{"id":38963,"date":"2026-07-14T09:11:00","date_gmt":"2026-07-14T13:11:00","guid":{"rendered":"https:\/\/www.drugpatentwatch.com\/blog\/?p=38963"},"modified":"2026-05-20T11:19:21","modified_gmt":"2026-05-20T15:19:21","slug":"sentiment-analysis-in-patent-litigation-how-to-predict-pharmaceutical-case-outcomes","status":"publish","type":"post","link":"https:\/\/www.drugpatentwatch.com\/blog\/sentiment-analysis-in-patent-litigation-how-to-predict-pharmaceutical-case-outcomes\/","title":{"rendered":"Sentiment Analysis in Patent Litigation: How to Predict Pharmaceutical Case Outcomes"},"content":{"rendered":"\n
\"\"<\/figure>\n\n\n\n

Pharmaceutical patent litigation produces more text per dollar at stake than almost any other legal domain. A single Hatch-Waxman case between AstraZeneca and a generic challenger can generate thousands of pages: the complaint, the answer, expert declarations, Markman briefs, the claim construction order, motions for summary judgment, expert reports, trial transcripts, and the final opinion. Add the parallel inter partes review at the Patent Trial and Appeal Board, and the paper trail doubles.<\/p>\n\n\n\n

All of that text contains signal. The language a judge uses when asking questions from the bench, the specific terms an Administrative Patent Judge reaches for when drafting an institution decision, the rhetorical posture a Federal Circuit panel adopts in an oral argument transcript \u2014 each of these carries probabilistic information about where the case is heading. For years, experienced patent litigators and IP analysts have parsed that language intuitively. Sentiment analysis and natural language processing make the process systematic.<\/p>\n\n\n\n

This article is a working guide to deploying sentiment analysis on pharmaceutical patent litigation documents. It covers the theoretical framework, the data sources, the models that perform best in legal text, the specific signals that predict outcomes at each litigation stage, and the commercial intelligence you can extract from the resulting scores. It is written for IP counsel, patent analysts, pharmaceutical business development teams, and portfolio managers who need to price litigation risk with precision, not intuition.<\/p>\n\n\n\n

\n\n\n\n

Why Pharmaceutical Patent Litigation Generates Unusually Rich NLP Data<\/strong><\/h2>\n\n\n\n

Before discussing methodology, it helps to understand why pharma is a particularly productive domain for this kind of analysis.<\/p>\n\n\n\n

The Volume and Consistency of ANDA Litigation Text<\/strong><\/h3>\n\n\n\n

Hatch-Waxman litigation under 35 U.S.C. \u00a7 271(e)(2) follows a rigidly structured procedural timeline. The 30-month stay after a Paragraph IV certification notification, the required jurisdiction in a handful of federal districts, and the bench-trial requirement all produce consistent, comparable bodies of text across hundreds of cases. That structural consistency is analytically valuable. When you train a model on claim construction orders from the District of Delaware between 2015 and 2024, the documents share enough procedural DNA that linguistic variation becomes meaningful signal rather than noise from format differences.<\/p>\n\n\n\n

The geographic concentration amplifies this. The overwhelming majority of ANDA complaints are filed in the District of Delaware and the District of New Jersey, the two jurisdictions where most pharmaceutical companies are incorporated or headquartered. You are, in practical terms, modeling a relatively small number of judges handling a high volume of structurally similar cases \u2014 a near-ideal NLP training environment.<\/p>\n\n\n\n

The Financial Stakes That Justify Investment in Prediction<\/strong><\/h3>\n\n\n\n

The economics justify sophisticated analysis. Between 2025 and 2030, an estimated $236 billion in global pharmaceutical revenue is at risk due to patent expirations. A single loss-of-exclusivity (LOE) event on a blockbuster can erase 80-90% of branded revenue within months of generic entry. Against that backdrop, spending $500,000 on an NLP-based litigation intelligence program that shifts outcome probability estimates by even 5 percentage points produces a return that dwarfs the investment.<\/p>\n\n\n\n

The Settlement Distortion That Makes Statistical Baselines Misleading<\/strong><\/h3>\n\n\n\n

Raw win\/loss statistics in ANDA litigation are almost useless as a predictive baseline. A 2024 analysis of terminated Hatch-Waxman cases found that innovator companies prevailed on the merits 20% of the time, whereas generic companies won a mere 2% of the time. These statistics, however, are profoundly misleading because they ignore the single most important factor in Paragraph IV litigation: settlement. The vast majority of cases resolve through negotiated agreements before a final merits ruling. In 2024, 39% of terminated matters ended in settlement, compared to 50% in 2023.<\/p>\n\n\n\n

What this means for sentiment analysis is that your predictive model needs to target two distinct outcomes: the probability of settlement (and its likely terms) and the probability of a specific merits ruling in the cases that go to decision. These require different feature sets and different text corpora. Conflating them produces garbage predictions.<\/p>\n\n\n\n

\n\n\n\n

What Is Sentiment Analysis in Legal Text? Definitions and Scope<\/strong><\/h2>\n\n\n\n

Standard Sentiment Analysis vs. Legal Sentiment Analysis: What Changes<\/strong><\/h3>\n\n\n\n

Sentiment analysis in its consumer application \u2014 classifying a product review as positive or negative \u2014 operates on a different linguistic register than legal opinion analysis. Consumer text is colloquial, subjective, and emotionally direct. Legal text is formal, performative, and deliberately hedged. A judge who describes a patent owner’s claim construction argument as ‘not without merit’ is communicating something very different from a consumer who says a product is ‘not bad.’<\/p>\n\n\n\n

In legal contexts, sentiment analysis expands to cover several distinct analytical tasks:<\/p>\n\n\n\n

    \n
  • Tone analysis<\/strong>: Is the judge’s language skeptical, deferential, or neutral toward a party’s argument?<\/li>\n\n\n\n
  • Argument strength scoring<\/strong>: Does the court’s opinion language signal that a particular legal theory was found persuasive or dismissed?<\/li>\n\n\n\n
  • Outcome prediction<\/strong>: Given the linguistic features of a brief, oral argument transcript, or procedural order, what is the probability of each possible outcome?<\/li>\n\n\n\n
  • Settlement signal extraction<\/strong>: Does the procedural posture and language of a case suggest the parties are moving toward negotiated resolution?<\/li>\n<\/ul>\n\n\n\n

    The Difference Between Sentiment Scoring and Predictive Modeling<\/strong><\/h3>\n\n\n\n

    These are related but distinct. Sentiment scoring assigns valence or intensity values to text \u2014 a Markman order that uses language associated with skepticism about the patent owner’s claim scope argument scores differently than one that uses language associated with deference. Predictive modeling takes those scores (and other features) as inputs into a classification or regression model that outputs a probability estimate for a specific outcome.<\/p>\n\n\n\n

    In practice, sentiment scores from judicial opinions are one signal among several. The best predictive models for pharmaceutical patent litigation combine sentiment features with structured metadata: the judge’s prior ruling history in ANDA cases, the technology center at the PTAB, the number of prior art references cited in an IPR petition, whether a Markman ruling was appealed, and the current stage of the litigation timeline.<\/p>\n\n\n\n

    Key NLP Terms Every Pharma IP Analyst Should Know<\/strong><\/h3>\n\n\n\n

    Token<\/strong>: The basic unit of text analysis, roughly equivalent to a word or subword unit after preprocessing.<\/p>\n\n\n\n

    TF-IDF (Term Frequency-Inverse Document Frequency)<\/strong>: A classical weighting scheme that scores how important a term is to a specific document relative to a corpus. Useful for identifying terms that are statistically distinctive in judicial opinions associated with favorable or unfavorable outcomes.<\/p>\n\n\n\n

    Word embeddings<\/strong>: Dense vector representations of words (Word2Vec, GloVe, FastText) that capture semantic relationships. ‘Invalid’ and ‘unpatentable’ will cluster together in a well-trained embedding space, which helps models generalize across variant terminology.<\/p>\n\n\n\n

    Transformer models (BERT, RoBERTa, LegalBERT)<\/strong>: Attention-based neural architectures pretrained on large corpora that produce contextual representations of text. A word’s representation changes based on the surrounding context, making these models significantly better at parsing the hedged, conditional language of judicial opinions than earlier approaches.<\/p>\n\n\n\n

    Fine-tuning<\/strong>: The process of taking a pretrained model and further training it on domain-specific data \u2014 in this case, pharmaceutical patent litigation opinions and PTAB decisions \u2014 to improve performance on in-domain tasks.<\/p>\n\n\n\n

    \n\n\n\n

    The Pharmaceutical Patent Litigation Pipeline: Where Sentiment Lives<\/strong><\/h2>\n\n\n\n

    Sentiment signals appear at multiple stages of pharmaceutical patent litigation. Understanding the pipeline is prerequisite to knowing which documents to analyze and what to look for.<\/p>\n\n\n\n

    Stage 1: Paragraph IV Notification and Initial Case Filing<\/strong><\/h3>\n\n\n\n

    The process starts with the generic or biosimilar manufacturer filing an Abbreviated New Drug Application (ANDA) with the FDA accompanied by a Paragraph IV certification asserting that listed Orange Book patents are invalid, unenforceable, or will not be infringed by the proposed generic product. The brand manufacturer receives a detailed notice letter explaining the legal and technical basis for the challenge.<\/p>\n\n\n\n

