Editing Life, One Protein at a Time
Generative models are no longer just proposing novel biomolecules from scratch; they are learning to edit trusted proteins, condition on experimental feedback, and design compounds from phenotypes rather than targets. But rigor lags ambition: publication bias, weak baselines, and immature infrastructure keep autonomous discovery aspirational. We pair the frontier results with the benchmarking and systems work needed to make them trustworthy.
Editing and Steering Proteins
Both entries address a shared limitation in current protein design methods: standard generative and comparison frameworks struggle to handle variable-length edits or to incorporate property constraints without retraining from scratch. Each proposes a general-purpose computational layer that sits alongside or on top of existing models—one for representing and manipulating sequences of arbitrary length, the other for guiding pretrained generative models toward desired traits—and both back their claims with wet-lab validation rather than purely computational benchmarks. They differ in scope and mechanism, however: the first focuses specifically on length-changing edits (insertions, deletions, resizing) using a language-model-derived fixed-dimension representation, while the second targets property steering (single- or multi-objective) across several classes of existing generative architectures without altering sequence length as its primary operation. Their demonstrated use cases also diverge—one shows miniaturized and expanded proteins with altered binding or fluorescence properties, while the other improves editing efficiency of a base editor beyond what directed evolution achieved, and releases accompanying software for broader adoption.
Miniaturizing and modifying natural proteins with Raygun
Raygun tackles a gap in computational protein design—recreating evolution's ability to insert and delete residues, not just substitute them—by encoding protein sequences as fixed-dimension probability distributions built from language model embeddings, making proteins of any length directly comparable. Governed by just two parameters (for substitutions and length change), the framework can shrink proteins by 10–25% (occasionally over 50%), expand them beyond natural size, or diversify their sequences while preserving predicted structure and functional sites. Cell-based experiments validated the approach: Raygun produced miniaturized fluorescent proteins shorter than 96% of entries in FPbase, shrank the biotin ligase TurboID, and generated expanded EGF variants with higher EGFR-binding affinity than wild type—demonstrating that a length-agnostic representation can support large-scale, coordinated sequence edits akin to natural protein evolution.
Property guidance for protein sequence generative models with ProteinGuide
ProteinGuide offers a unifying statistical framework for steering pretrained protein generative models—masked language models (ESM3), any-order autoregressive models (ProteinMPNN), and discrete diffusion/flow-matching models (MultiFlow)—toward desired properties without retraining, effectively enabling "on-the-fly" conditioning on experimental or predictive data. The authors demonstrate it for single- and multi-objective design (e.g., balancing competing properties like Pb versus Zn binding in PbrR), and validate it experimentally by coupling ProteinGuide with wet-lab data to improve the in vivo editing efficiency of an adenine base editor beyond what seven rounds of directed evolution had achieved. Code, a Python package (ProteinGen), and data are publicly released to support broader adoption across generative and predictive modeling workflows.
Discovery Pipelines and Their Data
Both items center on the raw material that fuels AI-driven drug discovery: the biological and chemical data used to guide or train generative models. Each treats data not as a neutral input but as a variable whose quality, structure, or completeness directly shapes what discovery pipelines can achieve. They diverge sharply in focus, however—one presents a specific generative method that substitutes morphological and textual signals for target information to design new molecules, while the other steps back to diagnose systemic weaknesses in the datasets and infrastructure underlying the field as a whole. The former offers a technical solution demonstrated through docking results, whereas the latter raises unresolved structural concerns, from publication bias to data fabrication, that condition whether such methods can be trusted or scaled.
MGMG: Cell Morphology-Guided Molecule Generation for Drug Discovery
MGMG proposes a target-agnostic molecule generation approach that pairs cellular morphological profiles from compound treatments with textual molecular descriptions, sidestepping the usual dependence on known drug targets. Morphology supplies phenotypic bioactivity cues while text provides structural guidance, and the combination boosts generation quality particularly when either signal alone is weak or uninformative. The method extends to genetic perturbations, allowing activator design directly from gene overexpression morphology without needing a reference compound structure. In silico docking shows MGMG-generated molecules achieve binding affinities comparable to reference compounds and retain key interactions while introducing structural diversity, despite no target information being used.
