Literature review is not going away as a component of target discovery. The published record contains decades of experimental findings that are not captured in any structured database, and a researcher who can navigate that record intelligently has access to evidence that is invisible to purely database-driven approaches. The question is not whether to do literature review but how much of it should remain manually curated versus computationally processed.
What Manual Literature Review Actually Does Well
Manual review by a domain expert has genuine, hard-to-replicate strengths. An experienced computational biologist reading a paper can assess experimental quality in ways that text mining cannot: whether the controls are appropriate, whether the cell model is relevant to the disease context, whether the effect size justifies the conclusions, whether a published finding has been independently replicated or represents a single laboratory's result. These judgments require understanding experimental design, knowing which assay artifacts are common in a given field, and recognizing when authors are overclaiming.
Manual review is also better at recognizing mechanistic context. A finding that a particular gene is upregulated in a disease tissue is more meaningful if a reviewer knows that the upregulation has been shown to be a secondary consequence of inflammation rather than a primary driver. That contextual knowledge is difficult to encode in a text mining system because it requires understanding the semantic structure of the claim and its relationship to other published claims, not just extracting the co-occurrence of entities.
Where Manual Review Falls Apart at Scale
The limitations of manual review become critical as the evidence base for a target expands. A target with 20 relevant papers can be reviewed carefully in a day. A target with 200 relevant papers requires a week of serious effort. A comprehensive review of the evidence for 50 candidate targets -- a realistic early-stage target list before prioritization -- would take months and is practically impossible for most research teams with existing workloads.
The scale problem is compounded by two systematic biases. First, reviewers under time pressure tend to default to recent, high-impact papers and underweight older findings that may be more carefully controlled. Second, reviewers have prior beliefs about which targets are likely to be valid, and those priors influence which evidence they attend to and how they interpret ambiguous results. These are not failings unique to less experienced reviewers; they are documented in cognitive science research on expert judgment under uncertainty and show up consistently across scientific domains.
What Automated Mining Does Well
Automated literature mining addresses the scale problem directly. A system that has indexed 40 million records and extracted entity relationships across that corpus can return all evidence connecting a target gene to a disease entity in seconds, across the full time span of the indexed literature. It will not weight recent papers more than older ones by default, it will not miss a key finding in a 2015 paper because the reviewer ran out of time, and it will apply consistent extraction rules across all records.
Automated systems are also well-suited to specific sub-tasks within literature review: entity extraction (identifying all mentions of a gene, protein, or disease across a large corpus), co-occurrence counting, citation network analysis, and recency-weighted association scoring. These tasks are tedious for human reviewers but tractable for text processing pipelines at scale.
The Genuine Limitations of Automated Mining
Automated literature mining has real limitations that are worth being explicit about rather than glossing over. Negation handling remains a significant challenge: a sentence like "knockdown of gene X had no effect on cell viability" contains the same entity co-occurrences as "knockdown of gene X significantly reduced cell viability" but has the opposite evidentiary implication. State-of-the-art natural language processing handles negation better than earlier systems, but it still fails on complex constructions and on domain-specific negation patterns that require understanding biological context.
Experimental quality assessment is largely out of scope for current mining systems. A paper with poor controls that reports a strong positive finding will contribute the same co-occurrence signal to an automated mining system as a well-controlled study from a rigorous group. The system has no way to distinguish them without access to structured metadata about experimental design quality that simply does not exist at scale.
The Right Integration
The practical answer is a staged workflow. Automated mining handles the breadth: indexing the full literature corpus, extracting entity associations, generating candidate evidence sets for a large initial target list. Manual review handles the depth: once a target list has been narrowed to a manageable number of high-confidence candidates, expert curation of the key papers provides the quality assessment and mechanistic context that automated systems cannot supply.
This is not a novel conclusion, but it is one that many groups implement poorly by doing one stage well and skipping the other. Teams that do only automated mining sometimes act on spurious associations that a reviewer would have immediately recognized as methodologically unreliable. Teams that do only manual review miss evidence that is outside the reviewer's specific expertise or publication recency window.
The Indexing and Currency Problem
Any literature mining system is only as current as its index. The lag between paper publication and database indexing varies by source: PubMed is typically current within days, but preprint servers and conference proceedings may be indexed inconsistently, and the specific relationship extraction pipelines that transform indexed papers into structured entity associations have their own processing latency. A target whose key supporting evidence was published as a bioRxiv preprint three months ago may not yet be fully represented in a mining system's structured evidence layer even if the paper text is available.
This currency gap is worth factoring into how much weight you place on automated mining for fast-moving research areas. In established fields where the evidence base is stable and well-indexed, automated mining is highly reliable. In areas where recent preprints are driving rapid development of target hypotheses, automated mining may need to be supplemented with manual tracking of the most relevant preprint servers. The right balance depends on the specific disease area and how quickly the evidence landscape is changing.