    These notice letters are analytically underutilized. They contain the challenger’s initial claim construction theory, its invalidity arguments, and its infringement non-infringement analysis. The specificity, confidence, and internal consistency of the legal theories in a notice letter \u2014 features measurable through NLP \u2014 correlate with subsequent litigation performance. A notice letter that relies on a single narrow prior art reference for its invalidity argument is structurally different from one that marshals multiple independent anticipation and obviousness grounds. The linguistic confidence of the argument, measured through hedging terms, modal verbs, and citation density, provides a rough first-pass quality signal before any court has opined on anything.<\/p>\n\n\n\n

    Stage 2: The 30-Month Stay and Early Litigation Filings<\/strong><\/h3>\n\n\n\n

    Brand manufacturers have 45 days after receiving the Paragraph IV notice to file a patent infringement suit. Filing suit triggers an automatic 30-month stay of FDA final ANDA approval. The stay gives the parties time to litigate. Early complaints, answers, counterclaims, and scheduling orders do not carry heavy predictive value by themselves, but they establish the initial framing: which patents are asserted, which invalidity grounds are raised, and which claim terms are identified as disputed.<\/p>\n\n\n\n

    Scheduling orders and local patent rules disclosures (infringement contentions, invalidity contentions) filed early in discovery are more useful. Invalidity contentions that cite large numbers of prior art references across multiple grounds (anticipation under 35 U.S.C. \u00a7 102, obviousness under \u00a7 103, written description and enablement under \u00a7 112) signal a better-resourced and more comprehensive challenge. NLP-based document classification can rank the strength of invalidity contentions relative to a training corpus of prior contentions in resolved cases.<\/p>\n\n\n\n

    Stage 3: The Markman Hearing and Claim Construction Order<\/strong><\/h3>\n\n\n\n

    The Markman hearing is the single most predictively loaded event in district court ANDA litigation. The Markman hearing is often the pivotal moment in a patent infringement case because the judge’s interpretation of claim language can determine liability and damages outcomes. A favorable claim construction can lead directly to a finding of infringement or non-infringement on summary judgment, narrow the issues for trial, or drive the parties toward settlement.<\/p>\n\n\n\n

    The claim construction order is the document with the highest density of predictive linguistic signal in the entire ANDA case docket. It contains the judge’s reasoning in full, including the arguments considered, the arguments rejected, the intrinsic and extrinsic evidence credited, and the specific definitions adopted. Each of those elements carries sentiment and argument-valence signals.<\/p>\n\n\n\n

    Stage 4: Summary Judgment and Trial<\/strong><\/h3>\n\n\n\n

    Cases that survive claim construction and reach summary judgment or trial produce opinions on validity and infringement. These opinions are the primary training data for outcome prediction models. District courts found patents invalid at trial 24% of the time in 2024, versus 60% of the time in 2023 \u2014 a dramatic shift driven partly by case selection effects and the types of patents reaching full trial.<\/p>\n\n\n\n

    Stage 5: The Parallel PTAB Track<\/strong><\/h3>\n\n\n\n

    Since the America Invents Act of 2011, pharmaceutical patent challengers have had a parallel route: filing an IPR petition at the PTAB seeking to cancel the Orange Book patent claims on prior art grounds. Since the AIA’s trial proceedings launched in September 2012, the PTAB has received more than 14,000 IPR petitions. Once a trial is instituted, approximately 70-80% of challenged claims in final written decisions are found unpatentable.<\/p>\n\n\n\n

    IPR proceedings generate distinct text corpora: the petition itself, the patent owner’s preliminary response, the institution decision, the patent owner’s response, the petitioner’s reply, expert declarations from both sides, oral argument transcripts, and the final written decision. Each document layer adds predictive information, and sentiment analysis at each layer refines the probability estimate for the final written decision outcome.<\/p>\n\n\n\n

    \n\n\n\n

    Building the Dataset: Which Documents to Collect and How<\/strong><\/h2>\n\n\n\n

    Primary Sources for Pharmaceutical Patent Litigation Text<\/strong><\/h3>\n\n\n\n

    The foundation of any NLP-based litigation prediction system is the training corpus. For pharmaceutical patent litigation, the primary document sources are:<\/p>\n\n\n\n

    PACER (Public Access to Court Electronic Records)<\/strong>: The federal judiciary’s public filing system. Contains all court filings in district court ANDA cases: complaints, motions, briefs, expert declarations, and judicial opinions. Access requires a fee-based account but bulk download is feasible.<\/p>\n\n\n\n

    USPTO Patent Trial and Appeal Board e-FOIA<\/strong>: The PTAB’s public portal contains all petitions, preliminary responses, institution decisions, trial briefs, and final written decisions from IPR and PGR proceedings. The USPTO makes structured statistics available that can be used to validate models, and the full text of all public documents is accessible.<\/p>\n\n\n\n

    Google Scholar \/ Westlaw \/ Lexis+<\/strong>: Published judicial opinions with citations and headnotes. Useful for Federal Circuit appeal opinions which have broad precedential impact.<\/p>\n\n\n\n

    DrugPatentWatch<\/strong>: A commercial pharmaceutical patent intelligence platform that aggregates Orange Book data, patent expiration timelines, ANDA filing activity, Paragraph IV certification history, and litigation case tracking. DrugPatentWatch is particularly useful for linking litigation events to specific drugs, active ingredients, and companies, which allows your NLP analysis to be stratified by therapeutic area, patent type (composition-of-matter vs. method-of-use vs. formulation), and challenger identity. The platform’s structured metadata provides the backbone against which NLP sentiment scores can be calibrated and validated.<\/p>\n\n\n\n

    How to Structure the Training Corpus<\/strong><\/h3>\n\n\n\n

    A well-structured training corpus for pharmaceutical patent litigation prediction has three components: the document text, the document metadata, and the outcome label.<\/p>\n\n\n\n

    Document metadata should include: case number, filing date, district, judge name, parties (brand and generic), drug name, active ingredient, patent number, patent type (composition, formulation, method-of-use), Orange Book listing status, whether an IPR was filed in parallel, litigation stage at document creation, and ultimate case outcome.<\/p>\n\n\n\n

    Outcome labels depend on what you are predicting. For Markman orders, the outcome might be coded as ‘claim construction favorable to patent owner,’ ‘claim construction favorable to challenger,’ or ‘split.’ For final trial opinions, the label is validity and infringement findings for each asserted patent claim. For IPR final written decisions, the label is claim-level patentability determination.<\/p>\n\n\n\n

    The most common labeling error in legal NLP is treating the party that ‘won’ the case as determinative without accounting for which specific claims survived. A brand manufacturer whose patent was found valid on infringement but invalid on obviousness ‘won’ on one issue and ‘lost’ on another. Models trained on case-level outcome labels without claim-level granularity perform worse and generalize less well.<\/p>\n\n\n\n

    Handling the Data Imbalance Problem in Patent Litigation<\/strong><\/h3>\n\n\n\n

    ANDA cases settle at high rates \u2014 around 39-50% of terminated matters in recent years. That means the subset of cases with judicial decisions on the merits is smaller than the full case population. Among decided cases, the outcomes are not balanced: innovator companies prevailed on issues 20% of the time versus generic companies at 2% of the time for the 2024 cohort of terminated Hatch-Waxman cases, reflecting the extreme dominance of settlement as an outcome.<\/p>\n\n\n\n

    Standard approaches to class imbalance \u2014 oversampling minority classes (SMOTE), undersampling majority classes, or adjusting class weights in the loss function \u2014 apply here. For pharmaceutical applications, the stakes asymmetry also matters: a false negative (predicting a patent will survive when it will be invalidated) and a false positive (predicting invalidation when the patent survives) have very different commercial consequences. Adjust your decision threshold accordingly.<\/p>\n\n\n\n

    \n\n\n\n

    NLP Models for Legal Text: What Works in Pharmaceutical Patent Cases<\/strong><\/h2>\n\n\n\n

    Why Standard Sentiment Models Fail on Patent Opinions<\/strong><\/h3>\n\n\n\n

    VADER, TextBlob, and other lexicon-based sentiment tools are calibrated on social media text, product reviews, and news articles. They assign sentiment scores based on word-level polarity dictionaries that do not transfer to legal language. The word ‘invalid’ in a tweet means something different than ‘invalid’ in a PTAB final written decision. Generic tools systematically misclassify the hedged language of judicial opinions, where high-stakes conclusions appear in passive constructions, subordinate clauses, and conditional framings that share surface features with neutral or even positive language.<\/p>\n\n\n\n

    The failure mode is specific to pharmaceutical patent opinions. A judge writing that a patent owner’s argument is ‘not without force’ is expressing limited but real credence \u2014 a trained legal reader assigns moderate-to-positive sentiment, while a lexicon-based tool may score the phrase negatively because of the word ‘not.’ A PTAB administrative patent judge who writes ‘petitioner has not demonstrated by a preponderance of the evidence that claim 1 would have been obvious’ is ruling for the patent owner \u2014 but the sentence contains the word ‘obvious,’ which lexicon tools often flag as a negative signal for the patent.<\/p>\n\n\n\n

    TF-IDF Baselines: Still Useful for Feature Engineering<\/strong><\/h3>\n\n\n\n

    Before deploying transformer models, TF-IDF remains useful for identifying the vocabulary that distinguishes winning-outcome opinions from losing-outcome opinions in your training corpus. Run TF-IDF across your corpus of Markman orders labeled by outcome direction. The terms that appear with high frequency in patent-owner-favorable opinions versus challenger-favorable opinions give you a first-pass vocabulary of predictive language. These terms become features for simpler gradient-boosted classifiers (XGBoost, LightGBM) that are fast to train, interpretable, and effective when training data is limited.<\/p>\n\n\n\n

