Proteins on Demand, Agents Under Test
Protein language models are graduating from generating novel sequences to editing the proteins scientists already trust—resizing them, tuning their phase behavior, and steering them toward target properties. Running underneath is a recurring theme: the payoff depends less on bigger models than on better data loops, coordination, and measurable reliability. We pair the latest life-sciences advances with hard looks at agentic systems and the data problems still holding AI-driven discovery back.
Editing the Proteins We Already Trust
These three projects share a starting point of a natural protein or biological behavior that already works, then use machine learning not to invent something new from scratch but to reshape, steer, or explain what already exists—resizing a scaffold, nudging a generative model's output, or predicting when a sequence will phase-separate. Each pairs a computational representation (length-agnostic embeddings, statistical guidance over pretrained generators, or embeddings fused with environmental parameters) with experimental validation to show the approach captures something mechanistically real rather than just statistically convenient. They diverge sharply in what aspect of protein behavior they manipulate: one changes overall sequence length while preserving function, another steers generation toward specified biochemical properties like binding or editing efficiency, and the third predicts and tunes condensate-forming behavior under varying physical conditions. They also differ in application target—engineering tools like fluorescent proteins and base editors versus disease-relevant phenomena like alpha-synuclein aggregation—reflecting distinct problems even though all rely on protein language model representations as a common substrate.
Miniaturizing and modifying natural proteins with Raygun
Raygun tackles a gap in computational protein design: engineering large-scale, coordinated sequence changes (substitutions plus insertions/deletions) that mimic natural evolution, rather than relying on substitution-only methods that become combinatorially intractable beyond a few dozen sites. Its key trick is representing proteins as fixed-dimension probability distributions derived from protein language model embeddings, making sequences of any length directly comparable and enabling controlled shrinking, expansion, or diversification via just two tunable parameters. Experimentally validated in cells, Raygun miniaturized fluorescent proteins (shorter than 96% of entries in FPbase) and the biotin ligase TurboID while retaining function, and expanded EGF to produce variants with higher EGFR-binding affinity than wild type—demonstrating that function can be preserved in a length-agnostic representation across substantial size changes (10–25%, sometimes over 50% reduction, or expansion beyond natural size).
Property guidance for protein sequence generative models with ProteinGuide
ProteinGuide offers a unifying statistical framework for steering pretrained protein generative models—spanning masked language models (ESM3), any-order autoregressive models (ProteinMPNN), and discrete diffusion/flow-matching models (MultiFlow)—toward user-specified properties without retraining the underlying model. The authors demonstrate on-the-fly conditioning using experimental data to boost single properties like stability or activity, as well as joint optimization of two competing objectives (e.g., improving Pb binding while reducing off-target Zn binding in PbrR). Most notably, combining ProteinGuide with iterative wet-lab data generation produced an adenine base editor with higher in vivo editing efficiency than one obtained through seven rounds of directed evolution, and the method is released as an accompanying Python package (ProteinGen) alongside code and data.
A machine learning framework for predicting and modulating condition-dependent protein phase separation
Most phase-separation predictors rely on sequence alone, ignoring that condensate formation shifts with concentration, temperature, and solvent conditions. LLPSense, developed by a KAIST team, addresses this by combining protein language model embeddings with explicit environmental parameters to yield condition-aware predictions, validated experimentally. Notably, the model uncovered previously unrecognized temperature-dependent reentrant phase behavior in SGTA and correctly predicted which α-synuclein mutations (linked to Parkinson's disease) enhance or suppress phase separation. The authors further show model-guided mutagenesis can be used to deliberately tune phase behavior, positioning LLPSense as a tool for both mechanistic disease research and designed condensate engineering.
The Data and Reliability Bottleneck
Closing the data loop in AI-driven drug discovery
AI is reshaping early-stage drug discovery by shifting hit identification from brute-force empirical screening toward predictive design, allowing companies to computationally filter out low-quality candidates before physical testing—but Cytiva's Paul Belcher argues this is straining lab infrastructure and exposing a deeper data problem. Because public training datasets skew almost entirely toward positive results (Belcher cites Elisabeth Bik's 2016 finding that ~4% of biomedical papers contained manipulated images, a problem now worsened by generative AI's ease of fabrication), models lack exposure to failed experiments and non-binding compounds, limiting their reliability and reinforcing bias; Belcher notes tools like Cytiva's Image Integrity Checker, which uses blockchain-style hash algorithms to detect tampered images, are gaining traction with publishers as a partial fix. He frames the end goal as autonomous "labs-in-the-loop" that continuously cycle prediction, testing, and optimization to feed better data back into models—but this depends on interoperable instruments and FAIR data flows that most labs currently lack, since many lab systems remain closed and siloed. No AI-designed drug has yet received full FDA approval, underscoring that these efficiency gains remain unproven at the clinical endpoint that drives most drug-development costs.
Engineering Agents That Coordinate
Both items treat multi-agent AI as an infrastructure and coordination challenge rather than a matter of improving individual model quality, arguing that reliability comes from the surrounding system—protocols, authorization, capacity planning—rather than from the agents' own reasoning. Each grounds this claim in empirical testing: one reports dramatic swings in group decision-making success tied to coordination protocols, the other runs thousands of workload experiments to isolate infrastructure effects from LLM variability. They differ in scope and register, however: the first focuses on cross-agent trust, identity, and communication standards (AGNTCY, Mycelium, CASA) aimed at an industry-wide architectural gap, while the second is a vendor-sponsored operational playbook offering concrete sizing and monitoring metrics for enterprise deployment. Correspondingly, one is pitched toward standard-setters addressing security and coordination failures across open systems, and the other toward operations leaders needing SLA-driven guidance on specific, already-codified workflows.
The path to artificial superintelligence
Cisco's Outshift argues that scaling AI further requires a horizontal axis, not just bigger models: getting independently capable agents to coordinate goals, memory, and reasoning without human intermediaries. Vijoy Pandey cites a study finding 41–87% failure rates among open-source multi-agent systems, framing this as an architectural gap rather than a prompting issue, and describes Outshift's stack—the connectivity layer AGNTCY (now under the Linux Foundation), the coordination protocol Mycelium, and the authorization tool CASA—as attempts to fix it. Internal testing reportedly showed structured coordination protocols raised successful group decisions from about a third to 93% across 14 scenarios, while Pandey claims roughly 90% of the time an agent can't confirm it's cleared for the task it was handed, underscoring new security risks like prompt injection and over-privileged access that these tools aim to address.
Building the enterprise environment for agentic AI
Intel ran thousands of agentic AI workload experiments—extending the open-source Terminal-Bench harness with profiling, telemetry, and deterministic record-replay of LLM responses—to isolate agent system performance from LLM variability across tasks like compilation, database operations, ray tracing, and ML training. The results reframe enterprise agentic AI as a systems-engineering problem rather than an inference problem, yielding five practical guidelines: size capacity by agents-per-vCPU density rather than raw agent count; monitor P95 task latency rather than average CPU utilization, since bursty agent compute patterns mask queuing delays; and default to scale-out infrastructure (favoring scale-up only for heavier per-agent compute or architectural constraints). The piece proposes six operational metrics—task success rate, cost per task, time per task, throughput, agent density, and latency—as a more actionable evaluation frame than model-centric benchmarks. It concludes that near-term enterprise value concentrates in workflows with existing codified rules and measurable SLAs (code generation, test farms, ticket triage, market analysis, security review), making the target adopter an operations-accountable leader rather than an experimental end user. Note: this is sponsored content produced by Intel, not MIT Technology Review's editorial staff.