LLMs reshape drug synergy and clinical data

Good Morning!

Eight papers, zero hype — just the biomedical agent work that moved the needle in the last 24 hours.

1. LLM meets single-cell RNA via verifiable RL

Li et al. bridged single-cell transcriptomics and large language models by discretizing gene-expression profiles into tokens using residual vector quantization, then fine-tuning with verifiable reinforcement learning so the model's reasoning steps can be checked against ground-truth cell-type labels. The approach sidesteps the continuous-to-token mismatch that has blocked prior scRNA-LLM integrations. Tested across multiple atlas datasets, RVQ-Alpha outperforms contrastive baselines on zero-shot cell annotation — a task that sits at the front of nearly every single-cell agent pipeline. Read More →

Why it matters: Verifiable RL gives single-cell agents a built-in audit trail for annotation decisions, which matters the moment those annotations feed downstream therapeutic target calls rather than just papers.

2. Graph nets predict combination drug efficacy

Song et al. applied residual graph isomorphism networks with attention mechanisms to drug-synergy prediction, encoding molecular graphs and cell-line features jointly to score pairwise drug combinations. The model outperforms prior GNN baselines on the DrugComb benchmark across multiple synergy metrics. For agent pipelines that triage combination candidates before wet-lab validation, this type of graph-attention scorer is the kind of module that slots directly into a hit-expansion loop. Read More →

3. Cross-attention scores RNA–protein binding generalizably

Catalano et al. trained a cross-attention model over paired RNA and protein sequences to predict interaction probability without relying on structure, achieving strong generalization to unseen RNA families — a persistent weak point for structure-dependent methods. The architecture is lightweight enough to score large interaction databases in batch, making it a practical candidate for agent-driven target-identification workflows. Read More →

4. FHIR format choice shifts LLM accuracy

Pator et al. showed that how FHIR records are serialized — JSON, XML, or simplified text — meaningfully changes LLM medication-reconciliation accuracy, with structured plain-text formats outperforming raw JSON on most models tested. Agents reading EHR data need a serialization standard, and this paper quantifies the cost of picking the wrong one. Read More →

5. Missing clinical data encoded as signal, not noise

Liang et al. modeled informative missingness in multimodal clinical time-series — treating the pattern of absent measurements as a feature rather than a gap — and learned dynamic treatment policies from the combined signal, improving outcomes on ICU decision benchmarks. Read More →

6. Differential privacy benchmarked on Dutch EHR notes

Miranda et al. compared DP de-identification methods on Dutch clinical text, finding that utility drops sharply at privacy budgets tight enough to satisfy GDPR — a concrete trade-off number for teams building agent pipelines over European EHR data. Read More →

7. ProDock pipes multi-target docking to storage

Phan et al. released ProDock, a workflow that runs consensus docking across multiple targets and writes results directly to a structured database, cutting the manual export step between docking runs and downstream analysis agents. Read More →

8. Agentic workflow classifies rock formations

Zhou et al. described GeoMind, an agentic workflow for lithology classification that invokes geological reasoning tools selectively — adjacent to biomedical discovery only in its agent architecture, included here for the tool-invocation design pattern. Read More →

Reply to talk back — this email comes to a human (newsletter@heurekalabs.co). Forward freely.

Agentic Discovery is a project of Heureka Labs · Unsubscribe