    Research predicting IPR institution outcomes has demonstrated that XGBoost models based on TF-IDF features from patent owner preliminary response briefs can produce useful predictive accuracy, and that these tree-based models have the advantage of interpretability \u2014 you can see which terms are driving the prediction, which in turn reveals what linguistic patterns the PTAB finds persuasive.<\/p>\n\n\n\n

    BERT and Its Variants: The Current Standard<\/strong><\/h3>\n\n\n\n

    Bidirectional Encoder Representations from Transformers (BERT) and its derivatives (RoBERTa, DeBERTa, ALBERT) have become the standard approach for legal text classification tasks. The key advantage is contextual encoding: a word’s vector representation is computed with respect to all surrounding words, allowing the model to correctly parse the hedged, negated, and conditionally qualified language that defines judicial opinions.<\/p>\n\n\n\n

    A systematic review of machine learning approaches to judicial decision prediction found that among deep learning techniques, LSTM networks and transformer models such as BERT are the most widely used and best-performing architectures, with prediction accuracy ranging from 81% to over 90% in well-structured legal domains.<\/p>\n\n\n\n

    For pharmaceutical patent litigation specifically, you have two fine-tuning options:<\/p>\n\n\n\n

    Domain-general legal BERT variants<\/strong>: LegalBERT (trained on U.S. Court opinions, contracts, and other legal text from the Pile of Law dataset) and CaseLaw-BERT (trained specifically on federal case law) both outperform base BERT on legal classification tasks without any task-specific fine-tuning.<\/p>\n\n\n\n

    Pharma-patent-specific fine-tuning<\/strong>: If you have a corpus of 500+ labeled ANDA opinions or PTAB final written decisions, fine-tuning LegalBERT on that corpus will outperform domain-general legal models on your specific prediction task. The general recommendation is: start with LegalBERT as your base, fine-tune on whatever in-domain labeled data you can assemble, and validate on held-out cases with known outcomes.<\/p>\n\n\n\n

    The Long Document Problem and Hierarchical Approaches<\/strong><\/h3>\n\n\n\n

    BERT’s maximum input length is 512 tokens \u2014 roughly 380 words. A Markman order in a complex ANDA case routinely runs 50-100 pages. A final written decision from the PTAB in a pharmaceutical IPR can exceed 80 pages. Standard BERT cannot process these documents in a single forward pass.<\/p>\n\n\n\n

    Three solutions exist. First, truncation: take the first 512 tokens or the last 512 tokens. This works poorly on judicial opinions where the most predictive language (the court’s legal conclusions) typically appears in the middle or end of the document. Second, windowed approaches: split the document into overlapping 512-token windows, run the model on each window, and aggregate the per-window predictions. Third, hierarchical models: encode individual sentences or paragraphs with BERT, then run a second-level model over the resulting sequence of paragraph-level representations. The hierarchical approach preserves document structure and outperforms windowing on long legal documents in most benchmarks.<\/p>\n\n\n\n

    For PTAB institution decisions specifically, a practical shortcut exists: these decisions typically contain a standard-format ‘Conclusion’ section that summarizes the board’s determination in 2-5 pages of dense, high-signal text. Targeting that section specifically for sentiment scoring, rather than the full document, can match or exceed whole-document model performance with a fraction of the computational cost.<\/p>\n\n\n\n

    \n\n\n\n

    Sentiment Signals at the Markman Hearing: Reading the Claim Construction Tea Leaves<\/strong><\/h2>\n\n\n\n

    Why the Markman Order Is the Most Predictive Document in ANDA Litigation<\/strong><\/h3>\n\n\n\n

    The claim construction order deserves special treatment. A favorable claim construction can lead directly to a finding of infringement or non-infringement on summary judgment, narrow the issues for trial, or drive the parties toward settlement. In practice, experienced ANDA litigators know whether their case improved or deteriorated the moment the Markman order drops \u2014 often before they have read it in full, just from the opening paragraphs. That intuition is precisely what NLP can systematize.<\/p>\n\n\n\n

    Linguistic Markers of Patent-Owner-Favorable Claim Construction<\/strong><\/h3>\n\n\n\n

    Claim construction orders that favor the patent owner tend to display several identifiable linguistic patterns.<\/p>\n\n\n\n

    The court credits the patent owner’s argument about the plain and ordinary meaning of disputed terms without importing limitations from the specification. Language like ‘we find no basis to depart from the plain meaning,’ ‘the intrinsic record does not compel the narrow construction proposed by defendant,’ or ‘we agree that the claims should be construed as the patentee drafted them’ are characteristic of patent-owner-favorable constructions.<\/p>\n\n\n\n

    Courts favorable to patent owners tend to quote the patent specification at length to demonstrate consistency between the claims and the patent’s written description \u2014 a rhetorical move that validates the patent’s internal coherence. The density of direct specification quotation in the court’s reasoning (measurable through NLP) correlates positively with patent-owner-favorable outcomes.<\/p>\n\n\n\n

    Expert testimony crediting. When the court’s opinion cites and accepts the patent owner’s expert witness at multiple points in its analysis, the sentiment trajectory of the document is favorable to the owner. The NLP task here is named entity recognition (identifying when the court refers to each party’s expert) combined with sentiment scoring of the surrounding context.<\/p>\n\n\n\n

    Linguistic Markers of Challenger-Favorable Claim Construction<\/strong><\/h3>\n\n\n\n

    Challenger-favorable Markman orders share their own vocabulary. The court typically imports limitations from the specification into the claims (‘we construe the term in light of the inventor’s own definition’), applies prosecution history estoppel (‘the patentee clearly disavowed this scope during prosecution’), or finds certain claim terms indefinite under 35 U.S.C. \u00a7 112.<\/p>\n\n\n\n

    The rhetorical tell of a challenger-favorable construction is a court that emphasizes what the patent does not cover: ‘the claims, properly construed, do not encompass the accused product’s formulation,’ ‘the patentee’s proposed construction would render the claim scope impermissibly broad,’ or ‘adopting the patent owner’s construction would read out preferred embodiments from the claim.’ Phrases anchored to limiting language \u2014 ‘does not,’ ‘cannot,’ ‘excluded,’ ‘disclaimed’ \u2014 cluster in negative-outcome opinions at higher frequency.<\/p>\n\n\n\n

    How Oral Argument Transcripts Add Predictive Value Before the Written Order<\/strong><\/h3>\n\n\n\n

    Markman hearings are bench proceedings and transcripts are typically filed on PACER within days of the hearing \u2014 often weeks before the written claim construction order issues. Those transcripts carry predictive signal about the eventual order. Judicial questioning patterns during Markman hearings are not neutral. When a judge asks multiple probing questions of one side’s argument while nodding through the other’s, the asymmetry of scrutiny is observable in the transcript.<\/p>\n\n\n\n

    Quantifiable features in oral argument transcripts: word count of judge’s questions directed to each party (more words per question to one party often signals skepticism); use of hypotheticals (judges typically introduce hypotheticals to stress-test arguments they are unconvinced by); latency of intervention (how quickly the judge interrupts a party’s argument, correlating with prior concern about that argument’s validity).<\/p>\n\n\n\n

    NLP can analyze how a specific technical or legal term has been used across thousands of other patents and court decisions to help build stronger arguments for a favorable claim construction and even predict how a particular judge might rule on a specific term. The oral argument transcript extends this analysis to real-time judicial language, letting you score the hearing outcome before the written order is drafted.<\/p>\n\n\n\n

    \n\n\n\n

    Predicting IPR Outcomes at the PTAB: NLP on Institution Decisions and Final Written Decisions<\/strong><\/h2>\n\n\n\n

    The Two-Stage PTAB Model: Institution vs. Final Written Decision<\/strong><\/h3>\n\n\n\n

    IPR prediction is not a single model problem. Institution and the final written decision are distinct decisions with distinct predictors, and conflating them produces a worse model than treating them separately.<\/p>\n\n\n\n

    The institution decision asks: has the petitioner shown a reasonable likelihood of prevailing on at least one challenged claim? In FY24, out of 1,288 petitions filed, 74% of patents were instituted \u2014 a figure that has remained relatively stable, with bio\/pharma petitions achieving a 73% institution rate. The base rate is high enough that institution prediction is primarily about identifying the cases that will be denied, which is the commercially more interesting question for patent owners assessing vulnerability.<\/p>\n\n\n\n

    The final written decision asks: has the petitioner proven by a preponderance of the evidence that the challenged claims are unpatentable? Nearly 80% of challenged claims are found invalid when the PTAB issues a final written decision. If disclaimed claims are factored in, more than 80% of challenged claims were invalidated in 2024. Here the base rate is high, and the commercially interesting question is identifying the patents that will survive \u2014 the ones worth the full defense investment.<\/p>\n\n\n\n

    What the IPR Petition Text Predicts About Institution<\/strong><\/h3>\n\n\n\n

    The IPR petition is the foundational document for institution prediction. Research using NLP on patent owner preliminary responses has demonstrated proof-of-concept predictive accuracy in IPR institution models. Working backward, the same logic applies to petition text.<\/p>\n\n\n\n

    High-institution-probability petitions share several NLP-measurable features:<\/p>\n\n\n\n

    Prior art citation density: petitions that cite multiple independent prior art references for each challenged claim, rather than relying on a single anticipatory reference, give the panel multiple paths to institution. NLP citation parsing can count and categorize prior art references automatically.<\/p>\n\n\n\n

    Claim chart specificity: well-structured petitions map each element of each challenged claim to specific passages in the prior art. Claim charts with granular element-by-element mapping, measurable through structural analysis of petition text, correlate with institution.<\/p>\n\n\n\n

