Skip to content

AI in drug discovery

Structure prediction, generative design and virtual cells: verifiable progress for AI in discovery and the clinic.

144 selectedRelated topicsMethods and resourcesSmall molecules and degradersOriginal research

Latest picks

61–80 of 144
Oct 5Mon
  1. youtube.com64

    CSHL interview: Elham Azizi on AI models for mapping the tumor microenvironment

    Cold Spring Harbor Laboratory interviewed Columbia's Elham Azizi about using machine learning, from probabilistic models to generative and foundation models, to characterize the tumor microenvironment from spatial transcriptomics, including deconvolution of mixed signals in sequencing-based platforms like Starfish. In ER-positive breast cancer, her cell-interaction algorithms suggest cytotoxic T cell infiltration correlates with worse outcomes, possibly because T cells increase tumor dependence on estrogen receptor programs, pointing to better treatment combinations. Her long-term goal is a TME foundation or world model that predicts patient trajectories across treatment doses and schedules, requiring large paired pre- and post-treatment cohorts and out-of-distribution generalization.

    Why it matters: Shows how spatial transcriptomics and foundation models could turn TME analysis into predictive treatment selection.

  2. youtube.com52

    Cross-Species Optic Nerve Single-Cell Atlas With AI Annotation Maps Glia Subtypes in Glaucoma

    A cross-species single-nucleus RNA-seq atlas of the retina and optic nerve covers about 150k nuclei across 5 species, integrating data into a shared space to identify conserved and species-specific cell types. AI-assisted annotation reached 99%, 95% and 90% accuracy at different confidence tiers, speeding manual curation. The work identifies AS1 and OPC3 glial subpopulations changed in the optic nerve and glaucoma context, pointing to glial roles in optic nerve damage.

    Why it matters: A 150k-nucleus, 5-species optic nerve atlas links glial subtypes to glaucoma and speeds annotation with AI.

  3. youtube.com72

    Jennifer Doudna says AI accelerates nuclease design but cannot replace chance discovery

    Jennifer Doudna discussed the Viper system, a new class of programmable nucleases discovered and named by her lab from studies of microbial defense systems, which can be engineered for gene editing and differs structurally from Cas proteins. She described its non-contiguous guide encoding, in which functional elements are not adjacent in sequence, making it harder for AI to learn without biological priors. She said generative and structural models can predict activity and narrow mutation search spaces, but chance observations still drive many breakthroughs.

    Why it matters: A CRISPR pioneer defines what AI can and cannot do in nuclease design for gene editing.

Oct 4Sun
  1. youtube.com64

    Maria Brbic maps multimodal AI integration strategy at CSHL AI in biology symposium

    At the CSHL AI in biology symposium, EPFL's Maria Brbic outlined a strategy for integrating multimodal biomedical data into unified AI representations: train strong single-modality encoders first, then bridge modalities with scarce paired data. Her lab's Luna model learns from spatial transcriptomics how tissue assembles and infers where dissociated single cells sit in tissue; a related work links single-cell transcriptomes to morphology via Iris. She also described a regularization method that makes small paired datasets trainable and warned that data diversity, not volume, plus leakage control, determines model capability.

    Why it matters: A leading lab's concrete recipe for bridging scarce paired multimodal biology data directly affects how foundation models get built and validated.

  2. youtube.com60

    Anshul Kundaje on Interpretable Genomic Machine Learning as a Baseline for AlphaGenome

    In a CSHL Leading Strand interview, Stanford's Anshul Kundaje argues for lightweight, interpretable sequence-to-activity models trained per assay, which learn enzyme-preference bias first and strip it away to leave latent biology, a framework that helped AlphaGenome avoid learning such bias. These models capture DNA grammar beyond GC content, such as cooperative binding of two proteins on the same side that breaks at 10bp spacing, and nucleosome-scale spacing near 150bp that drives nucleosome eviction. They also serve as strong baselines for large models and can find causal mutations in rare-disease patients, including variants visible only in early developmental or disease-specific cell contexts.

    Why it matters: A leading genomics ML researcher explains why interpretable lightweight models remain essential alongside models like AlphaGenome.

  3. youtube.com75

    Single-cell immune dynamics predict infliximab response before treatment in IBD and rheumatoid arthritis

    Shai Shen-Orr argues the immune system gains new functions by rearranging interactions among existing cell types rather than adding new types, and applies single-cell dynamics to predict drug response. In an IBD cohort, fibrosis-organizing pathway genes differentially expressed in intermediate monocytes predicted infliximab response with AUC 0.78. The same gene set predicted infliximab response in rheumatoid arthritis blood across three independent datasets, suggesting non-response is conserved across diseases and specific to the therapy.

    Why it matters: A pre-treatment gene signature predicts infliximab response across two diseases and three independent datasets.

