hiring a Computational Biologist (Toronto) At Shift Bioscience, we’re uncovering the biology of cell rejuvenation to develop safe, effective therapies that reverse cellular aging 🧬 . Our platform integrates single-cell biology and machine learning to identify and

What is the state of research on the emerging grand challenge of virtual cell modeling? 1/n open.substack.com/pub/behindbiom…

Do deep generative models in single-cell omics really work for perturbation prediction? Some benchmark studies say yes: 🔗 arxiv.org/pdf/2408.10609 🔗biorxiv.org/content/10.110… Others say no: 🔗 arxiv.org/abs/2410.13956 🔗 BMC Genomics: bmcgenomics.biomedcentral.com/articles/10.11… To move beyond this

Are you trying to build a so-called AI virtual cell? 🔬 Yet ... the mean still outperforms your perturbation-response prediction model. 😮‍💨 Our paper— just accepted to the 𝗚𝗲𝗻𝗕𝗶𝗼 𝗪𝗼𝗿𝗸𝘀𝗵𝗼𝗽 @ 𝗜𝗖𝗠𝗟 —dives into 𝘸𝘩𝘺 the mean excels and what you can do about it. Paper ▸ arxiv.org/abs/2506.22641 Code ▸ github.com/shiftbioscienc… 🧪 𝗘𝘅𝗽𝗲𝗿𝗶𝗺𝗲𝗻𝘁𝗮𝗹 𝗳𝗮𝗰𝘁𝗼𝗿𝘀 We modelled perturbed single-cell RNA-seq data 𝘪𝘯 𝘴𝘪𝘭𝘪𝘤𝘰 to see which experimental variables inflate mean-baseline performance under common metrics (MSE and Pearson-Δ, computed for all genes and for the top-20 DEGs). Two stood out: 𝗰𝗼𝗻𝘁𝗿𝗼𝗹 𝗯𝗶𝗮𝘀 and 𝘄𝗲𝗮𝗸 𝗽𝗲𝗿𝘁𝘂𝗿𝗯𝗮𝘁𝗶𝗼𝗻 𝗲𝗳𝗳𝗲𝗰𝘁𝘀. Analyses of the datasets Replogle ’22 and Norman ’19 confirmed these trends: ↑ bias + ↓ perturbation effect = ↑ mean performance. 🏜️ 𝗠𝗲𝘁𝗿𝗶𝗰 𝗺𝗶𝗿𝗮𝗴𝗲𝘀 Pearson-Δ is usually referenced to control cells, introducing a systematic bias that lets the mean perturbed profile perform well. Our fixes: • 𝗪𝗲𝗶𝗴𝗵𝘁𝗲𝗱 𝗠𝗦𝗘 (𝗪𝗠𝗦𝗘), a DEG-weighted metric that can also be easily incorporate into model training. • 𝗪𝗲𝗶𝗴𝗵𝘁𝗲𝗱 𝗥² Δ, referenced to the 𝘮𝘦𝘢𝘯 𝘰𝘧 𝘱𝘦𝘳𝘵𝘶𝘳𝘣𝘢𝘵𝘪𝘰𝘯𝘴 instead of the control, capturing both the 𝘥𝘪𝘳𝘦𝘤𝘵𝘪𝘰𝘯 and 𝘮𝘢𝘨𝘯𝘪𝘵𝘶𝘥𝘦 of Δ without the control bias. 🔍 𝗖𝗮𝗹𝗶𝗯𝗿𝗮𝘁𝗲𝗱 𝗯𝗮𝘀𝗲𝗹𝗶𝗻𝗲𝘀 For fair evaluation, we suggest three anchors: 1. 𝗡𝗲𝗴𝗮𝘁𝗶𝘃𝗲 – control mean. 2. 𝗡𝘂𝗹𝗹 – mean of all perturbations; the bare-minimum a model should beat. 3. 𝗣𝗼𝘀𝗶𝘁𝗶𝘃𝗲 – 𝘵𝘦𝘤𝘩𝘯𝘪𝘤𝘢𝘭 𝘳𝘦𝘱𝘭𝘪𝘤𝘢𝘵𝘦: predict one half of a perturbation’s cells from the other half; an empirical upper bound set by the dataset noise. 🧠 𝗙𝗼𝗰𝘂𝘀𝗲𝗱 𝘁𝗿𝗮𝗶𝗻𝗶𝗻𝗴 = 𝗯𝗲𝘁𝘁𝗲𝗿 𝗯𝗶𝗼𝗹𝗼𝗴𝘆 Swapping MSE for WMSE lifted a perturbation model such as GEARS out of mode-collapse to capturing real biological variation—even in difficult zero-shot, unseen-gene settings. Centering Δ on the mean of all perturbations, using DEG-weighted losses, and benchmarking against these calibrated baselines offers a robust recipe for perturbation modelling. With 𝗔𝗿𝗰 𝗜𝗻𝘀𝘁𝗶𝘁𝘂𝘁𝗲’𝘀 𝗻𝗲𝘄 𝗩𝗶𝗿𝘁𝘂𝗮𝗹 𝗖𝗲𝗹𝗹 𝗖𝗵𝗮𝗹𝗹𝗲𝗻𝗴𝗲 kicking off, the community needs rigorous metrics and baselines more than ever—making this paper particularly timely. Shout-out to the co-first authors @Henrymiller2012 and @g27182818 for accelerating the discovery of cell rejuvenation genes @ShiftBioscience!