    Argument type distribution: petitions built primarily on obviousness combinations under \u00a7 103 with multiple supporting references tend to perform better at institution than petitions relying heavily on single-reference anticipation arguments, which are facially simpler but easier for patent owners to distinguish. NLP classification of argument type within petitions is feasible with a relatively small labeled corpus.<\/p>\n\n\n\n

    Reading the Patent Owner Preliminary Response for Settlement Signals<\/strong><\/h3>\n\n\n\n

    The patent owner’s preliminary response (POPR) is an optional filing that allows the patent owner to argue against institution before the PTAB decides whether to proceed. It is analytically revealing not only because it previews the patent owner’s defense theory, but because its tone and length signal the patent owner’s internal assessment of vulnerability.<\/p>\n\n\n\n

    A POPR that focuses primarily on procedural arguments (real-party-in-interest challenges, time-bar arguments, discretionary denial grounds) rather than substantive invalidity rebuttal often indicates the patent owner lacks strong substantive defenses. A POPR that makes extensive substantive arguments, cites multiple claim differentiation principles, and attaches detailed expert declarations indicates the patent owner believes the merits support survival.<\/p>\n\n\n\n

    The ratio of procedural to substantive argument text in a POPR, measurable through NLP topic modeling and section classification, is a useful institutional settlement predictor: high procedural focus correlates with higher settlement probability before the final written decision, as the patent owner’s negotiating position is weaker on the merits.<\/p>\n\n\n\n

    Final Written Decision Language and Post-Decision Appellate Risk<\/strong><\/h3>\n\n\n\n

    Final written decisions from the PTAB that invalidate patent claims use specific linguistic constructions that signal the strength (or vulnerability) of the invalidation on Federal Circuit appeal. A decision that extensively addresses the patent owner’s rebuttal arguments, explicitly distinguishes the challenged claims from the asserted prior art’s differences, and acknowledges factual disputes that it resolves in the petitioner’s favor is more likely to withstand appeal than one that summarily dispatches the patent owner’s positions.<\/p>\n\n\n\n

    The Teva v. Sandoz standard (2015) requires the Federal Circuit to review the PTAB’s underlying factual findings for clear error rather than de novo, which means the linguistic texture of the final written decision affects appellate durability. While the ultimate determination of claim meaning remains a question of law reviewed de novo, underlying factual findings are subject to more deferential clear error review. Sentiment analysis of the final written decision’s treatment of disputed facts \u2014 did the panel engage carefully or summarily? \u2014 correlates with subsequent Federal Circuit affirmance rates.<\/p>\n\n\n\n

    \n\n\n\n

    Judge-Specific Sentiment Modeling: Building Judicial Profiles<\/strong><\/h2>\n\n\n\n

    Why Individual Judge History Is the Most Powerful Predictor<\/strong><\/h3>\n\n\n\n

    The single most powerful structured predictor of ANDA case outcome is the presiding judge’s prior ruling history in pharmaceutical patent cases. This is not controversial \u2014 experienced ANDA litigators file where they want to draw favorable judges, which is why case filings are concentrated in Delaware and New Jersey. What NLP adds is the ability to go beyond ‘judge X has ruled for patent owners in 7 of 10 cases’ to ‘judge X’s opinions display these specific linguistic patterns when ruling for challengers versus owners, and here is how the current Markman briefing maps to those patterns.’<\/p>\n\n\n\n

    Building judge-specific sentiment profiles requires: collecting all prior ANDA opinions by the target judge, labeling each by outcome direction at the claim level, fine-tuning a model on the judge’s historical output, and scoring new opinions and orders from that judge against the historical baseline.<\/p>\n\n\n\n

    Geographic and Jurisdictional Effects on Outcome Language<\/strong><\/h3>\n\n\n\n

    District-level patterns matter independent of individual judges. The District of Delaware and the District of New Jersey are hubs for ANDA litigation because most pharmaceutical companies are incorporated in Delaware and headquartered in New Jersey. Both districts have developed specialized local patent rules that produce consistent procedural document formats \u2014 useful for NLP preprocessing.<\/p>\n\n\n\n

    Delaware, with its concentration of IP-specialized judges, tends to produce Markman orders with more extensive technical analysis and longer, more detailed reasoning. New Jersey opinions vary more by judge. Federal Circuit appeal outcomes for ANDA cases originating from Delaware versus New Jersey differ systematically \u2014 a factor to incorporate into your hierarchical model.<\/p>\n\n\n\n

    PTAB Administrative Patent Judge (APJ) Panel Analysis<\/strong><\/h3>\n\n\n\n

    At the PTAB, the three-APJ panel composition for each IPR has historically been a meaningful predictor. The PTAB’s Art Unit 1600 (biotechnology and organic chemistry, covering most pharmaceutical patents) has a defined set of APJs with varying track records on specific invalidity theories. An APJ who has written multiple final written decisions crediting secondary indicia of nonobviousness (commercial success, long-felt need, unexpected results) will read an IPR petition challenging a pharmaceutical patent on obviousness grounds differently than one who has systematically discounted those arguments.<\/p>\n\n\n\n

    NLP-based APJ profiling \u2014 extracting each judge’s language patterns around key legal standards from their prior opinions \u2014 allows you to model not just whether the PTAB will institute but which obviousness or novelty framing is most likely to succeed before the specific panel assigned to your case. This is the level of analysis that sophisticated PTAB practitioners already perform qualitatively. NLP makes it quantitative and scalable.<\/p>\n\n\n\n

    \n\n\n\n

    Settlement Prediction: Using Litigation Language to Forecast Case Resolution<\/strong><\/h2>\n\n\n\n

    The Settlement Signal in Procedural Filings<\/strong><\/h3>\n\n\n\n

    Settlement in ANDA litigation is not random. Cases settle when the parties’ expected value calculations converge \u2014 when the brand manufacturer’s expected loss from continued litigation (patent invalidation, immediate generic entry, revenue cliff) approaches the generic manufacturer’s expected gain from settlement (authorized generic arrangement, defined entry date, elimination of litigation cost and risk). NLP can estimate when that convergence is approaching.<\/p>\n\n\n\n

    Procedural signals that correlate with impending settlement:<\/p>\n\n\n\n

    Joint motions for case schedule extensions: when parties jointly request time extensions, particularly after a Markman order that partially favored the challenger, the extension often reflects settlement discussions. The language of the joint motion (how the parties characterize the reasons for the extension, whether they mention the case’s complexity versus simply requesting administrative time) carries signal.<\/p>\n\n\n\n

    Reduced filing volume: a case that had weekly discovery disputes goes quiet. NLP monitoring of docket activity frequency is a simple but effective settlement predictor.<\/p>\n\n\n\n

    Changes in brief tone: post-Markman briefs from the brand manufacturer that begin acknowledging the challenger’s infringement non-infringement arguments rather than flat-out dismissing them sometimes signal a softening negotiating posture.<\/p>\n\n\n\n

    What Happens to Settlement Timing After an Adverse Markman<\/strong><\/h3>\n\n\n\n

    An adverse Markman ruling (for either party) dramatically changes the settlement calculus. Cases in which the patent owner received a narrow claim construction that excludes the generic product from its scope \u2014 or in which the challenger received a broad construction that captures the generic product \u2014 typically either settle within 60-90 days of the order or proceed to aggressive summary judgment motion practice. Monitoring post-Markman filing patterns and the language of those filings gives you a two-to-three month settlement probability window that is highly actionable for portfolio risk management.<\/p>\n\n\n\n

    Pay-for-Delay and Antitrust Exposure: Sentiment Signals in Settlement Agreements<\/strong><\/h3>\n\n\n\n

    Paragraph IV settlements between brand and generic manufacturers are scrutinized by the FTC under the antitrust framework established in FTC v. Actavis (2013). Agreements that include value transfers from the brand to the generic beyond simple authorized generic arrangements or defined entry dates are subject to antitrust challenge. The Actavis ‘rule of reason’ analysis requires courts to evaluate the overall competitive effects of the settlement.<\/p>\n\n\n\n

    For NLP purposes, settlement agreements themselves (when publicly available through regulatory filings or antitrust proceedings) can be analyzed for linguistic markers of value transfer: how the settlement describes any supply arrangement, whether the patent owner’s acknowledgment of the generic’s non-infringement position is broad or narrow, and how the entry date provision is structured. These are commercially important signals because the antitrust risk attached to a settlement directly affects its practical value to the generic manufacturer.<\/p>\n\n\n\n

    \n\n\n\n

    Case Studies: Sentiment Analysis Applied to Real Pharmaceutical Litigation<\/strong><\/h2>\n\n\n\n

    Case Study 1: Reading the Markman Order in Teva Pharmaceuticals USA, Inc. v. Sandoz, Inc.<\/strong><\/h3>\n\n\n\n

    The Teva v. Sandoz dispute over Copaxone (glatiramer acetate, 20 mg\/mL) involved claim construction of the term ‘average molecular weight’ \u2014 a technically contested term where multiple scientific measurement methods produced different numerical values, and each method supported or undermined infringement.<\/p>\n\n\n\n

    The district court’s Markman order in this case is a useful training document because it was subsequently reversed on appeal to the Federal Circuit and then reviewed by the Supreme Court, producing a three-level record of judicial opinion on identical claim language. NLP analysis of the three-opinion sequence reveals how the same disputed term was characterized at each level: the district court’s careful deference to extrinsic expert testimony about scientific measurement convention, the Federal Circuit’s de novo skepticism, and the Supreme Court’s correction of the appellate standard.<\/p>\n\n\n\n