  4. youtube.com68

    Peking University Team Trains Protein Generative Model Claiming to Beat AlphaFold3 on Complexes

    Peking University's Iching Gao presented a protein generative model trained from scratch on Huawei Ascend 910A hardware with MindSpore rather than PyTorch or Jax. With only 2-3 residue contact or interface restraints, the model reportedly outperforms AlphaFold2 and AlphaFold3 on complexes, though these are speaker claims needing independent validation. The talk also covered joint structure-sequence generation, illustrated with ubiquitin and SUMO, and a Chinese-language agent platform, Galileo, that orchestrates some 100 software tools for tasks like peptide drug candidate design.

    Why it matters: Independently trained model claims to beat AlphaFold3 on complexes with minimal restraints, pending replication.

  5. youtube.com62

    CSHL Symposium on AI in Biology: Interview with Marinka Zitnik

    Marinka Zitnik of Harvard tells a CSHL Leading Strand interview that AI for biology should use "skeptical reasoning" and report uncertainty rather than force conclusions when evidence conflicts or is lacking. She says literature-only models inherit heavy bias toward the most popular genes and must draw on raw perturbation data such as perturb-seq chemical screens. Her foundation model Proton, without additional training, was applied to bipolar disorder and then Alzheimer's disease, checking expression differences of approved-drug targets in relevant cell types for drug repurposing.

    Why it matters: Explains a concrete approach to uncertainty-aware AI and multimodal hypothesis generation in drug discovery.

  6. youtube.com60

    GV20 Founder Shirley Liu on AI Rebuilt Cancer Drug Discovery and Antibody AI's Shortfalls

    GV20 Therapeutics founder Shaoli (Shirley) Liu described how AI is reshaping anticancer drug discovery in an ImmuneZoom lecture. She argued antibody AI lags small-molecule AI because antibody structure prediction is hard, with 10^18 possible sequences and no experimentally validated de novo target-antibody design yet. Her STEAD platform retrospectively validated targets including EGFR, HER2, PD-L1 and PD1, predicted anti-PD1 antibodies with near-matching affinity before seeing clinical sequences, and has a molecule, IGS8, in phase 1 that turns cold tumors hot, with a phase 2B planned.

    Why it matters: A leading AI-drug-discovery founder lays out what antibody AI can and cannot yet do, with pipeline timelines.

  7. youtube.com68

    Rino Rappuoli traces vaccine R and D from empiricism to AI-driven design

    In a Global ImmunoTalks lecture hosted by Stanford's Bali Pulendran, Rino Rappuoli, former head of R&D at GSK and Novartis and founder of reverse vaccinology, reviewed how vaccine development moved from empiricism to AI-driven design. He described reverse vaccinology's computational screening that identified 2158 genes from the meningitis B genome, the RSV vaccine approved in 2023 after decades of failure, and synthetic biology enabling vaccines built from a sequence alone, with AI as the natural next step in this computational lineage.

    Why it matters: A leading vaccinologist explains why AI is the endpoint of vaccine design's two-century technological lineage.

  8. youtube.com68

    Martin Steinegger interview: organizing the protein universe with AI structure prediction at billion scale

    In a CSHL interview, AlphaFold2 and ColabFold co-author Martin Steinegger describes organizing the protein universe at scale using AI-predicted 3D structures rather than sequences. AlphaFold DB holds about 240 million structures and BioHub has released 1.1 billion metagenomic predictions; clustering the universe by domain combinations yields roughly 2 to 3 million groups for finding distant homologs, functional motifs and enzymes.

    Why it matters: A leading practitioner explains how billion-scale predicted structures turn distant homology and motif discovery into routine analysis.

Oct 3Sat
  1. mp.weixin.qq.com60

    State method predicts cell perturbation responses across contexts without large-scale experiments

    A Cell paper proposes a method called State for predicting how cells respond to perturbations across different cellular contexts, without requiring large-scale single-cell perturbation experiments. Current approaches such as single-cell CRISPR screening or small-molecule perturbation deliver transcriptome-level readouts in a given context, but their high cost limits scaling across contexts, and existing deep learning models often fail to outperform simple linear models when generalizing perturbation effects across contexts.

    Why it matters: State could replace costly large-scale single-cell perturbation experiments with computational prediction across contexts.

Oct 2Fri
  1. Nature Biotechnology65

    Deep learning perturbation models can outperform baselines on calibrated metrics

    Standard perturbation-modeling benchmarks can be misleading because control bias and signal dilution let uninformative mean baselines score highly; in Replogle22 K562 GWPS, with only 3.25 DEGs per perturbation on average versus 102.09 in Norman19, the technical-duplicate positive control underperformed the mean baseline on MSE in roughly 95% of perturbations. The authors propose an interpolated duplicate baseline blending both controls and a meta metric, the dynamic range fraction (DRF), to quantify how well benchmarking metrics distinguish positive from negative controls across 14 datasets. This reframes evaluation of transcriptomic perturbation models away from MAE, MSE and Pearson of control-referenced deltas alone.

    Why it matters: It supplies positive and negative controls and a calibration metric that change how deep learning perturbation models should be benchmarked.