RT @dives86: @ShiftBioscience proposes an improved ranking system for virtual cell models to accelerate gene target discovery and drive rej…

🆕 New LEVITY episode is live! Why tune in? Because a 24-person Cambridge startup may have beaten the billion-dollar OSK race with a single gene found by AI-powered virtual cells - and the early data look game-changing. (Links as always below.) Just about the hottest thing in

Thankyou @peterottsjo and @DrPatrickLinden for hosting me on the LEVITY @reachlevity podcast to discuss our preprint 'A single factor for safer cellular rejuvenation' and beyond! youtube.com/watch?v=vdOVKZ…

RT @dives86: @SheekeyScience breaks down the @ShiftBioscience preprint 'A single factor for safer cellular rejuvenation' highlighting both…

Genuine epigenetic rejuvenation in primary cells has long been the holy grail. A groundbreaking preprint reveals that over-expression of a single (secret) gene overcomes this barrier: greatly reduced age estimates across in fibroblasts and keratinocytes according to validated epigenetic clocks including the Skin&Blood clock (Horvath 2018) and the original pan-tissue clock (Horvath 2013). In keratinocytes, this gene decreased the pan-tissue clock by nearly ten years for each month of treatment! Longitudinal sampling confirmed age REVERSAL. This gene seems to outperform even the Yamanaka factors (OSKM) while crucially avoiding pluripotency induction and its associated cancer risks.  Lucas Paulo de Lima Camillo,  Daniel Ives,  Brendan M. Swain (2025) A single factor for safer cellular rejuvenation. biorxiv.org/content/10.110…

@eythorarnalds @lucascamillomd 1. Confirm a broader range of cell types can be rejuvenated 2. Demonstrate rejuvenated cells exhibit the same behaviour as younger cells 3. Progress to proof-of-concept rejuvenation studies in mouse models

@lucascamillomd @ShiftBioscience this is amazing 🙌🏻 - what are the next steps ?

Rare moments make you stop, stare, and imagine their future impact. I’m thrilled to share one of those moments from our work at @ShiftBioscience. Over the last few years, the data often felt unreal. Today we compare the single-gene SB000 (Shift Bioscience 000) with the gold-standard Yamanaka factors OSK(M). Key highlights -⏳ Efficacy – in fibroblasts, SB000 is comparable to OSK in transcriptomic rejuvenation, including decreased senescence markers, and it even outperforms the cocktail across dozens of epigenetic clocks. - ⛑️ Safety – SB000 preserves transcriptomic signatures of fibroblast identity and no pluripotent colonies are observed, in sharp contrast to OSK(M). - 🌐 Generalisability – Rejuvenation extends to cells from a different germ layer. In keratinocytes, PCHorvath2013 fell by over 10 years and DunedinPACE by more than 20% over six weeks. This is only a glimpse of what we’re cooking at Shift with @BrendanMSwain , @dives86, and the rest of the team. Super excited for the future of radical cellular rejuvenation! Paper: biorxiv.org/content/10.110…

Single-gene rejuvenation target promises safer cell age reversal @ShiftBioscience #longevitybiotech #aging longevity.technology/news/single-ge…

RT @dives86: The dream of waking up in the near future and being substantially younger at a cellular level is something I have the rare pri…

I’m excited to welcome Lord David Prior and Sir Tony Kouzarides to the team at @ShiftBioscience! Joining as Chair of the Board, David brings a wealth of Board experience, including as Chairman of NHS England, to support the Company’s long term mission to comprehensively reverse

I'll be speaking at Founders Longevity Forum in London on 10th June at OXO2! Let’s connect in London — whether you're building something new in Longevity biotech, investing in the future, or just curious about where the field is going next. Register your interest here:

I’m excited to be part of a #SynBioBeta2025 panel discussing breakthroughs in epigenetic reprogramming and how this could transform Longevity biotech. Let’s connect in San Jose — whether you're building the next big thing in Longevity biotech, investing in the future, or just

1/ We're thrilled to announce Shift Bioscience @ShiftBioscience as a sponsor of Vitalist Bay - our 8-week longevity zone dedicated to combating aging at the cellular level using groundbreaking AI technology. 🔬🧬💻