    Training a model on this multi-level sequence teaches it to distinguish the linguistic markers of ‘deferred to expert’ reasoning (associated with district court outcomes) from ‘de novo construction’ reasoning (associated with Federal Circuit correction). That distinction directly predicts appeal risk in claim construction \u2014 a commercially critical variable.<\/p>\n\n\n\n

    Case Study 2: IPR Sentiment Analysis in the Eliquis (Apixaban) Defense<\/strong><\/h3>\n\n\n\n

    Bristol-Myers Squibb and Pfizer’s Eliquis (apixaban) is one of the most commercially significant patent cases in recent pharmaceutical history. BMS reported over $13 billion in 2024 Eliquis revenue; a scenario in which generic entry occurs on schedule in 2026 produces an NPV impact in the range of $25-35 billion in lost future cash flows.<\/p>\n\n\n\n

    Generic manufacturers including Sigmapharm, Aurobindo, MSN Pharmaceuticals, and others filed Paragraph IV ANDAs, triggering the 30-month litigation stay. Courts upheld some apixaban patents and invalidated others, producing a complex litigation landscape that BMS and Pfizer have actively managed through both litigation and negotiated settlement discussions.<\/p>\n\n\n\n

    The Eliquis litigation record demonstrates the value of portfolio-level sentiment analysis rather than single-case analysis. With dozens of challengers, multiple patents asserted, parallel IPR petitions, and complex settlement dynamics, tracking the sentiment trajectory of individual case filings gives a more accurate NPV estimate for the exclusivity portfolio than any single-case prediction. Sentiment scores from the Eliquis Markman orders \u2014 which construed claim terms related to daily dosage, pharmaceutical composition, and crystalline form \u2014 would have provided early portfolio signals about which patents were most vulnerable to challenger strategies.<\/p>\n\n\n\n

    Case Study 3: Serial Litigation and NLP-Based Pattern Detection<\/strong><\/h3>\n\n\n\n

    A tactic documented in recent academic literature involves brand manufacturers filing successive lawsuits built on new but substantively similar patents after an initial case settles or is resolved. In the case of Astellas’s overactive bladder drug mirabegron (Myrbetriq), after an initial Hatch-Waxman case settled in 2020 with generic entry expected in 2024, Astellas pursued four additional lawsuits, each built on new but substantively indistinguishable patents. These tactics have delayed broad competition, leaving only two firms to launch in 2024 under the threat of massive damages.<\/p>\n\n\n\n

    NLP-based patent claim comparison \u2014 measuring semantic similarity between the claims of successive patents asserted in serial litigation \u2014 can identify when new litigation is built on claims that are substantively identical to previously adjudicated ones. High semantic similarity between new patent claims and previously invalidated or expired claims is a flag for both FTC antitrust scrutiny and accelerated invalidity defense. Tools like sentence-BERT applied to patent claim text can quantify this similarity with high precision.<\/p>\n\n\n\n

    \n\n\n\n

    Integrating Sentiment Scores with Quantitative Patent Valuation Models<\/strong><\/h2>\n\n\n\n

    The Standard DCF-Plus-Litigation-Probability Framework<\/strong><\/h3>\n\n\n\n

    Standard pharmaceutical patent valuation uses a discounted cash flow model with the patent expiry date as the terminal exclusivity date, adjusted by probability-weighted scenario trees for each litigation or PTAB challenge. The PTAB introduces a probability-weighted scenario tree that bifurcates the expected value of the patent between two paths: the ‘PTAB challenge absent’ path, in which the patent proceeds to its nominal expiry, and the ‘PTAB challenge present’ path, in which the patent faces institution and a final written decision before its nominal expiry.<\/p>\n\n\n\n

    Sentiment analysis contributes to this framework by dynamically updating the probability assigned to each scenario as new documents become available. Before an IPR is filed, the patent’s valuation relies on structural patent quality metrics. After the petition is filed but before institution, sentiment analysis of the petition text updates the institution probability. After institution, sentiment analysis of the board’s first substantive order updates the final written decision probability. The result is a continuously updated probability estimate rather than a static expert-opinion-based probability that may not be revised for months.<\/p>\n\n\n\n

    Translating Sentiment Scores into LOE Date Adjustments<\/strong><\/h3>\n\n\n\n

    The most commercially actionable output of pharmaceutical patent litigation sentiment analysis is an adjusted probability distribution over LOE (loss of exclusivity) dates. Rather than a point estimate (‘this patent expires in 2028’), you produce a probability density over LOE dates that incorporates: nominal expiry, probability of IPR invalidation (and the expected IPR timeline), probability of district court validity\/infringement findings, probability of settlement and its likely entry date terms.<\/p>\n\n\n\n

    The LOE date distribution directly feeds into revenue forecasting models. A brand manufacturer’s finance team can update its long-range plan as new litigation documents shift the probability mass of the LOE distribution. An institutional equity analyst covering the company can monitor the same signals to identify when the market’s embedded LOE date assumption diverges from the updated prediction.<\/p>\n\n\n\n

    What This Means for Generic Manufacturer Go\/No-Go Decisions<\/strong><\/h3>\n\n\n\n

    From the generic manufacturer’s perspective, the commercial application of sentiment analysis inverts: you want to know when a brand manufacturer’s patent portfolio is more vulnerable than the market believes. A Markman order that produces a sentiment score indicating a narrow claim construction favorable to challengers, before the market has priced that outcome into the brand manufacturer’s stock, represents an information advantage. Generic manufacturers and pharmaceutical-focused hedge funds that build NLP monitoring infrastructure gain a systematic edge in timing ANDA filings, IPR petitions, and at-risk launch decisions.<\/p>\n\n\n\n

    \n

    ‘Patent data is the exhaust of pharmaceutical strategy. Every filing, every ruling, every prosecution history amendment is a record of what companies believe their innovations are worth and which claims they can actually defend. The question is whether you have the tools to read that exhaust accurately and quickly.’ \u2014 DrugPatentWatch analysis on pharmaceutical patent intelligence, 2024.<\/p>\n<\/blockquote>\n\n\n\n

    \n\n\n\n

    Practical Implementation: How to Build a Pharma Patent Sentiment System<\/strong><\/h2>\n\n\n\n

    Step-by-Step Architecture for a Working System<\/strong><\/h3>\n\n\n\n

    A functional pharmaceutical patent litigation sentiment monitoring system requires five components working together.<\/p>\n\n\n\n

    1. Document ingestion pipeline<\/strong>: automated collection from PACER (using the CourtListener API or direct PACER bulk access), PTAB e-FOIA, and DrugPatentWatch for structured patent and litigation metadata. The pipeline should trigger on new filings for a defined watch list of drugs, patents, or companies.<\/p>\n\n\n\n

    2. Document classification and preprocessing<\/strong>: identifying document type (complaint, Markman order, institution decision, final written decision, etc.), extracting the relevant sections for analysis, and cleaning text for model input. For PTAB documents, the analysis section, discussion of claims, and conclusion section should be extracted separately and scored independently.<\/p>\n\n\n\n

    3. Sentiment and argument scoring<\/strong>: running fine-tuned LegalBERT (or a pharma-specific BERT variant) on the preprocessed sections. Output a vector of scores: overall patent-owner-sentiment, overall challenger-sentiment, argument-confidence scores for specific legal theories (obviousness, anticipation, written description, enablement), and flagging of high-significance phrases.<\/p>\n\n\n\n

    4. Metadata integration and model inference<\/strong>: combining NLP-derived scores with structured metadata (judge identity, PTAB panel composition, patent type, claim count, number of asserted prior art references) in a gradient-boosted classifier or logistic regression model trained on historical outcomes.<\/p>\n\n\n\n

    5. Alert and reporting layer<\/strong>: generating automated alerts when a new filing shifts a watched patent’s predicted outcome probability by more than a defined threshold, with a summary of the specific linguistic features driving the change. Integration with portfolio management systems or financial models should propagate updated LOE probabilities automatically.<\/p>\n\n\n\n

    Build vs. Buy: Commercial Legal NLP Platforms vs. Custom Systems<\/strong><\/h3>\n\n\n\n

    Several commercial platforms now offer legal NLP capabilities applicable to patent litigation: Cipher (acquired by Clarivate), PatSnap, IPlytics, and Docket Navigator all provide varying degrees of NLP-based patent analysis. None are pharmaceutical-litigation-specific out of the box. The choice between building a custom system and integrating a commercial platform depends on the size of the patent portfolio at risk, the frequency of litigation events, and the internal technical capacity of the organization.<\/p>\n\n\n\n

    For a brand manufacturer with a $10 billion+ revenue portfolio facing multiple active ANDA challenges, the economics of a custom system are compelling. For a generic manufacturer evaluating targets for ANDA filing, a commercial platform augmented with custom pharmaceutical sentiment models may be the more practical path. DrugPatentWatch’s structured litigation tracking, combined with your own NLP layer applied to the underlying filing documents, is a pragmatic hybrid that many mid-sized pharmaceutical companies can deploy without a large ML engineering team.<\/p>\n\n\n\n

    Validation: How to Know Your Model Is Working<\/strong><\/h3>\n\n\n\n

    Validation for legal prediction models requires careful design. Standard cross-validation on a random split of your labeled data overfits to time-specific judicial language patterns, because legal language evolves and specific judges’ writing styles change over their careers. Use temporal validation: train on all cases from years T-10 through T-2, validate on T-2 through T-1, and test on T-1 through T. This tests whether your model generalizes to new cases that occurred after training, which is the actual deployment scenario.<\/p>\n\n\n\n