Closing the data loop in AI-driven drug discovery
AI is reshaping early-stage drug discovery by replacing brute-force physical screening with predictive design, letting companies generate and filter candidate molecules computationally before committing to lab work—but Cytiva's Paul Belcher argues this shift is exposing a data bottleneck rather than solving it. Public training datasets are hitting a "data wall": they weren't built for AI, lack structure and diversity, and suffer from severe publication bias, since negative results (failed compounds, non-binders) are almost never shared, leaving models unable to learn what doesn't work. Data fabrication compounds the problem—Belcher cites Elisabeth Bik's 2016 finding that ~4% of biomedical papers contained manipulated images, a risk now amplified by generative AI—prompting tools like Cytiva's blockchain-based Image Integrity Checker to verify image authenticity for publishers. Belcher's endpoint vision is the fully autonomous "lab-in-the-loop," cycling prediction, testing, and optimization in real time, but this depends on interoperable instruments and FAIR data infrastructure that most labs still lack; notably, no AI-designed drug has yet received full FDA approval.
Benchmarks and Infrastructure
Both items grapple with the gap between how systems are conventionally evaluated and what actually determines their performance in practice, arguing that standard metrics — compute-scaling assumptions in one case, LLM-centric benchmarks in the other — obscure the real bottlenecks. Each also pushes toward more operationally grounded measures: DEG-weighted losses and stronger baselines for perturbation models, and task success rate or P95 latency for agent deployments. They differ sharply in domain and source, however: the ICML piece surveys academic research across biological foundation models, multimodal spatial-transcriptomics work, and benchmarking critiques, while the Intel piece is a single vendor-sponsored infrastructure study focused on enterprise agent capacity planning. The former is diagnostic and exploratory, cataloging open questions about what models learn, whereas the latter is prescriptive, offering concrete deployment recommendations like scale-out defaults and agent-density planning.
Reflections on ICML 2026
AI for biology cemented itself as a major ICML theme this year, jumping to 5th place among subject areas (2.5% of accepted papers) and earning three dedicated workshops plus a keynote from Aviv Regev. Several trends stood out: biological foundation models increasingly rely on custom tokenization schemes (dnaHnet, scBIG, STACK, ST-JEPA) rather than adapting language-model conventions, though evidence that transcriptomic FMs lack clean data-scaling laws suggests sparse learnable signal, not compute, may be the real bottleneck. Perturbation prediction for transcriptomics drew heavy attention from both academia and industry, with flow-matching approaches (PerturbDiff, scBIG) and neural-ODE-based methods (PerturbODE, which also infers gene regulatory networks) all claiming state-of-the-art results, underscoring a need to identify which strategies actually win. A parallel wave of benchmarking and position papers interrogated "what models actually learn," including work showing linear autoencoders can beat nonlinear ones on test loss despite missing recoverable latent structure, and a proposed signal-dilution explanation (plus a DEG-weighted loss fix) for why mean baselines still outperform large FMs on perturbation prediction — reinforcing calls for simpler, more rigorous baselines. Multimodal work centered on linking spatial transcriptomics to histopathology (MoLF, ST-JEPA, Regev's SCHAF and PerturbPair preprints), pointing toward future
Building the enterprise environment for agentic AI
Intel's systems-level study of agentic AI workloads—based on thousands of experiments using an extended version of the open-source Terminal-Bench harness—argues that enterprise agent deployment is a capacity-planning and infrastructure problem, not merely an LLM inference challenge. Key findings: capacity should be planned by agent density (agents per vCPU) rather than raw agent count, since bursty compute patterns make average CPU utilization a poor monitoring signal (P95 task latency is proposed as a better leading indicator); scale-out is recommended as the default deployment strategy over scale-up, reserved for workloads with heavier per-agent compute needs. Intel also proposes six operational metrics (task success rate, cost per task, time per task, throughput, agent density, latency) to replace LLM-centric evaluation, and suggests the technology's near-term value lies in automating already-codified workflows like ticket triaging, regression testing, and security review rather than open-ended experimentation. Note this is Intel-sponsored content, not MIT Technology Review's independent reporting.