  2. youtube.com78

    MIA: Samuel King, Generative Genome Design; primer by Brian Hie

    At the Broad Institute MIA seminar on September 16, 2026, Brian Hie presented the Evo 1 and Evo 2 genome language models, and Samuel King demonstrated generative genome design based on Evo 2 on ΦX174 with synthesis validation.

    Why it matters: The talk showed an Evo 2 and flow-matching-based generative genome design method, producing an infectious complete phage on ΦX174, informing genome-scale de novo design.

Oct 1Thu
  1. Nature59

    Function-preserving watermarking of AI-generated proteins

    SynthIDBio embeds detectable, function-preserving watermarks directly into AI-designed proteins. SynthIDBio-sequence integrates watermarking into ProteinMPNN and produced biologically watermarked, functional binders in vitro: low nanomolar binders for the SARS-CoV-2 receptor binding domain and subnanomolar binders for VEGF-A and PD-L1, with no effect on binding affinity distributions or hit rates. SynthIDBio-structure fine-tunes AlphaFold 3 to emit watermarked structure predictions with near-perfect detectability and negligible loss of global accuracy, establishing provenance tracking for AI-generated sequences and structures without central databases.

    Why it matters: First demonstration of watermarking AI-designed proteins that preserves experimentally verified binding function and structural accuracy.

Sep 30Wed
  1. reddit.com59

    MIT used AI to develop new recipe for mRNA vaccines so they no longer need freezers and can sit at room temp for a full year or survive about 100F heat for two months which could finally let these vaccines reach poor countries without cold storage and even work as easy skin patches instead of shots.

    MIT engineers used machine learning to screen a new lipid nanoparticle formulation that keeps mRNA vaccine nanoparticles stable at room temperature for about a year and at roughly 100°F for about two months, aiming to eliminate cold-chain dependence. The post links to an MIT News report; the exact temperature ranges and stability data should be checked against the original paper. Commenters debated whether the method counts as modern large-model AI and discussed potential effects on vaccine logistics in low-income countries.

    Why it matters: The post explains in plain language how machine learning screens lipid nanoparticle formulations to improve mRNA vaccine heat stability, comparable with the MIT paper for assessing cold-chain and distribution impact.

  2. bioRxiv Genomics57

    Disease relevance and replicability of deep learning gene expression prediction

    Deep learning models that predict gene expression from DNA sequence owe their high overall correlation mostly to capturing whether genes are on or off, not to disease-relevant changes in expression among active genes; performance rises with how bimodal a gene's expression distribution is. The authors also found inconsistencies between published code and reported performance, and conclude that personalized genomic prediction in disease remains far from solved.

    Why it matters: Reported performance of popular gene-expression prediction models overstates their usefulness for disease research.

Sep 29Tue
  1. medicalxpress.com56

    New genetic testing guidance could improve patient diagnosis worldwide

    RNA splicing tests can be trusted to classify genetic variants when well designed, according to evidence-based recommendations published in Genome Medicine. Reviewing more than 41,000 variants across 5,458 genes, the authors found minigene assays agreed fully or highly with patient RNA testing in 89% of comparable cases, while only one of seven high-throughput datasets met clinical standards. Variants producing 80% or more abnormal RNA strongly indicated a disease-causing effect, and 20% or less argued against one.

    Why it matters: It sets quantitative thresholds and design standards labs can use now to resolve uncertain genetic test results.

  2. Nature Medicine29

    Reproducibility in biomedical research

    Nature Medicine has published a review on reproducibility in biomedical research, posted online on 29 September 2026. The review examines current efforts to harmonize reproducibility terminology and implementation, the barriers that persist, and how the research ecosystem is responding, noting that reproducible practices are still not consistently applied despite being foundational to trustworthy science.

    Why it matters: A Nature Medicine review maps why reproducible practices remain inconsistently implemented across biomedical research.

Sep 28Mon
  1. Nature Biotechnology56

    AI-guided optimization for thermostable mRNA vaccines

    AGENT, an AI-driven framework combining high-throughput experimentation with Bayesian optimization, engineers thermostable solid-state mRNA vaccines. The formulations retained full bioactivity after 2 months at 37 °C and elicited immune responses non-inferior to fresh vaccines. Microneedle patches enabled cold-chain-free delivery in rodents and nonhuman primates.

    Why it matters: Cold-chain-free mRNA vaccines with retained bioactivity could change vaccine distribution logistics.

CBACBA Frontier Insights (CFI)Chinese Biopharmaceutical Association (CBA)www.cba-usa.org

Disclaimer CBA Frontier Insights uses AI-assisted methods to aggregate, summarize, and analyze publicly accessible biopharmaceutical information. Content is provided for informational and educational purposes only and does not represent the official views of original sources or affiliated organizations. AI-generated content may contain errors; please verify important information against primary sources. Intellectual property rights remain with their respective owners. For copyright concerns or corrections, please contact us through the website’s feedback page.