    Calibration matters as much as discrimination. A model that says ‘70% probability of patent invalidation’ should be right about 70% of the time when it says that. Pharmaceutical patent litigation models trained on small datasets have a tendency to be overconfident. Isotonic regression or Platt scaling on the model’s output probabilities corrects for this and produces calibrated probability estimates that are suitable for direct input into financial valuation models.<\/p>\n\n\n\n

    \n\n\n\n

    Risks, Limitations, and Ethical Guardrails<\/strong><\/h2>\n\n\n\n

    When Sentiment Analysis Gets Pharmaceutical Patent Cases Wrong<\/strong><\/h3>\n\n\n\n

    Sentiment analysis models fail in pharmaceutical patent litigation in several predictable ways. Understanding these failure modes prevents over-reliance on model output.<\/p>\n\n\n\n

    First, unprecedented legal theories. The PTAB’s approach to biosimilar patent obviousness has shifted substantially as APJ panels grapple with the complexity of large-molecule pharmaceutical chemistry. A model trained on small-molecule ANDA opinions will not generalize well to biosimilar IPRs, where the technical arguments, the relevant prior art, and the legal standards interact differently. Always segment your training corpus by patent type and molecule class.<\/p>\n\n\n\n

    Second, regime changes. The Federal Circuit’s 2024 decisions on obviousness in pharmaceutical patent cases \u2014 particularly around secondary considerations of nonobviousness \u2014 shifted the legal landscape in ways that require retraining models that incorporated prior judicial language on those issues. NLP models are not self-updating; their predictions drift as the legal environment evolves. Scheduled retraining on recent decisions is required.<\/p>\n\n\n\n

    Third, judge assignment effects. When a highly experienced ANDA specialist judge in Delaware retires or is elevated to the Federal Circuit, cases reassigned to a less experienced judge produce outcomes that are harder to predict from prior court language patterns alone. Model performance typically degrades in the 12-18 months after significant judicial assignment changes.<\/p>\n\n\n\n

    Fourth, the FTC and regulatory intervention effect. The FTC’s recent campaign against improper Orange Book listings \u2014 challenging more than 100 patents in November 2023 and expanding to over 300 additional patents in April 2024 \u2014 creates a category of litigation events driven by regulatory pressure that does not have strong historical precedent. The FTC challenged patents as improperly listed, targeting asthma inhalers, epinephrine autoinjectors, and other drug-device combination products, warning that patents covering delivery devices should not be listed in the Orange Book if they do not claim the approved drug substance or product directly. Models trained entirely on pre-FTC-campaign data will not have seen this dynamic.<\/p>\n\n\n\n

    Ethical Use of Predictive Models in Patent Litigation<\/strong><\/h3>\n\n\n\n

    Predictive models do not create legal conclusions; they estimate probabilities. Using a sentiment-derived outcome probability as a substitute for legal analysis \u2014 rather than as one input into a broader legal and commercial judgment \u2014 produces liability risk. Model outputs should always be reviewed by qualified patent counsel before informing litigation strategy decisions.<\/p>\n\n\n\n

    There is also a data bias concern specific to pharmaceutical patent litigation. Cases that settle never produce labeled outcome data on the merits. The training corpus is therefore biased toward cases with unusual merit-level outcomes \u2014 the ones where settlement failed or was strategically avoided. Models trained on this subset will systematically misestimate probabilities in cases where settlement is the expected outcome for rational commercial reasons unrelated to legal merit.<\/p>\n\n\n\n

    Privilege and Confidentiality in NLP Systems<\/strong><\/h3>\n\n\n\n

    All documents analyzed in your NLP system must be either public (filed in court or at the PTAB and accessible without restriction) or covered by appropriate privilege and confidentiality controls if they include attorney-client communications or work product. The ingestion pipeline should have explicit document classification to separate public filings from privileged documents, with the latter excluded from training data and stored in isolated systems. This is not primarily a legal compliance issue (analyzing your own privileged documents for internal analysis is permissible) but a practical operational security issue: training data that includes privileged internal assessments should not be on the same system as training data from publicly filed documents.<\/p>\n\n\n\n

    \n\n\n\n

    The Biosimilar Dimension: Applying Sentiment Analysis to BPCIA Litigation<\/strong><\/h2>\n\n\n\n

    How BPCIA Patent Litigation Differs from Hatch-Waxman for NLP Purposes<\/strong><\/h3>\n\n\n\n

    The Biologics Price Competition and Innovation Act (BPCIA) creates a patent dance framework for biosimilar litigation that differs structurally from Hatch-Waxman. Where Hatch-Waxman triggers a single 30-month stay and routes disputes to a defined district court forum, the BPCIA involves a pre-litigation information exchange, mandatory patent lists from both parties, and the possibility of immediate patent infringement actions as well as declaratory judgment suits. The result is a more complex procedural record that requires different NLP preprocessing to extract predictive signals.<\/p>\n\n\n\n

    The most analytically rich documents in BPCIA disputes are the patent list exchanges themselves \u2014 the brand manufacturer’s initial patent list, the biosimilar developer’s responsive list, and the parties’ subsequent narrowing exchanges. NLP analysis of how these lists change through the exchange process (which patents are added, which are dropped, which are contested by both parties) reveals the parties’ internal legal assessments of patent strength before litigation begins.<\/p>\n\n\n\n

    Biosimilar Market Entry Timing: Sentiment Signals in Consent Judgment Language<\/strong><\/h3>\n\n\n\n

    BPCIA cases \u2014 particularly those involving blockbuster biologics like adalimumab (Humira), etanercept (Enbrel), and bevacizumab (Avastin) \u2014 have produced consent judgments and settlement agreements that define market entry dates in complex, contingent terms. The language of these agreements, when publicly disclosed through 8-K filings or antitrust discovery, contains valuable commercial signals about the parties’ assessments of their respective legal positions at the time of settlement.<\/p>\n\n\n\n

    A consent judgment that allows immediate biosimilar entry carries very different sentiment than one that requires the biosimilar to wait until near the reference product’s patent expiry. The distance between the settlement entry date and the nominal patent expiry date, combined with the language the parties use to characterize the patents (whether the brand acknowledges any invalidity risk or simply grants a license), is a direct indicator of the patent owner’s legal confidence.<\/p>\n\n\n\n

    \n\n\n\n

    The Inflation Reduction Act Variable: How New Pricing Mechanics Change Litigation Sentiment<\/strong><\/h2>\n\n\n\n

    IRA Drug Price Negotiation and Its Effect on Patent Defense Economics<\/strong><\/h3>\n\n\n\n

    The Inflation Reduction Act of 2022 changed the economic calculus for pharmaceutical patents in ways that are still propagating through the industry. The IRA’s Medicare drug price negotiation provisions, which the CMS began exercising in 2023-2024 targeting ten high-cost Medicare Part D drugs, created a new competitive layer on top of the existing Hatch-Waxman framework.<\/p>\n\n\n\n

    For NLP-based patent litigation prediction, the IRA creates a new class of documents: CMS negotiation-related filings, industry comments on the negotiation process, and the strategic communications of companies facing both IRA negotiation and patent expiry simultaneously. These documents carry market-structure sentiment that affects the brand manufacturer’s patent defense incentive calculus. A drug facing IRA price negotiation in 2026 and patent expiry in 2028 may have materially lower patent defense value than a drug facing patent expiry in 2032 without IRA exposure. Monitoring the convergence of IRA negotiation language with patent litigation language is a newer but important analytical layer.<\/p>\n\n\n\n

    \n\n\n\n

    The Future of Sentiment Analysis in Pharmaceutical Patent Prediction<\/strong><\/h2>\n\n\n\n

    Large Language Models as Active Litigation Analysis Tools<\/strong><\/h3>\n\n\n\n

    The emergence of large language models (LLMs) capable of reading and reasoning about long legal documents changes the NLP toolkit available for pharmaceutical patent litigation analysis. Where fine-tuned BERT models classify documents against learned patterns, LLMs can be prompted to explain why a specific Markman order represents a favorable or unfavorable claim construction, identify the specific arguments that swayed the court, and generate comparative analysis against prior cases with similar factual patterns.<\/p>\n\n\n\n

    The limitation remains accuracy on highly technical pharmaceutical chemistry and claim language. General-purpose LLMs trained on broad internet corpora have inconsistent performance on the specific technical terms that define pharmaceutical patent claims. Fine-tuned models remain preferable for the technical claim analysis component, while LLMs add value in the legal reasoning layer \u2014 explaining the significance of a specific claim construction ruling in plain language for non-lawyer stakeholders.<\/p>\n\n\n\n

    Real-Time Oral Argument Monitoring and the ‘Sentiment Before the Decision’ Opportunity<\/strong><\/h3>\n\n\n\n

    Several federal district courts and the Federal Circuit now make oral argument audio available promptly after hearings. Oral argument transcripts, combined with audio sentiment analysis (detecting hesitation, skepticism, and engagement from vocal tone), represent the closest thing to a real-time prediction signal available before a judicial ruling. The Federal Circuit oral argument in a pharmaceutical patent appeal, if listened to through an NLP\/audio analysis lens, often makes clear which way the panel is leaning a week or more before the opinion issues. For investors managing positions in companies with pending Federal Circuit patent appeals, that lead time is commercially significant.<\/p>\n\n\n\n

    Causal vs. Predictive Models: The Next Research Frontier<\/strong><\/h3>\n\n\n\n

    Most NLP-based litigation prediction systems are predictive: they identify patterns in historical data that correlate with outcomes, without claiming to explain why those patterns exist. The more ambitious research question is causal: which specific arguments, framed in which specific ways, actually change judicial outcomes rather than merely correlating with outcomes because they appear in the types of cases that tend to win for other reasons?<\/p>\n\n\n\n

    Causal inference in legal NLP is methodologically difficult because you cannot run randomized experiments on judges. Quasi-experimental methods \u2014 comparing outcomes when the same legal argument appears before different judges, or when similar factual situations are argued with stylistically different language \u2014 offer partial identification of causal effects. For pharmaceutical patent litigation, where the same claim types (polymer crystal form, pH-dependent release mechanism, enantiomeric purity) recur across dozens of cases with different judges and different outcomes, the data density is sufficient to support rudimentary causal analysis. The practical upshot would be evidence-based guidance on how to argue specific pharmaceutical patent claim types before specific judges \u2014 a significant competitive advantage for litigation counsel.<\/p>\n\n\n\n

    \n\n\n\n

    Key Takeaways<\/strong><\/h2>\n\n\n\n
      \n
    • Pharmaceutical patent litigation generates structured, consistent text at high volume, making it one of the most productive domains for NLP-based outcome prediction.<\/li>\n\n\n\n
    • Sentiment analysis works differently in legal text than in consumer text. Standard tools (VADER, TextBlob) fail on judicial opinions. Fine-tuned LegalBERT or domain-adapted transformer models are the current standard.<\/li>\n\n\n\n
    • The Markman claim construction order is the single highest-density predictive document in an ANDA case. Sentiment analysis of that order \u2014 before the parties have processed its implications \u2014 offers a genuine analytical lead time.<\/li>\n\n\n\n
    • IPR prediction requires two separate models: one for institution probability (where base rates are high and the value is in identifying denials) and one for final written decision outcomes (where base rates favor invalidity and the value is in identifying surviving patents).<\/li>\n\n\n\n
    • Settlement prediction is a separate task from merits prediction. Procedural filings, filing frequency, and post-Markman brief tone all carry settlement signal that a well-designed system should monitor.<\/li>\n\n\n\n
    • Raw win\/loss statistics in ANDA litigation are misleading because settlements dominate outcomes. A 2024 analysis found innovator companies prevailed on the merits 20% of the time and generics only 2%, figures that ignore that the vast majority of cases never reach a merits ruling. Model on settlement-adjusted outcomes, not raw court win rates.<\/li>\n\n\n\n
    • Judge-specific profiles are the most powerful predictor after patent quality. NLP-based judicial language modeling outperforms simple win-rate statistics because it captures how a judge reasons, not just how a judge votes.<\/li>\n\n\n\n
    • The commercial output of a working system is an updated probability distribution over LOE dates \u2014 not just a prediction of ‘win or lose’ \u2014 which feeds directly into patent portfolio NPV models and revenue forecasts.<\/li>\n\n\n\n
    • Validation must use temporal splits, not random splits. Models that look good on random held-out data often fail in production because they have learned time-specific judicial language patterns that do not generalize.<\/li>\n\n\n\n
    • DrugPatentWatch’s structured pharmaceutical patent and litigation metadata provides the backbone for linking NLP sentiment scores to drugs, companies, and commercial outcomes \u2014 translating model predictions into business intelligence.<\/li>\n<\/ul>\n\n\n\n
      \n\n\n\n

      Frequently Asked Questions<\/strong><\/h2>\n\n\n\n

      What documents should I prioritize for sentiment analysis in a new ANDA case?<\/strong><\/h3>\n\n\n\n

      Start with the Paragraph IV notice letter (previews the challenger’s legal theory), the brand’s Markman opening brief and the court’s claim construction order (highest predictive density), and any IPR petition filed in parallel. If resources are constrained, the Markman order alone provides the most prediction value per document-analysis hour.<\/p>\n\n\n\n

      Can sentiment analysis predict which PTAB panel will grant institution?<\/strong><\/h3>\n\n\n\n

      With panel composition data and a sufficiently large corpus of prior decisions by the assigned APJs, yes \u2014 at better-than-base-rate accuracy. Bio\/pharma petitions ran a 73% overall institution rate in FY24, but specific APJ combinations with prior obviousness-skeptical language profiles in their published decisions are identifiable through NLP and have below-average institution rates on pharmaceutical patent claims.<\/p>\n\n\n\n

      How does claim construction reversal risk affect the LOE probability model?<\/strong><\/h3>\n\n\n\n

      Federal Circuit de novo review of claim construction means district court Markman rulings are not the end of the prediction task. Your model should score Federal Circuit reversal risk separately: rulings that relied heavily on extrinsic evidence (expert testimony) rather than intrinsic evidence (the patent’s own specification and prosecution history) are more vulnerable to Federal Circuit revision under the post-Teva v. Sandoz mixed standard of review.<\/p>\n\n\n\n

      What is the training data minimum for a reliable pharmaceutical patent litigation NLP model?<\/strong><\/h3>\n\n\n\n

      For a PTAB institution decision prediction model, 200+ labeled petition-outcome pairs can produce a functional model, though 500+ significantly improves generalization. For Markman outcome prediction, the dataset is smaller (fewer cases go to Markman) and you may need to augment with transfer learning from non-pharmaceutical ANDA cases or combine district courts to build sufficient volume. Below 100 labeled examples, stick to structured metadata models with NLP features as supplements rather than the primary input.<\/p>\n\n\n\n

      How should I handle the 30-month stay expiration when modeling LOE risk?<\/strong><\/h3>\n\n\n\n

      Model the 30-month stay as a probability-conditional time horizon. If IPR institution probability is high (above 60% based on petition sentiment), model a parallel track where the PTAB final written decision could precede the 30-month stay expiration and invalidate the Orange Book patent before district court trial. In that scenario, the brand’s automatic stay protection disappears. The probability-weighted expected LOE date incorporates both the district court and PTAB timelines.<\/p>\n\n\n\n

      What does it mean when a PTAB institution decision uses more hedged language than usual?<\/strong><\/h3>\n\n\n\n

      Higher-than-baseline hedging in a PTAB institution decision \u2014 language like ‘petitioner has presented some evidence,’ ‘we are not persuaded at this stage that,’ or extensive acknowledgment of the patent owner’s counterarguments \u2014 signals a closer call that may not hold through the full trial. Petitioners whose institution decision shows hedged rather than confident language for specific claims should treat those claims as less secure going into the final written decision stage, even though institution was granted.<\/p>\n\n\n\n

      Can this approach be applied to European Patent Office (EPO) opposition proceedings?<\/strong><\/h3>\n\n\n\n

      Yes, with important caveats. EPO opposition decisions are written in English, German, or French (with official translations available). The procedural format is consistent enough that NLP-based opposition prediction is feasible. The legal standards differ from U.S. practice (inventive step analysis under EPO guidelines is not identical to obviousness under U.S. \u00a7 103), so models trained on USPTO or PTAB data do not transfer to EPO proceedings without substantial retraining on EPO-specific text.<\/p>\n\n\n\n

      How does the FTC’s Orange Book delisting campaign affect patent litigation sentiment models?<\/strong><\/h3>\n\n\n\n

      The FTC’s 2023-2024 campaign against improper Orange Book listings introduced a new category of litigation risk: patents that were previously assumed to generate an automatic 30-month stay may be delisted and lose that protective function. NLP monitoring of FTC administrative filings and related district court challenges to Orange Book listings is now a required component of a complete pharmaceutical patent litigation intelligence system.<\/p>\n\n\n\n

      What sentiment signals predict whether a generic manufacturer will pursue at-risk launch?<\/strong><\/h3>\n\n\n\n

      At-risk launch decisions follow an adverse Markman for the brand manufacturer, a favorable PTAB institution decision, and internal sentiment that the challenger’s claim construction theory is strong enough to withstand appeal. The specific NLP signal is the challenger’s public legal communications (press releases, 8-K filings, earnings call language about the challenged drug) becoming more confident in tone post-Markman. Generic manufacturers who have secured a favorable claim construction and whose legal communications shift from ‘we believe we have strong arguments’ to ‘we have obtained a favorable ruling’ are signaling internal confidence in an at-risk launch.<\/p>\n\n\n\n

      How do I score the strength of secondary considerations of nonobviousness arguments in IPR proceedings?<\/strong><\/h3>\n\n\n\n

      Secondary considerations (commercial success, long-felt need, failure of others, unexpected results) appear in patent owner responses and expert declarations. NLP scoring for these arguments requires identifying the claim type (commercial success vs. unexpected results are linguistically and legally distinct), the specificity of the nexus argument (does the evidence show that commercial success is attributable to the claimed invention rather than marketing?), and the quality of the quantitative evidence referenced. Patent owner responses that include specific sales figures, market share data, or clinical outcome comparisons tied directly to claim language score higher on secondary-considerations strength than those making generic commercial success assertions.<\/p>\n\n\n\n

      What is the biggest mistake analysts make when applying sentiment analysis to patent litigation?<\/strong><\/h3>\n\n\n\n

      Treating case-level outcome as the prediction target without claim-level granularity. A brand manufacturer whose patent is found valid on some claims but invalid on others, or valid but not infringed, ‘won’ and ‘lost’ simultaneously. A prediction model that classifies the case as a ‘brand win’ misrepresents the commercial outcome if the surviving claims do not cover the accused generic product. Always label and predict at the claim level, then aggregate to case level with explicit logic about which claim outcomes determine the commercial result.<\/p>\n\n\n\n

      \n\n\n\n

      Sources<\/strong><\/h2>\n\n\n\n
        \n
      1. Womble Bond Dickinson. (2025, January 15). 2024 Hatch-Waxman year in review<\/em>. https:\/\/www.womblebonddickinson.com\/us\/insights\/articles-and-briefings\/2024-hatch-waxman-year-review<\/li>\n\n\n\n
      2. National Law Review. (2025, January 15). 2024 Hatch-Waxman litigation trends and key Federal Circuit decisions<\/em>. https:\/\/natlawreview.com\/article\/2024-hatch-waxman-year-review<\/li>\n\n\n\n
      3. DrugPatentWatch. (2026, February 23). Top Paragraph IV litigation trends and what they mean for pharma<\/em>. https:\/\/www.drugpatentwatch.com\/blog\/top-paragraph-iv-litigation-trends-and-what-they-mean-for-pharma\/<\/li>\n\n\n\n
      4. DrugPatentWatch. (2026, February 11). Uncovering the success patterns in modern Paragraph IV litigation<\/em>. https:\/\/www.drugpatentwatch.com\/blog\/uncovering-the-success-patterns-in-modern-paragraph-iv-litigation\/<\/li>\n\n\n\n
      5. DrugPatentWatch. (2026, March 8). Win the patent fight: Hatch-Waxman litigation strategies for brand and generic manufacturers<\/em>. https:\/\/www.drugpatentwatch.com\/blog\/win-the-patent-fight-hatch-waxman-litigation-strategies-for-brand-and-generic-manufacturers\/<\/li>\n\n\n\n
      6. DrugPatentWatch. (2025 May). The Hatch-Waxman playbook: Paragraph IV certifications, 180-day exclusivity, and the $467 billion generic drug race<\/em>. https:\/\/www.drugpatentwatch.com\/blog\/the-hatch-waxman-playbook-paragraph-iv-certifications-180-day-exclusivity-and-the-467-billion-generic-drug-race\/<\/li>\n\n\n\n
      7. DrugPatentWatch. (2026, March 24). The Patent Trial and Appeal Board: The definitive analyst’s guide to IPR strategy<\/em>. https:\/\/www.drugpatentwatch.com\/blog\/understanding-the-patent-trial-and-appeal-board-ptab-a-comprehensive-overview\/<\/li>\n\n\n\n
      8. DrugPatentWatch. (2025, November 19). How safe is your drug patent from PTAB challenges? A strategic guide for pharma leaders<\/em>. https:\/\/www.drugpatentwatch.com\/blog\/how-safe-is-your-drug-patent-from-ptab-challenges-a-strategic-guide-for-pharma-leaders\/<\/li>\n\n\n\n
      9. PTAB Law Blog. (2025, January 6). Trial statistics trends at the PTAB: 2024 edition<\/em>. https:\/\/www.ptablaw.com\/2025\/01\/06\/trial-statistics-trends-at-the-ptab-2024-edition\/<\/li>\n\n\n\n
      10. IPWatchdog. (2025, January 12). The PTAB’s 70% all-claims invalidation rate continues to be a source of concern<\/em>. https:\/\/ipwatchdog.com\/2025\/01\/12\/ptab-70-claims-invalidation-rate-continues-source-concern\/<\/li>\n\n\n\n
      11. Finnegan, Henderson, Farabow, Garrett & Dunner. (2024). Trends in PTAB trials involving drug and biologic patents<\/em>. https:\/\/www.finnegan.com\/en\/insights\/blogs\/at-the-ptab-blog\/trends-in-ptab-trials-involving-drug-and-biologic-patents.html<\/li>\n\n\n\n
      12. Polsinelli. (2021, August 13). A brief overview of pharmaceutical IPRs and statistical outcome<\/em>. https:\/\/www.polsinellionpostgrant.com\/blog\/2017\/6\/2\/a-brief-overview-of-pharmaceutical-iprs-and-statistical-outcome<\/li>\n\n\n\n
      13. Park, B., Bhardwaj, N., Yi, S., & Hsu, Y. (2024). Predicting patent litigation risk using RoBERTa<\/em>. Stanford CS224N Final Project. https:\/\/web.stanford.edu\/class\/archive\/cs\/cs224n\/cs224n.1244\/final-projects\/BrianParkNikitaBhardwajSimoneYiYiHsu.pdf<\/li>\n\n\n\n
      14. Mart\u00ednez-Gil, J., Freudenthaler, B., & Natschl\u00e4ger, T. (2022). Survey of text mining techniques applied to judicial decisions prediction<\/em>. Applied Sciences, 12<\/em>(20), 10200. https:\/\/www.mdpi.com\/2076-3417\/12\/20\/10200<\/li>\n\n\n\n
      15. Suzgun, M., et al. (2022). Predicting institution outcomes for inter partes review (IPR) proceedings at the PTAB by deep learning of patent owner preliminary response briefs<\/em>. Applied Sciences, 12<\/em>(7), 3656. https:\/\/www.mdpi.com\/2076-3417\/12\/7\/3656<\/li>\n\n\n\n
      16. Mazzeo, M. J., Hillel, J., & Zyontz, S. (2013). Explaining the ‘unpredictable’: An empirical analysis of U.S. patent infringement awards. International Review of Law and Economics, 35<\/em>, 58\u201372.<\/li>\n\n\n\n
      17. Sierra IP Law. (2025, June 12). Markman hearing \u2014 determining the meaning of patent claims<\/em>. https:\/\/sierraiplaw.com\/markman-hearing\/<\/li>\n\n\n\n
      18. Teva Pharmaceuticals USA, Inc. v. Sandoz, Inc., 574 U.S. 318 (2015).<\/li>\n\n\n\n
      19. Feldstein, M. J. (2024). IPR and PGR statistics for final written decisions issued in September 2024<\/em>. Finnegan. https:\/\/www.finnegan.com\/en\/insights\/blogs\/at-the-ptab-blog\/ipr-and-pgr-statistics-for-final-written-decisions-issued-in-september-2024.html<\/li>\n\n\n\n
      20. Hernandez, A., et al. (2024). Serial patent litigation: an emerging strategy to delay entry of generic competition<\/em>. Nature Medicine<\/em>. https:\/\/www.ncbi.nlm.nih.gov\/pmc\/articles\/PMC12757684\/<\/li>\n\n\n\n
      21. DrugPatentWatch. (2025, July). 5 ways to predict patent litigation outcomes<\/em>. https:\/\/www.drugpatentwatch.com\/blog\/5-ways-to-predict-patent-litigation-outcomes\/<\/li>\n\n\n\n
      22. DrugPatentWatch. (2025, April). The algorithmic adjudicator: How big data and AI are revolutionizing pharmaceutical patent litigation<\/em>. https:\/\/www.drugpatentwatch.com\/blog\/the-algorithmic-adjudicator-how-big-data-and-ai-are-revolutionizing-pharmaceutical-patent-litigation\/<\/li>\n\n\n\n
      23. USPTO Patent Trial and Appeal Board. (2024). Trial statistics<\/em>. https:\/\/www.uspto.gov\/patents\/ptab\/statistics<\/li>\n<\/ol>\n","protected":false},"excerpt":{"rendered":"

        Pharmaceutical patent litigation produces more text per dollar at stake than almost any other legal domain. A single Hatch-Waxman case […]<\/p>\n","protected":false},"author":1,"featured_media":39092,"comment_status":"open","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"_lmt_disableupdate":"","_lmt_disable":"","site-sidebar-layout":"default","site-content-layout":"","ast-site-content-layout":"default","site-content-style":"default","site-sidebar-style":"default","ast-global-header-display":"","ast-banner-title-visibility":"","ast-main-header-display":"","ast-hfb-above-header-display":"","ast-hfb-below-header-display":"","ast-hfb-mobile-header-display":"","site-post-title":"","ast-breadcrumbs-content":"","ast-featured-img":"","footer-sml-layout":"","ast-disable-related-posts":"","theme-transparent-header-meta":"","adv-header-id-meta":"","stick-header-meta":"","header-above-stick-meta":"","header-main-stick-meta":"","header-below-stick-meta":"","astra-migrate-meta-layouts":"default","ast-page-background-enabled":"default","ast-page-background-meta":{"desktop":{"background-color":"var(--ast-global-color-4)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"tablet":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"mobile":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""}},"ast-content-background-meta":{"desktop":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"tablet":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"mobile":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""}},"footnotes":""},"categories":[10],"tags":[],"class_list":["post-38963","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-insights"],"modified_by":"DrugPatentWatch","_links":{"self":[{"href":"https:\/\/www.drugpatentwatch.com\/blog\/wp-json\/wp\/v2\/posts\/38963","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.drugpatentwatch.com\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.drugpatentwatch.com\/blog\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.drugpatentwatch.com\/blog\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.drugpatentwatch.com\/blog\/wp-json\/wp\/v2\/comments?post=38963"}],"version-history":[{"count":1,"href":"https:\/\/www.drugpatentwatch.com\/blog\/wp-json\/wp\/v2\/posts\/38963\/revisions"}],"predecessor-version":[{"id":39338,"href":"https:\/\/www.drugpatentwatch.com\/blog\/wp-json\/wp\/v2\/posts\/38963\/revisions\/39338"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.drugpatentwatch.com\/blog\/wp-json\/wp\/v2\/media\/39092"}],"wp:attachment":[{"href":"https:\/\/www.drugpatentwatch.com\/blog\/wp-json\/wp\/v2\/media?parent=38963"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.drugpatentwatch.com\/blog\/wp-json\/wp\/v2\/categories?post=38963"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.drugpatentwatch.com\/blog\/wp-json\/wp\/v2\/tags?post=38963"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}